This company is hard to value, and the outlook has gotten better pretty much every quarter since this time last year. If it was easy, Nebius would have traded above $100 that whole time.

So yes, this is a groundbreaking quarter, and you’re going to see quite a few changes in my approach based on a few factors:

• Improved unit economics

• Higher confidence in management’s ability to execute

• Astoundingly prudent financial engineering

All of this materially changes the valuation.

I’ve run a series on Nebius for some time now called the Four Horsemen of the Nebius Apocalypse. Back when it was trading around $70, sceptics were everywhere, and there were four main risks to the business which, if they came to fruition, were totally legitimate harbingers of debt and doom.

For now the horsemen are nowhere to be seen, and the apocalypse feels like a forgotten chapter in Nebius’s history. All four risks have seen material mitigations and improvements, and I believe this is the pivotal moment where Nebius starts being taken seriously as a potential AI-native hyperscaler.

A quick word on where AI actually is right now

Before we get into the numbers, I want to share something that’s been doing the rounds and gave me pause.

Courtesy of @TheRealDanSaedi on twitter

78% of the world’s population has never used AI. Another 21% are casual free-tier chatbot users. Paid users sit at 0.72%. People actually using AI to do meaningful work? 0.12%.

Amongst all the hype, it just shows how early AI consumption really is. We are debating bubble valuations on AI infrastructure while less than one in a hundred people has ever paid for the damn thing.

The total addressable market for AI compute isn’t priced into anything yet. Not into Nvidia, not into Microsoft, and definitely not into Nebius. This is the context in which Q1 2026 numbers should be read. And if you need more information on the AI bubble, check out my recent article on the path for Anthropic getting to $1 trillion in revenue by 2030 - that’s the backdrop I’m looking at this against.

Anyway, let me explain what I’m doing today.

My plan for this article

You will have read the news on the acquisitions Nebius has made, and you will have read the highlights from the earnings. I’m not going to bother repeating any of that.

I’m going to give you my analysis on where things stand, what this means for Nebius’s outlook, and what price we should be paying for this business.

The article is structured as follows:

1. The updated valuation - my new base case of $340, up from $194 last quarter, and what’s driving the move.

2. Status of the Four Horsemen in the wake of the current quarter.

3. My new bull case - what happens if Nebius executes at the upper end of its potential.

4. Portfolio planning and positioning going forward.

Let’s do it.

1. The Updated Valuation

I’ve been particularly conservative on Nebius up until this point to ensure my valuation has a significant margin of safety, in what has been a new and very uncertain market. Unit economics and profitability were theoretical for most of last year, and I did not want to assume hero numbers until I had evidence to back the claim.

A year has now passed since my first piece on Nebius at $51, and we have meaningful quarters of operating data. The thesis has moved from theory to demonstration. I’m also moving the terminal year of the model out to 2031 - which is favourable on the valuation as a lot of the value comes after the heavy capex spend is done.

So here is where we land:

A base case fair value of $340 against a current share price of $207.27. That’s +63% upside, with multiples held at the same level as my March piece (27.5x terminal). Every dollar of the move from March’s $194 PV comes from operating numbers, not re-rating. The maths is now better-supported by the data, not by being more aggressive with assumptions.

For context on where this sits with the street: Morgan Stanley raised their price target yesterday from $126 to $144 provisionally, while explicitly noting they are going back to the drawing board on their entire neocloud valuation framework. Their new working framework has a bull case of $400 and a bear of $70 - which they themselves describe as creating “a solid risk/reward.” When the most conservative shop on the street starts using the phrase “compelling bull case” about Nebius, you know the narrative has shifted.

I’m in that same direction of travel but with slightly higher conviction I would say, I’ll explain why.

4 GW is no longer a pipe dream. Last quarter, my base case assumed 2.5 GW of operational capacity by 2030. Today, Nebius has already contracted over 3.5 GW of power, and management has raised the year-end 2026 guide to more than 4 GW. The NVIDIA partnership disclosed on the call goes further:

Arkady Volozh: “...the NVIDIA strategic investment is meaningful across several dimensions, beyond the $2 billion of equity and line of sight of five gigawatts of capacity commitment by the end of 2030.”

So 4 GW operational by 2030 is no longer my optimistic anchor - it’s the conservative read of what Nebius is now lining up. My base case has been built around what has already been delivered or contracted, not around speculative future wins.

Worth flagging: the capex raise from $16-20B to $20-25B isn’t inflation hitting the business. Management was crystal clear that “the impact of component inflation in our 2026 program was around low single digits” - they locked in 2026 inputs back in 2025. The raise is them deciding to invest more aggressively into 2027 capacity that already has customer commitments. That’s exactly the kind of capex raise I want to see.

Four customers competing for every GPU coming online. This number, from Mark Boroditsky on the call, is the single best demand-side data point in the print. Combined with the Meta deal restructure clarified in the shareholder letter, it dramatically de-risks the unit economics assumption.

The $27B Meta deal now has a structure that didn’t quite land in March: $12B firm five-year compute purchase beginning early 2027, plus $15B in contracted capacity that Nebius can sell to Meta on pre-agreed terms or to AI cloud customers at market rates. In other words, Meta is providing a floor on a chunk of capacity while Nebius retains the ability to sell that capacity to higher-paying customers if demand materialises. They get a risk-free attempt to sell their vertical stack to enterprise customers, and if those customers don’t show up, Meta absorbs the capacity. Heads they win on price, tails they win on guaranteed offtake.

The combination of 4:1 oversubscription and the Meta deal structure is what makes me comfortable raising the per-MW revenue assumption from $10m to closer to $11m in the base case. And realistically, they’ll sell at a higher revenue/MW premium to enterprise customers. The bull case I’ll get to in a moment leans into this even harder.

Unit economics have dramatically improved. Q1 group adjusted EBITDA was $129.5m (32% margin), versus a $53.7m loss in the year-ago quarter. AI cloud adjusted EBITDA margin reached 45% in Q1, up from 24% in Q4 2025. Yes, 45%. I had to check that as well.

(One honest disclosure: reported net income of $621m looks great but is flattered by a $781m non-cash gain on Nebius’s ClickHouse stake revaluation. The operating story is real; the headline net income number is not the right thing to anchor on.)

To be honest with you, this is the number I struggled to model in my framework. When I set the November scorecard, I assumed Nebius would reach 12-15% net margins by 2030. Q1 AI cloud EBITDA is now telling us the trajectory is steeper. Volozh, in his letter, was explicit about it:

“...our trajectory toward 20-30% EBIT margins.”

That’s a quoted statement from the CEO about EBIT, not EBITDA. After tax and depreciation, that maps to net margins in the high-teens to low-twenties. My base case now uses 18% net margin terminal - still below the upper end of what Volozh is suggesting. The bull case uses more. We’re not building heroic numbers here; we’re building defensible ones.

Depreciation is collapsing as the thesis plays out. This one matters more than people realise. Michael Burry’s argument against neoclouds last year was that GPUs become obsolete and useless after four years, and that companies using long depreciation schedules were artificially inflating returns. The implication was that the whole space was a bubble.

The data has not been kind to that thesis. Q1 depreciation as a percentage of revenue came in at 53% - already at the level my March model assumed we wouldn’t reach until 2027. Old GPUs are not collapsing in value; they’re holding pricing on inference workloads, and Nebius has been clear that pricing actually rose on new generations in Q1 while holding firm on older ones. The bear thesis on GPU depreciation hasn’t materialised because inference demand is real, growing faster than training, and tolerates older hardware just fine.

This isn’t a victory lap on the Burry argument. It’s a single quarter and the bear thesis isn’t necessarily wrong, just yet to play out as expected. But the data is now consistently moving in the right direction, and the 4yr to 5yr useful life extension that Nebius made in Q1 isn’t financial engineering disguised as economics - it’s accounting catching up to reality.

Q1 convertibles plus Microsoft prepayment plus asset-backed financing comments make funding massively de-risked. This is the change I find most encouraging, and it’s worth dwelling on because dilution was my biggest concern in March.

Nebius raised $6.3 billion in Q1 without touching the ATM - $4.3B from convertibles at 1.25-2.625% rates (extraordinarily cheap given current market conditions), plus $2.0B from the NVIDIA prefunded equity warrants. The Microsoft prepayment, sitting in deferred revenue, jumped from $1.6B at year-end to $4.8B at Q1 end - a $3.2B cash injection that funds capex without affecting the P&L or share count.

On the call, management telegraphed the next leg of financing very explicitly:

Arkady Volozh: “We are actively progressing potential debt transactions, including asset-backed financing and corporate-level debt. We plan to start tapping into these financing options by raising mid-single-digit billions of dollars in the near term.”

For readers unfamiliar with asset-backed financing: this is essentially borrowing against the specific cash flows of Nebius’s contracted customer revenue. The Microsoft and Meta contracts are AAA-rated equivalents from a credit perspective, so debt secured against them prices very tightly. Crucially, it doesn’t dilute shareholders. Nebius is monetising the contracts upfront and using the proceeds to build capacity, while customer payments service the debt.

Management’s stated funding hierarchy is now clear: prepayments first, asset-backed financing second, corporate debt third, ATM as last resort. The ATM remains unused. I’m baking in some dilution in 2027 to fund the residual capex gap, but it’s modest compared to what I feared in March. Some dilution but not an egregious amount, as I’d put it.

The cumulative effect of these changes on the model is what gives us the $340 base case. None of it requires heroic assumptions, all of it is grounded in something Nebius said or showed on Q1.

2. How The Four Horsemen Look in Light of Q1

Three months ago I published the Dilution Paradox piece, which acknowledged a Q4 print that mostly went well but flagged convertible-note dilution as the killer that could break per-share returns. Since then, two-thirds of the 2026 capex year has been funded without equity, the dilution number is contained, and the operating data has rolled in better than I projected.

Here’s the updated scorecard for the base and bull cases:

All four horsemen still pass. Three of them pass cleanly. The fourth - dilution - I’m calling “PASS with reservations” because the financing story still has 2027 and beyond to play out. Let me walk through each one.

2.1 Revenue Growth

November concern: Can Nebius bring 2.5 GW+ online and convert it to $10m per MW?

Where we stand: Comfortably ahead. Q1 2026 printed $399m against consensus near $317m - a 26% beat, with revenue growth of +75% sequential. Annualised run-rate hit $1.92B at Q1 month-end, up from $1.25B at year-end - a 54% sequential jump that re-accelerated from Q4’s pace. Pipeline up 3.5x quarter-on-quarter (and importantly, that excludes hyperscaler deals like Meta). And underneath all of that, the demand-side number that defends the entire pricing thesis: 4+ customers competing for every GPU Nebius brings online. Arkady put it more bluntly on the call: “Demand is high. Everything we build is sold.”

Contracted power has been raised four times in nine months: >1 GW in August, >2.5 GW in November, >3 GW in February, >4 GW today. Year-end 2026 guidance is now >4 GW with the NVIDIA-partnership aiming for 5 GW by 2030.

Impact on base case: Drives the move from 2.5 GW to 4 GW capacity, which contributes roughly +$120 to the PV. The single largest driver of the new base case.

What I’d be wrong about: Connected (operational) capacity by year-end is still guided to 800 MW - 1 GW, with the majority arriving in Q3. If Q3 doesn’t deliver the promised capacity step-up, the FY guide cuts and the first horseman will require us to do some remodelling of our timelines.

2.2 Depreciation as % of Revenue

November concern: 66% in FY25, scaling down to 20-25% by 2030. Could GPU depreciation crush margins, particularly with the bear thesis (e.g. Burry) suggesting GPUs become worthless after four years?

Where we stand: Q1 actual depreciation was 53% of revenue. To put that in perspective, my March base case model said depreciation wouldn’t reach 53% until 2027. Nebius got there in one quarter. The improvement isn’t an accounting trick from the useful-life extension, though that helps - the bigger driver is that revenue is scaling faster than the depreciable asset base. Old GPUs are still earning revenue. New GPUs are commanding higher prices. The trajectory is moving faster in the right direction than I dared model, and to be honest could even imagine.

Impact on base case: Enables the terminal net margin assumption to move from 15% to 18% without re-rating the multiple. Worth roughly +$50 to the PV.

What I’d be wrong about: This is currently the safest of the four horsemen, but one quarter of dramatic improvement doesn’t make a trend. The 4yr to 5yr useful life extension is mechanically helping the reported number. I’ve kept the February haircut on the exit multiple (27.5x rather than 30x) as discipline against this question being weighted too heavily. If the trajectory reverses in Q2 or Q3, the haircut absorbs it.

2.3 Net Income Margin

November concern: Can Nebius turn profitable by 2027 and reach 12-15% margins by 2030?

Where we stand: Q1 AI cloud adjusted EBITDA margin was 45%. Group adjusted EBITDA margin was 32%. Group adjusted EBITDA of $129.5m flipped from a year-ago loss of -$53.7m. Cost of revenue dropped from 49% to 26% of revenue. SG&A from 120% to 36%. R&D from 72% to 17%. Every single line of operating leverage is moving the right way.

Volozh’s “trajectory toward 20-30% EBIT margins“ is the most important forward-looking statement of the print. Translating to net margins after depreciation and tax lands you in the high-teens to low-twenties range. My base case at 18% terminal net margin sits in the middle of that. The bull case I’ll get to uses more.

Impact on base case: Justifies the 15% → 18% terminal margin move. Bull case at 25% net is consistent with the upper end of management’s stated EBIT trajectory.

What I’d be wrong about: Pricing power is one quarter old. Q1 saw pricing rise on new-gen GPUs and hold on older-gen, and the 4:1 customer-per-GPU ratio suggests this should hold for at least a few more quarters. But one quarter is a data point, not a trend. The $11-12.5m per MW that drives the bull case is the assumption to watch.

And to be honest about the 45% AI cloud EBITDA - it’s unusually strong, and management told us not to extrapolate it. From the call: “We expect margins in Q2 to go a little bit lower, returning to Q1 levels in Q3, and stepping even higher in Q4.” In other words, Q2 will print softer because of hiring and acquisition-related costs landing before the capacity that supports them comes online. That’s a timing dynamic, not a thesis problem - but it means the 40% FY26 EBITDA guide is the right level to model against, not 45%.

2.4 Shares Outstanding

November concern: Will dilution destroy per-share returns or remain controlled?

Where we stand: This is the horseman with the most nuance, and the one that prevents this from being a clean victory lap. The good news is real:

• Q1 raised $6.3B without touching the ATM - $4.3B convertibles at 1.25-2.625%, plus $2.0B NVIDIA prefunded warrants.

• Deferred revenue jumped from $1.58B at year-end to $4.78B at Q1 end - a $3.2B increase that is the Microsoft prepayment flowing into the cash balance and funding capex.

• Asset-backed financing of mid-single-digit billions is being progressed against the Microsoft and Meta contracts.

• The ATM (dilution mechanism), despite being available, remains completely unused.

But the hard news is that 2027 capex is genuinely large, and even with all four non-equity financing levers pulled at scale, some residual gap is likely. My base case now assumes roughly 25 million shares of ATM-style dilution in 2027 to bridge it - diluted share count progressing from 309M today to 365M by 2031. That’s “some dilution but not loads,” and it costs the base case about 14 percentage points of upside.

Impact on base case: -$37 in base case PV from the additional dilution path. The status is still PASS, but I’ve moved it to “PASS with reservations” until the asset-backed financing programme is announced and sized in subsequent quarters.

What I’d be wrong about: Three things sit under this horseman.

First, the 2027 financing gap. 2026 funds without equity (~$8B deployable cash + ~$3B operating cash flow + ~$5-7B asset-backed + ~$2-5B corporate debt covers the $20-25B capex). 2027 is tighter: $20-26B capex against $4-6B operating cash flow leaves a $15-20B gap. Asset-backed scales but a residual gap is likely. If asset-backed markets close or hyperscaler counterparties get downgraded, that 25M shares could double and base case PV drops toward $300.

Second, owned capacity is now >75% of contracted, a structural pivot from leased. Better long-term economics, but a 1.2 GW Pennsylvania build represents billions of sunk capex before any revenue. Permitting delays, supply chain disruption, or local opposition could mean capital deployed against revenue that arrives late.

Third, the basic-vs-diluted share count gap is wide. Most coverage of Nebius uses 253M basic shares. Diluted today is 309M, which means market cap is $64B not $52B at $207.27. The base case PV of $340 implies a $105B fair market cap. None of this is hidden, but the gap between basic and diluted is big.

What could bring the apocalypse back

I want to be specific about this because otherwise people are going to be screenshotting this essay and mailing it to me if things go wrong. Three things would force me to seriously rethink:

1. Q3 capacity step-up doesn’t arrive. If connected MW doesn’t track to the 800 MW - 1 GW year-end guide, the FY revenue cuts, ARR exit cuts, and Horseman 1 properly fails for the first time since the framework was set.

2. GPU pricing softens for two consecutive quarters. One quarter of rising new-gen pricing is encouraging. Two would be a trend. Sustained softening would unwind the margin trajectory and force a re-rate.

3. Asset-backed financing markets close. If credit conditions tighten or hyperscaler counterparties get downgraded, Nebius would have to fund the 2027 capex gap through equity, and the dilution path materially worsens.

For now, none of these are flashing. But they are the watchlist items for Q2 and Q3.

3. The Bull Case: $1,000 at 6 GW

I want to be upfront about something. The bull case I’m laying out below would have felt absurdly aggressive six months ago. It would have felt aggressive even three months ago. Today, after Q1, it feels like a credible scenario - not the central estimate, but a path that requires execution rather than miracles.

Read more

The market has noticed. Sceptics are everywhere now, with many calling tops warning that AI infrastructure is wildly overstretched, that the bubble is forming, that the maths doesn’t work.

I engaged with the bubble question directly back in November, and have so far been proven to be somewhat on the money.

The framework I used then was simple: own the bottleneck.

Five months on, and after huge gains in AI infrastructure stocks, Anthropic’s $15B of net ARR added in three weeks tells me the bottleneck is not going anywhere.

In fact, the last 6 months has been crazy, yet I honestly think you aint seen nothing yet.

I think Anthropic specifically has a credible path to $1 trillion ARR by 2034, with $600 billion as my base case for 2030.

Two reasons:

• First, the compute commitments. Anthropic has signed more than $380 billion in compute spend and over 11 gigawatts of capacity. Amazon. Google. Microsoft. NVIDIA. Fluidstack. SpaceX. The deals stretch to the early 2030s. To put 11 GW in context: that’s roughly the entire installed electricity generation capacity of Ireland. Five times the United Kingdom’s total data centre footprint today, and nearly twice what the UK government has set as its most ambitious 2030 target.

• Second, the maths. To make commitments of this scale economic at reasonable inference margins, Anthropic needs annual revenue somewhere north of $300 billion by the end of the decade. The contracts force the trajectory. There is no scenario in which Anthropic signs these deals and the revenue doesn’t materially follow - short of bankruptcy, which Amazon, Google, and the rest of the consortium would have to have wildly miscalibrated to allow.

So that’s where we are. $45 billion ARR today. A revenue requirement north of $300 billion by 2030. A stated ambition of $1 trillion by 2034.

This isn’t a moonshot call, it’s a data driven, evidence based, logical conclusion.

And it is the calibrated framework I anchor every AI infrastructure valuation on.

Power. Compute. Memory. Networking. Robotics. Every name in the trade - Nebius, IREN, Micron, the hyperscalers - only makes sense at current multiples if the frontier AI labs hit certain revenue thresholds. In my opinion, Anthropic specifically clears those thresholds. By a lot.

That’s what I’m going to spend the rest of the article on. Let me show you the argument, and you can tell me how crazy I am at the end….

Here’s what I’ll cover

1. Why AGI by 2029 is the assumption everything else depends on

2. The maths the $380 billion of contracts force on Anthropic

3. How four revenue lines build to $600B by 2030, and $1T not long after

4. The bears I take seriously

5. What this means for the rest of the AI infrastructure trade

The step change: why AGI by 2029 is what makes these numbers rational

I think AGI by 2029 is a sensible take.

Everything I’m about to argue rests on that - so let me explain why I think it, before I ask you to follow me into the maths.

The core reason is that three independent forecasts arrive within this range from completely different starting points - and Ray Kurzweil who I’ve mentioned many times on substack, pinpoints this exact date with a decent track record. See below:

Ray Kurzweil in 1999, anchored to a Law of Accelerating Returns argument about compute curves, lands at 2029. Dario Amodei in October 2024 (Machines of Loving Grace), anchored to scaling-law extrapolation from training runs inside Anthropic, lands at 2026-27. Demis Hassabis in 2024, anchored to capability benchmarks at DeepMind, lands at 2029-34.

Read more

I’ve followed that principle, and it’s worked well.

The average return from my publications is over 45% in my first year. (See here) But my actual portfolio has returned much higher - some of the losers I cut early, and I went much bigger on the high conviction winners like Nebius and Google.

• So what’s my process? How have I been limiting downside and maximise returns? Well here I’m going to tell you at a broad level, so you can do it too, or you can critique my process.

• I’ll provide the process, the tooling, and the principles.

It’s six stages, and they roughly look like this:

Each stage has its own job, and at the end of each you should have something concrete - a shortlist, a watchlist, a thesis, a number. The point of the funnel isn’t to generate work, it’s to give you the time and confidence to do nothing once you’re in a position. Conviction comes from process.

Here’s how each stage works.

01 - Discovery

This is the process of finding companies that you might be interested in. You can go about this in 2 ways:

1. Start with a business, then assess its investability

2. Start with an idea, a theme, a narrative, and find businesses within that domain

I mostly go for option 2, so I’m going to use that as the example here, but the principles apply to both entry points.

Desk based brainstorm

Sit down, and think. Sit on the bus and think, walk and think. But think.

I’m thinking about where the world will be in 5 years, and how certain trends might play out. Some examples:

• AI is not a bubble. Most knowledge workers will eventually be replaced by GPUs.

• The GLP-1 drug trend is just getting going, its going to have a profound impact when it hits mass scale, not just for obesity but for lots of disorders.

• Going back to the moon is the first step to a new space race, which will be capitalised on by private entities - with companies who can launch things into space being the owners of the bottleneck.

I’ll generally map these out as bubbles on an A4 sheet, and mark which are my strongest. I’ll then start noting companies which apply to my ideas off the top of my head.

I don’t know every company in the world, so this is generally where desk based research ends.

LLM

Utilising an LLM is the next logical step.

I have created skills, set prompts, and templates that I use to kick things off, but you can go with something like the following to get started:

“You are an investor interested in maximising returns over the next 5-10 years. You have a brief of finding who will benefit the most from [Trend X from Desk Based Research], find me 10 companies and rank them.”

Claude is my go-to, but I will generally repeat this same prompt in Gemini and ChatGPT until I have a synthesised list of potential ideas ready to begin monitoring.

Substack

Substack is a great place to find new ideas too. Simply search space economy, or GLP-1, or Artificial Intelligence - you’ll find talented writers posting about the topics you’re interested in, and often they’ll have great ideas about companies.

X

Same as above. Follow the right people, search keywords, stay up to date.

At the end of Discovery you should have: a list of around 10–15 companies tied to your strongest themes.

02 - Shallow Dive

By this point you should have a list of companies that fit your world view over the next 5 to 10 years. The list might be 15 long.

If we performed a deep dive on each of these, you would never invest in anything - you’d be still analysing your list after the narrative has played out. So I perform a shallow dive: a top level look at the companies to try and prioritise a few to investigate further (around 5).

Tools here are as follows:

• LLM (Claude, ChatGPT, Gemini)

• Substack

• Web based research

This isn’t a checklist phase - it’s a feel phase. Read a couple of Substack pieces on the company. Skim a recent earnings call. Pull up a five-year price chart. Check recent news.

Start with what you think is the best business idea, the one that resonates with you most, and begin discounting your way through the list when others don’t.

This phase is about finding companies you can hold onto when the going gets tough - they need your full confidence in the overall outlook.

Read substack articles written by other people covering the businesses as a first port of call, and then check their work with an LLM. Check for recent news articles or analysis, synthesise the information with an LLM to ensure you understand it.

Get a feel for which companies you want to go further with, and ensure it isn’t every business on the list - take your best ideas for the next phase.

At the end of Shallow Dive you should have: around 5 companies you’d genuinely consider owning.

03 - Monitor

This is about finalising your shallow dive, and beginning the process of really getting to know each business.

Watchlists - TradingView or Yahoo Finance

You’ll need a watchlist to track these. Observe the price movements, see how it reacts to news, learn more about each business.

I use TradingView, and I have a MONITOR heading to track these businesses at all times.

You can use yahoo finance too, which I’ve used in the past, and there’s many other watchlists you can use. If you’re a paid Rootfolio subscriber, the screener does this job too - and tracks your conviction score against price movement, which TradingView won’t do.

This watchlist should be aggressively curated. If you learn something new about a business you don’t like, take it out of monitoring, go back to the shallow dive and find a new one. The watchlist isn’t a graveyard - names that have sat there for six months without earning a deeper look should be cut.

Google Alerts

Google alerts work well too. Just configure them to ping to your mailbox daily and all interesting news it finds will be sent to your mailbox to read.

It’s easy to unsubscribe too.

X / Substack

Keep following the commentary for these businesses.

Stress test your initial hypotheses, and try learn from other investor’s analysis. X is a great place for this to do when you get spare time, substack is more of a desktop process for deep focus.

At the end of Monitor you should have: a live shortlist where you understand how each name moves and reacts.

04 - Deep Dive

Now its time to get to the next level.

For businesses getting to this point, the following should be true:

• You believe in the long term vision of the market the business is in

• You believe the business can capitalise on the opportunity they intend to capture

• You understand the company at a high level

I will generally start with company reports. Reading them and getting a feel is a great way to do an initial smoke test. Look at the balance sheet, look at their recent growth, look at the transcript of the earnings calls and try and see if you can see this company executing as you would expect.

NotebookLM

A nice way to do this is to dump all your sources into NotebookLM. This is an LLM that only uses the sources you give it, so it ensures total accuracy, and doesn’t hallucinate. I’ll generally dump a few earnings reports, some transcripts, and any other supporting materials in there, and then start querying my assumptions and drilling down into the financials to understand the company better.

LLM

Using an LLM to sense check your research is also useful. Ask it to stress test your hypothesis, let it analyse the balance sheet to check your assumptions, let it forecast growth based on what you’ve learnt.

The aim here should be understanding enough about the financials to begin valuing the company. You need to understand at a minimum:

1. The balance sheet

2. Expected dilution

3. Expected revenue growth

4. Expected margins

At the end of Deep Dive you should have: a thesis you could defend out loud, and the financial inputs you need for a valuation.

05 - Valuation

Until relatively recently, I was writing valuations using DCF models in excel. They took days to get right, with dozens of inputs.

Over the last 2 years however, I’ve begun using lightweight DCF tools for my initial valuation analysis. By lightweight I mean: forecast horizon, growth rate, terminal margin, exit multiple, discount rate. No line-by-line cash flow buildouts, no working capital schedules. The thesis decides the inputs, the tool does the math.

There’s quite a few out there, but I always found myself using the outputs as a starting point, then checking the results and tweaking the outcomes by hand in excel and verified with an LLM.

In the present day, verifying the numbers with an LLM is a no-brainer. It provides a more accurate result and required a fraction of the time required - but you need a starting point.

I decided to build my own valuation tool on rootfolio.ai - paid subscribers can use it for free, others can either sign up or find something similar online.

The example below uses Apple: AAPL.

First taking a look at historical performance to understand what’s in the realm of possibility, and what sort of base are we starting from.

You should already have studied the balance sheet, and have a good idea of the financials of the company, but if you want to study further, these are all available in the tool.

Then we’ll start building out some assumptions about the company. Rootfolio uses averages over the projection period.

You can come up with bear, base and bull scenarios to see how different valuations look.

Notice the warnings underneath the inputs - “Growth assumption (10%) exceeds most recent annual growth (6.4%)”, “Profit margin (28%) exceeds highest recorded margin (26.9%)”, “Terminal value accounts for 78% of total DCF”. The tool flags when an assumption is more aggressive than recent history, or when the model is leaning too hard on terminal value to do the heavy lifting. That’s the part that catches you when you’re being too optimistic.

From there you get a rough idea of how everything lines up.

You have your market and business thesis, you’ve done some due diligence on the company, you’ve learnt about the company, you have a rough idea on how this translates to financial numbers, and now you have a present value.

In the example above, we find that with our assumptions for Apple, on a free cash flow basis, there is a big enough margin of safety for us to consider buying.

If all these valuations were a deep red, and suggested no value in the business, I might put it back into the early discovery shortlist. If the valuation is way ahead of its share price already, this might be a good point to go no further.

In this case, I’ve done my rough valuation and I now want to refine it further - checking my numbers, getting feedback through an LLM, and consulting my notes from NotebookLM.

I’ve selected ‘advanced options’ here, which allows me to add share dilution in. I can get a rough feel for whether buybacks will continue by asking my NotebookLM. If buybacks are authorised and management has signalled they will buy back shares, and they have been buying back shares, you can safely assume the trend could continue.

And as we go a level down, the valuation becomes more refined.

Screenshot it, throw it back into your LLM of choice, then validate the feedback against another LLM. Stress test it, pick holes in it. Eventually you should be close to something you have confidence in.

Manually writing out each row in excel, or some of the rows, can be helpful. But for many cases I find this approach works well.

At the end of Valuation you should have: a base case fair value you actually believe, and a bull case with a named catalyst.

06 - Decision Making

Lets be honest, this is the hard bit.

Making good decisions in this game has made people billionaires.

Finding undervalued companies is not the hard bit. Prioritising the right companies and positioning accordingly is where the real money is made.

Unfortunately I don’t have a one size fits all and prints money answer to this question, but I do have a tool and method that helps.

For nearly 5 years, I used a concept called UAV - Upside Adjusted Valuation - to prioritise businesses based on key factors: base valuation, bull valuation (upside), moat, management, narrative, balance sheet. It looked a little bit like what we have below.

This provided a composite score for each company, based on all my KPI’s for investment criteria. My composite score had specific weightings, but the concept can be tailored to your own needs.

UAV worked well for me but it lived in a spreadsheet, which meant the data was stale the moment a price moved. So I rebuilt it in Rootfolio, refined the scoring, and renamed it Root Score. Same idea - composite score, business quality plus valuation upside - just live, against current prices.

With a ranked score system, you get a clear view at any time which companies are asymmetric opportunities, and which companies show no real value in buying.

As you come up with valuations, and score the business based on your due diligence, a tool like this helps to monitor your stocks and inform your decision making.

If you don’t want to create your own excel sheet, Rootfolio’s stock screener is an option that you could consider. You can build your own screener too — just select a stock, add your criteria, and off you go.

At the end of Decision Making you should have: a ranked list of names, sized to conviction, with the discipline to do nothing on the ones that don’t qualify.

Bringing it together

The 45%+ in my publications, and the higher number in my actual portfolio, didn’t come from being clever. It came from process.

The point of every stage above is to earn the right to do nothing - to hold names you have conviction in through the volatility, and to skip the ones you haven’t earned conviction on. Most people lose money in the markets because they’re working backwards from a stock to a thesis. The funnel above works forwards: theme, candidates, shortlist, thesis, valuation, position.

You don’t need Rootfolio to do this. You can run the whole process with a notebook, TradingView, an LLM and a spreadsheet. I did it that way for years.

But if any of the stages above have felt like the one where you keep getting stuck - for me it was always the valuation step, and then the decision step - that’s why I built Rootfolio. It compresses the parts of the process that are repetitive and error-prone into something live and consistent. Paid subscribers get full access; everyone else can sign up at rootfolio.ai and try it.

If you’ve got a better process, I’d love to hear it. Reply to this email, or message me on substack.

I’ve made two changes worth flagging. First, the About page is updated – it now properly introduces who I am and what Root Logic is. You can read it here. Second, paid subscribers now get full access to Rootfolio – the platform I built to run my own stockpicking process, including the Root scoring system, a live valuation tool, portfolio tracking, and a few other things. If you’re already a paid subscriber and want to claim your Rootfolio access, email me at substack@harrydoyle.uk with the Google account you’d like to use.

Right – on to Rocket Lab.

At today’s $78, Rocket Lab fails every reasonable valuation test I can run. The standard Bull case earns me my discount rate, nothing more. The Vision case requires the company to become the Nvidia of space. And yet I bought more last week. Here’s why that isn’t crazy.

Aswath Damodaran has owned Nvidia since 2018. He’s also sold parts of it twice – once in early 2024 around $80, again later when his updated valuation still came in well below market price. Each time, the stock kept running. By late 2025 it crossed $5 trillion in market cap, and Damodaran was still calling it overvalued. His model wasn’t wrong – the assumptions were defensible, the maths was honest – but he continually called the market’s bluff that the TAM wouldn’t increase, and it just keeps going.

He’s open about leaving money on the table. He’s also clear about what he calls the Big Market Delusion: the tendency for investors in genuinely large markets to systematically overpay, because each one believes they’ve picked the winner. Most participants in big-market delusions lose. Nvidia is the exception, not the rule.

I’m telling you this upfront because I own Rocket Lab. It’s currently about 10% of my portfolio – bigger than I’d normally size a venture-style position. I want to walk you through how I justify that with a clear conscience, given what Damodaran’s framework actually says. He’s somebody I greatly respect – which is why disagreeing with him here is worth doing carefully.

This isn’t a license to ignore valuation. It’s a recognition that valuation alone can’t see bottlenecks before they’re priced in. The Big Market Delusion punishes investors who pay up for indistinguishable players in big markets. It can also reward the rare ones who pay up for the bottleneck. That’s the question this article is really asking: is Rocket Lab the bottleneck, and is the bottleneck worth what they’re charging for it?

So this is me showing my working. If you read to the end and disagree, the position size on my side should give you pause – but the maths I’m about to walk through should give me pause too. That’s the honest tension at the heart of any vision bet.

The 2035 setup

The McKinsey/World Economic Forum number you’ve probably seen is $1.8 trillion by 2035. That’s the entire space economy. It’s also misleading if you’re using it to size Rocket Lab’s opportunity.

Inside that $1.8 trillion, McKinsey splits the world into two pieces:

• Backbone applications ($755 billion by 2035) – satellites, launch vehicles, components, GPS infrastructure. The hardware and the services that ride on it.

• Reach applications (~$1.05 trillion by 2035) – Uber, weather services, food delivery, parcel tracking. Industries enabled by space tech but not part of it.

Rocket Lab competes in the backbone. Uber doesn’t pay them. Neither does DoorDash. So when you see articles applying capture rates to the headline $1.8 trillion, you’re looking at a sleight of hand. The honest TAM for what Rocket Lab does is $755 billion – still enormous, still the right kind of secular tailwind, but materially smaller than the marketing version.

Within that backbone, four pillars matter for the 2035 thesis:

Megaconstellations. Networks of hundreds or thousands of satellites in low Earth orbit working as a single system – Starlink is the obvious example. Amazon’s Kuiper is deploying. China has multiple state-backed equivalents launching. Western governments are building their own military versions for surveillance and communications. Each constellation needs thousands of satellites built, launched, and refreshed every few years as they fall back to Earth. This is the meat and potatoes – high-cadence, high-volume manufacturing and launch. Rocket Lab’s Neutron rocket and Space Systems division are aimed squarely at this market.

Defence and national security. The Space Development Agency is the US military’s recently-created arm for procuring satellite networks. Rocket Lab’s $816 million SDA prime contract is the proof point that they’ve crossed from “small launch novelty” into the kind of customer list historically reserved for Lockheed and Boeing. Behind this sits a structural tailwind: Western governments increasingly do not want SpaceX as their only option. Sovereign launch redundancy is a real budget item now.

Space-based solar power. Speculative but real. Meta announced a partnership with Overview Energy on 27 April for up to 1 GW of orbital solar capacity, with commercial delivery targeted for 2030. (Worth flagging: this is Meta–Overview, not Meta–Rocket Lab. Some commentary has conflated the two. RKLB is at most an indirect beneficiary as a potential launch and satellite supplier.) Commercial deployment at meaningful scale by 2035 is a possibility, not a base case.

Orbital data centres, lunar economy, asteroid mining. This is where the science-fiction stuff lives. Hyperscalers have filed for plans to put compute in orbit. Nvidia announced its Vera Rubin Space-1 system. None of this is at commercial scale, none of it is profitable. By 2035? Possible at the margins. By 2045? Quite likely. This is where Rocket Lab’s optionality lives – not its base case.

The honest framing: the first two are happening regardless. The third is plausible. The fourth is the decade-long lottery ticket.

They own the bottleneck. (Or part of it).

Here’s where I think the consensus actually has it right. There is no credible alternative to SpaceX in the medium-lift, fully-integrated space services category. Blue Origin is a decade behind on launch cadence. Stoke and Relativity are pre-revenue at scale. Chinese players don’t sell to Western governments. ULA is a legacy cost structure dressed up as a competitor.

But “credible #2” undersells what Rocket Lab actually is. They own the bottleneck.

Every megaconstellation, every defence satellite, every orbital data centre, every space-based solar array – none of it gets to space without launch capacity. Satellite manufacturers can scale. Payload designers can scale. Software can scale. Launch cannot scale on demand. It’s vertically integrated, capital-intensive, regulator-bound, physics-bound. You can’t add a launch provider the way you can add an AWS region. And in the West, there are exactly two companies that can do it at scale by 2030: SpaceX and Rocket Lab. That’s it.

Owning the bottleneck of a growing economy is a structurally different kind of business. It’s why the freight railroads compounded for a century. It’s why TSMC commands its margins despite being one of many semiconductor businesses. It’s why Visa and Mastercard – the duopoly that owns the payment rails – trade at 30–35x earnings in steady state and have for years, even though they grow at single digits. When demand explodes through your chokepoint, you get to set terms.

Now apply this to Rocket Lab. The 2035 space economy doesn’t run without launch capacity. The West can’t politically allow SpaceX to be the only company providing it. Bookings out to 2030+ are already largely sold before they exist. The duopoly isn’t a far-off scenario – it’s already forming, and the regulatory and capital barriers to a third entrant are enormous.

That changes how you should think about the valuation, and we’ll come back to it.

What RKLB actually has, today:

• $602M in 2025 revenue, up 38% year-on-year, with 21 Electron missions flown at a 100% mission success rate

• Space Systems now generates two-thirds of revenue ($403M of $602M) – meaning the company is more a satellite manufacturer than a launch provider, which most retail investors haven’t internalised

• A $1.85 billion backlog, up 73% year-on-year

• $1.1 billion in cash and marketable securities at year-end

• An $816M SDA prime contract win, the Mynaric acquisition closed in April, and a Neutron rocket explicitly designed to launch megaconstellation payloads in bulk

This is a real business. It’s just not yet a profitable one. GAAP R&D ran $270.7M in 2025, up 55% on Neutron development. Net loss for the year was $198M, with adjusted EBITDA loss of $101M. Real revenue, real losses, real burn.

The bear case – and why it has more legs than people admit

In February 2025, Bleecker Street Research published a short report alleging Rocket Lab had “materially misled investors” about Neutron’s mid-2025 launch timeline. They predicted the rocket wouldn’t fly until mid-2026 to mid-2027.

A year later, Rocket Lab’s own guidance now puts Neutron’s first flight in Q4 2026 – directly inside Bleecker’s predicted window. The Stage 1 tank ruptured during pressure testing in January 2026 and a manufacturing defect was identified. The mid-2025 target has now slipped twice.

Bleecker’s other claims are now harder to dismiss:

• They argued Neutron would need significant additional capital to reach profitability. Rocket Lab raised over $1.3 billion via ATM equity offerings across 2025 – the cash need is real, even if the exact figure was up for debate.

• They argued Neutron pricing was being undercut against the stated $50–55M sticker. We don’t yet have a clean way to verify or refute this.

• They argued Electron alone couldn’t justify the valuation. Hard to disagree.

A delay isn’t fatal. Many great hardware companies have shipped late. But each slip eats into the addressable SDA contract windows, the National Security Space Launch onboarding, and the ability to grab share before SpaceX fully prices Falcon 9 to suppress competition. Time isn’t free here.

Three valuations, working back from 2035

For a vision piece, I think it makes sense to anchor every scenario on the same target year: 2035. This forces honesty. You can’t hide behind a 5-year horizon for the boring case and a 10-year horizon for the exciting one. Same year, same TAM, same discount rate. The only thing that changes is what share of the market Rocket Lab captures – and how profitably.

I’m using a 12% discount rate across all three scenarios. That’s the same rate I used for Nebius. The logic: this is a venture-style bet with a decade-long horizon and binary execution risks. It deserves the higher hurdle.

Every scenario assumes meaningful dilution (40–60% over the decade) to fund Neutron production, Space Systems expansion, and likely further M&A. The current ~605M share count includes 46M preferred shares.

The three scenarios

Two things worth pausing on.

There’s no truly “conservative” scenario here. Even the bottom case requires a 44% revenue CAGR for a decade. If the McKinsey TAM is right, even modest market capture implies extraordinary growth from RKLB’s $602M base.

The Vision case requires Rocket Lab to take an enormous slice of the non-SpaceX share. Today, SpaceX is roughly 80% of the global commercial launch market. The remaining 20% is split between China, Russia, ULA, Arianespace, RKLB and a long tail. A 12% capture of the global backbone TAM in 2035 implies Rocket Lab takes more than half of all non-SpaceX activity globally. Possible, but you should sit with what that requires.

Reading the table

At today’s $78:

• The Conservative case implies you’re paying roughly 2.4x fair value. If you believe RKLB will only capture 3% of the backbone TAM, you should not own this stock at this price.

• The Bull case implies modest upside (~11%) over a decade. That’s roughly the discount rate itself – you’d be earning your required return for the risk, no more. A reasonable outcome, but not the reason you’d take a venture-style position.

• The Vision case implies ~3x upside over ten years – roughly 24% annualised in absolute terms. That’s the bet that justifies the position.

But here’s the thing I’ve been hinting at throughout: I don’t think the Vision case is actually the upper bound. The bottleneck argument I made earlier has implications for the model that the standard Vision inputs don’t capture. If you take the duopoly framing seriously, the multiple, the margin, and the capture rate could all be too low. There’s a fourth scenario – call it Mega-Vision, the bottleneck-priced-properly case – that runs the maths differently and arrives somewhere uncomfortable for anyone trying to be conservative.

That, plus the action framework for what to actually do with all this, plus RKLB’s current Root score and how it ranks against the other names in my screener, sits below the paywall.

Risks I’m watching

• Neutron slipping into 2027. Bleecker may yet be vindicated. A failed first flight would be worse.

• A meaningful capital raise. $1.1B in cash is healthy, but not infinite given the ongoing burn and Mynaric integration costs. A bad market for a secondary offering would force unfavourable terms.

• SpaceX pricing aggression. Falcon 9 rideshare already gutted Electron’s growth thesis. Falcon 9 priced to suppress Neutron is the thing that scares me most.

• Vision happens, but RKLB isn’t the winner. Stoke, Relativity, Blue Origin, Chinese players. The 2035 space economy could materialise without Rocket Lab being the #2 – and the article’s whole valuation collapses if that happens.

• The McKinsey TAM being wrong. $755B backbone by 2035 is one forecast. If it lands at $400B instead, every scenario halves.

So why am I in at 10%?

Honest answer: because I think the asymmetry is still in my favour, even after running the numbers.

A few reasons I’ve sized this larger than a typical venture position:

The downside isn’t zero. Even the Conservative case has Rocket Lab worth $32 a share. That’s a 60% drawdown from here, which is real money – but it’s not a SPAC-style wipeout. The $1.85B backlog, the SDA relationships, the Space Systems business, and the $1.1B in cash all give the company multiple years of runway even if Neutron disappoints. Real business, real revenue.

The #2 position is structural, not just current. Sovereign launch redundancy isn’t a passing concern. Governments will not consent to SpaceX as the sole provider of strategic launch capacity, and no other Western company is closer to filling that role. Standard models undershoot when a company owns the supply constraint of a generational build-out – which is why I think even my Vision case is too conservative.

The Bull case alone delivers an acceptable outcome. At $87 PV against $78 today, even being roughly right about the Bull case earns me my discount rate. I don’t need the Vision case to play out to avoid losing money – I need it to play out to make real money.

That said, I want to be clear about what 10% isn’t. It isn’t “max conviction, sized to dominate the portfolio.” My core holdings – the Games Workshops and Alphabets – are companies where I think the base case delivers strong returns. RKLB is sized large because the asymmetry is unusual, not because the base case is exceptional. If the Bull case starts pricing in the Vision case (i.e. the stock runs to $130+), I’ll trim. If Neutron fails its first flight, I’ll reassess hard.

Damodaran has owned Nvidia since 2018, even as he sold portions of it for being overvalued. He didn’t have to be right about the destination to participate in the journey. Sizing matters, but so does showing up.

The numbers say Rocket Lab is overvalued. They’re not wrong. The Bull case earns me the discount rate, the Vision case requires heroic execution, and a careful analyst applying Damodaran’s framework would tell me to wait for a better entry. If I do that, I might catch a 30% pullback. I might also miss the company that owns the chokepoint of the next generational build-out, while waiting for a price the market has no intention of giving me.

So I added on the dip below $80. Not because the maths supports it cleanly – the maths doesn’t. I added because numbers alone can’t price what Rocket Lab is becoming, and I’d rather be roughly right about the structural argument than precisely right about the wrong question.

That doesn’t make it a recommendation. It makes it a disclosed position with a worked-through thesis, an honest acknowledgment of where my own framework breaks down, and a willingness to be wrong slowly. Take it for what it is.

For now, the position stays. The May 7 earnings, the Neutron tank rebuild, and the first flight will tell me whether to add, hold, or trim.

Performance since inception

Live performance is tracked at Rootfolio – paid subscribers get full access.

Read more

That is what I want to unpack in this piece.

Contents:

1. Thesis Reminder

2. 2026 So Far

3. Checking The Thesis

4. The Four Horsemen Valuation

5. Updated Base Case

6. Updated Bull Case

7. Portfolio Impact

Thesis Reminder

When I first wrote about Nebius, the investment case was still largely conceptual, but the core idea was simple:

• Nebius was building an AI-native infrastructure and platform business

• demand for AI compute and related tooling would expand dramatically

• the team had the operational pedigree to execute

• and on a sum-of-the-parts basis, the valuation looked too low

That was already an attractive setup: a company exposed to the fastest-growing part of technology, led by a management team I trusted, trading at what looked like a meaningful discount to intrinsic value.

But at that stage, much of the thesis was still unproven. We had not yet seen what execution would look like in practice, how quickly demand would convert into real contracts, or whether management could scale the business while preserving discipline.

In other words, the thesis was compelling, but still largely theoretical.

That has changed.

2026 So Far

Let me walk through the key items that matter for the thesis, with the most recent item being the most interesting to me.

Mar 3: Independence AI Factory - 1.2GW Hyperscale Facility

What happened: Independence, Missouri approved a 1.2GW AI factory on ~400 acres. This is hyperscale tier - comparable to the largest cloud facilities globally. First of its kind at this scale in the US for a non-hyperscaler. PILOT payments exceed $650M over 20 years, structured to avoid residential rate increases.

Why it matters: This single site represents 40% of the >3GW 2026 target. More importantly, Arkady said: “This is our first project of this scale, but not the last.” They’re not building small. They’re going directly to hyperscale infrastructure.

Thesis impact: Validates point #2 (demand durability). You don’t get a 1.2GW facility approved for speculative demand - this is backed by customer commitments.

Mar 16: Meta Deal #2 - $27B Over 5 Years

What happened: $12B firm commitment, up to $15B optional capacity. NVIDIA Vera Rubin GPUs, delivery starts early 2027. Combined with the November 2025 $3B deal (delivered February 2026), Meta’s total contracted spend is $30B.

Why it matters: This is the largest AI infrastructure contract announced to date. The $15B optional capacity is the key - Meta has a backstop, but Nebius can sell that capacity to third parties if demand materialises. This is smart risk allocation.

The $3B original deal is already generating revenue (delivered Feb 2026, currently in servicing stage). The new $27B deal validates that Meta didn’t just test Nebius - they’re betting on them for the next five years.

Thesis impact: Validates point #2 (demand). $30B from one customer through 2031 is not speculative demand.

Mar 17-20: Convertible Note Offering - $4.6B Raised

What happened: $2.59B in 2031 notes at 1.250%, $1.75B in 2033 notes at 2.625%. Total raised: $4.6B. Conversion prices: $180-183.

Why it matters: This is absurdly cheap capital. Compare 1.25-2.625% to a 9% equity cost: they’re saving $260-300M per year in financing costs. The notes convert when the stock performs (my base case has it at $263 by 2028), so this is effectively delayed equity at much better terms than issuing stock today.

But - and here’s the trade-off - those notes convert to 25.3M shares. That’s the dilution that drops my valuation from $210 to $194.

Thesis impact: A mixed bag. They chose the cheapest capital available. But it adds shareholder dilution. We’ll speak more about the impact of this later.

Checking The Thesis

I touched on the thesis from my original article, it was raw, conceptual and somewhat speculative.

A year later, much of that top-level thesis uncertainty has been quelled. In its place, the key risks are now more specific: execution against an enormous buildout, customer concentration, financing structure, and the extent to which Nebius can turn infrastructure scale into platform economics.

1. AI-native infrastructure stack: VALIDATED

June 2025: Theoretical claim about building across the full stack.

March 2026: NVIDIA partnership with co-development. Physical AI platform with validated customers. CrowdStrike integration for enterprise security. Token Factory for inference. Tavily for agentic search.

This isn’t GPU rental. This is a full platform with vertical-specific solutions and switching costs.

2. Explosive demand for 5+ years: VALIDATED

June 2025: The question was whether AI compute demand would materialise at scale and whether customers would pay Nebius prices.

March 2026: $49.4B in contracted revenue through 2031. That’s $19.4B Microsoft, $3.0B Meta original (already generating revenue), and $27B Meta new (firm + optional). The Independence AI Factory (1.2GW) got approved because demand is locked in.

78-79% of 2026-2027 revenue is already contracted (conservative view, firm deals only). 96% coverage in 2027 if Meta exercises partial optional capacity.

Demand isn’t speculative anymore. It’s contracted.

3. Team execution capability: VALIDATED

June 2025: Known based on Yandex track record.

March 2026: They delivered seven major developments in five weeks. Meta $3B deal delivered on time in Feb 2026. Microsoft first tranche delivered on schedule in November 2025. Everything built is immediately sold. Capacity sold out in Q3, Q4, and already sold out in Q1 2026.

Execution risk has moved from “can they deliver?” to “can they build fast enough?”

4. Financial discipline: ONE YELLOW FLAG

June 2025: Partial validation based on Yandex legacy.

March 2026: Margins improving (19% to 24% in core AI cloud). Positive adjusted EBITDA in Q4. They chose convertible notes at 1.25-2.625% instead of equity at 9% cost.

But - the convertibles add 25.3M shares. Management said they’re “mindful about shareholder dilution,” and the ATM programme is unused with no plans to use it. They prefer debt and asset-backed financing.

This is smart financing that creates dilution. That’s the trade-off. I’m flagging it yellow, not red, because the capital is being deployed into capacity that generates contracted revenue.

5. Cheap on all valuation metrics: STILL VALIDATED

June 2025: $15.4B sum of parts vs $12B market cap.

March 2026: At $117 current price, base case shows $194 fair value (+81% upside). Sum of parts is now higher (ClickHouse at $15B valuation, Nebius AI cloud growing). Still trading below intrinsic value.

What changed: Four of five thesis points are now validated vs two in June 2025. Conviction went from 40% to 80%. But the per-share valuation dropped because of dilution from convertible notes.

That is the core paradox with Nebius today.

I am more confident in the business than I was six or nine months ago. The strategic position looks stronger. The commercial evidence is better. The execution looks more real.

But I am slightly less optimistic about the per-share math, because scaling a business of this kind requires vast amounts of capital, and some of that capital is now more likely to arrive in forms that dilute shareholders.

Conviction in the company is up, but my confidence that my shares will realise the full opportunity is slightly down. To many, including me, this is absolutely a worthwhile trade-off.

The Four Horsemen Valuation

If you’ve read my previous Nebius pieces, you know I value this business through four variables: revenue growth, depreciation as a percentage of revenue, net income margins, and share dilution. I believe these are the 4 metrics that deserve the most attention - deterioration in any of them materially impacts the valuation of the business.

Let’s walk through each one.

Revenue Growth

The $49.4B in contracted revenue tells you most of what you need to know here. Microsoft ($19.4B), Meta original ($3B delivered February 2026), and Meta new ($27B starting 2027). That’s three deals from two customers covering 43-54% of my five-year revenue model.

The conservative view - firm contracts only - shows 78-79% coverage in 2026-2027.

The optimistic view, where Meta exercises some of its $15B optional capacity, pushes 2027 to 96% coverage. Either way, the near-term revenue is locked in.

What about 2028-2030? That’s where execution matters. The conservative case needs $31.3B in incremental wins. The optimistic case needs $22.7B. Can they do it? Well, they’re already at >2GW of contracted power in February and guiding to >3GW by year-end. The Independence AI Factory alone is 1.2GW. They’re not slowing down.

Revenue growth risk has moved from “will anyone buy this?” to “can they build fast enough?” That’s a much better problem to have.

Depreciation as % of Revenue

Here’s where it gets interesting. Nebius extended their depreciation schedule from 4 years to 5 years in Q1 2026. This isn’t an accounting trick - it’s a statement about GPU economics. If you’re confident a H100 keeps earning revenue in year 5, you depreciate over 5 years. If you think it’s worthless after 4, you depreciate faster.

The market data backs them up. H100 lease rates haven’t collapsed like everyone expected. They’ve held, and in some cases increased. Inference workloads need GPUs too, and there aren’t enough to go around. So the GPUs keep earning past the initial training window.

My model has depreciation falling from 76% of revenue in 2025 to 23% by 2030. That’s not because of the accounting change - it’s because revenue grows 47x while depreciation only grows 14x. The 5-year schedule just makes the math cleaner.

Conservative adjustment: Back in February, I normalised the benefit by reducing exit multiples 5x and capping net margins to avoid double-counting. The depreciation improvement is real, but I’m not giving them credit twice.

Net Income Margins

Base case: 7.5% in 2027, ramping to 15% by 2030.
Bull case: 9.5% in 2027, ramping to 19.5% by 2030.

The bull case hits AWS-level margins from platform economics. Token Factory, Tavily, Aether - the software layer that makes customers sticky and lets you charge more per GPU hour. The base case is more conservative: they scale the infrastructure but don’t fully monetize the platform layer.

Interest expense from the convertible notes? Trivial. $68M in 2027, declining to zero as the notes convert when the stock performs. That’s 0.8% of 2027 revenue and 0.3% by 2030. It barely moves the needle.

Share Dilution

And here’s where the paradox lives.

The business got better. They likely raised $4.6B in convertible notes at 1.25-2.625% interest rates. That’s absurdly cheap capital. Compare it to equity at a 9% cost: they’re saving $260-300M per year in financing costs.

But those notes convert to 25.3M shares when the stock performs. And my model already assumed some dilution from organic growth (stock comp, misc grants). So we go from 253M shares today to 330M by 2030.

Result: Base case PV drops by $16. Not because the business got worse - it got better. But the value is now spread across 8% more shares.

The maths is straightforward: same 2030 net income ($3.75B), same multiple (27.5x), but 330M shares instead of 305M. That’s $11.36 EPS instead of $12.30. Apply the same 9% WACC discount, and you get $194 instead of $210.

This is good financing that creates shareholder dilution. Both things are true.

But my main takeaway here is that it shows that management is less averse to dilution than they recently suggested on the latest earnings call - and that’s why my 2030 valuation has dropped. We will need a huge amount of financing to to bring the CapEx commitments to life, it looks like more of that financing could be via dilution than I previously hoped.

Updated Base Case

What changed from my February valuation? Two things: share dilution and interest expense. That’s it.

Everything else - revenue path, depreciation schedule, margin assumptions, exit multiple - stayed exactly the same. The business didn’t change. My view on the opportunity didn’t change. The only difference is how many shareholders are splitting the pie.

What Stayed the Same

Revenue projections: unchanged. Still $3.2B in 2026 ramping to $25B by 2030.

Depreciation schedule: unchanged. Still 53% in 2027 falling to 23% by 2030.

Margin trajectory: unchanged. Still 7.5% net margins in 2027 ramping to 15% by 2030.

Exit multiple: unchanged. Still 27.5x P/E.

The business model didn’t change. The capacity deployment path didn’t change. My view on GPU economics didn’t change. This is still the same conservative case I built in February - Microsoft and Meta contracts plus modest incremental wins, normalised depreciation, 15% terminal margins.

What Changed

Share count: 305M → 330M (+25.3M from convertibles)

The convertible note offering dilutes you. The notes convert when the stock performs - conversion prices are $180-183. My base case has the stock at $263 by 2028, so the notes convert mid-year and those 25.3M shares are in the count by 2030.

This was the right financing decision. Raising $4.6B at 1.25-2.625% interest beats diluting equity at 9% cost. But it still dilutes you.

Interest expense: Small drag, declining fast

• 2027: -$68M (notes outstanding all year)

• 2028: -$34M (notes convert mid-year at $263)

• 2029-2030: $0 (fully converted to equity)

The interest expense is immaterial. It’s 0.27% of 2030 revenue. It rounds to nothing.

The Maths

February base case:

• 2030: $3,750M NI, 305M shares, $12.30 EPS

• Exit: $12.30 × 27.5x = $338

• Fair value today: $210

March base case:

• 2030: $3,750M NI, 330M shares, $11.36 EPS

• Exit: $11.36 × 27.5x = $312

• Fair value today: $194

Impact: -$16 per share (-7.6%)

The entire move is dilution. The business is worth more in absolute terms - they raised capital at cheap rates to fund contracted capacity. But that value gets spread across 8% more shares.

Current price: $117.62
Base case fair value: $194
Upside: +65%

That’s the paradox. Seven major positive developments in Q1. Higher conviction. Lower per-share value. You’re buying into a bigger business with more shareholders.

Updated Bull Case

A few major changes here. Expected dilution is up, P/E compression is faster. Let me explain…

On Multiples and Honesty

Let me be direct: 42.5x was too giddy.

When I wrote the original bull case, I was looking at a business growing 250% year-on-year with platform economics and thinking “this deserves a premium multiple.” Fair. But 42.5x in 2030 assumes the market is still pricing Nebius like a hypergrowth startup five years from now. That’s wishful thinking.

If we get to 2030 and revenue is still climbing at 45% while margins are expanding, we might command a 40x P/E. Maybe. But that multiple needs to be earned, not assumed. And the path from here to there involves a lot of capex, a lot of execution risk, and a lot of dilution.

So I’ve dropped the exit multiple to 35x. This still implies a premium business - AWS trades 25–30x, Snowflake compressed from 50x peaks to 30–40x as it matured. At 35x, you’re paying for quality but not for perfection. And critically, the bull case still has 2030 revenue growing at +45%, which justifies the premium over the base case 27.5x.

5 GW Is No Longer Speculation

In February, my bull case assumed 4,000 MW. At the time, management was hinting at “>3GW contracted power” and I added 1GW of headroom. It was speculative.

That changed on March 10th. NVIDIA invested $2B for an 8.3% stake in Nebius and the two companies announced an explicit partnership to deploy more than 5 GW of AI capacity by end of 2030. Six days later, Meta signed a $27B deal - $12B in dedicated Vera Rubin capacity starting early 2027, plus $15B in backstop purchases over five years.

The bull case is no longer “what if they build more?” It’s “can they execute on what they’ve already committed to?” That’s a fundamentally different risk profile.

So the bull case is now 5,000 MW. And this is where the revenue question gets interesting.

The Capex Reality

Here’s where I got it wrong in February.

I modelled the old bull case with $53B in total capex to deploy 4,000 MW. That’s $13.25M per MW. But my base case uses $16.84M per MW for 2,500 MW deployment. Why would the bull case be cheaper per megawatt? It shouldn’t.

When you deploy more aggressively, costs don’t fall - if anything, they rise. You’re competing for power contracts, construction crews, GPU allocations, and cooling infrastructure in a market where every hyperscaler is building simultaneously.

The corrected bull case uses consistent economics: $80.1B total capex for 5,000 MW, or $16.02M per MW. That’s roughly in line with the base case per-MW cost - which is the intellectually honest answer. You don’t get a volume discount on data centres.

Revenue Per MW: Where the Bull Case Lives

In my base case, I use $10M per MW. That’s the blended rate for a customer mix dominated by Microsoft and Meta - large hyperscaler contracts at bulk pricing. At 170MW today, Nebius is generating $7.35M/MW. As utilisation matures, that trends toward $10M. Nothing controversial.

But here’s the thing: not all megawatts are priced the same.

Hyperscaler deals - Microsoft, Meta - are high-volume, long-term contracts. They’re sold at bulk rates, probably $8–9M per MW. These customers are essential: they provide revenue visibility, de-risk the capex, and establish credibility. But they’re not the highest-margin business.

Cloud and enterprise customers - AI startups scaling from hundreds to tens of thousands of GPUs, enterprises deploying inference workloads - pay more per GPU hour. They use the full software stack (Token Factory, Tavily, Aether), they value flexibility over volume discounts, and they generate $12–13M per MW. On the Q4 call, management was explicit: they’re focusing on customers with “premium workloads and use cases” and will remain “opportunistic” with large hyperscaler deals.

The base case at 2.5GW is mostly hyperscaler - Microsoft and Meta contracts cover the majority of that capacity. Blended rate: $10M/MW. Revenue: $25B.

The bull case adds 2.5GW on top. If that incremental capacity skews toward cloud/enterprise customers at premium pricing, the blended rate across 5GW rises to $11M/MW. Revenue: $55B.

This is the key distinction. The margin premium in the bull case doesn’t come from cheaper capex - per-MW costs are identical. It comes from higher revenue per MW, driven by customer mix.

The Margin Mechanics

This is important, so let me spell it out.

Capex per MW is $16M in both cases. Depreciation per MW is the same in both cases - you’re depreciating the same assets at the same rate. But depreciation as a percentage of revenue is lower in the bull case, because you’re generating more revenue per MW.

Base case 2030: $5,700M depreciation / $25,000M revenue = 23% Bull case 2030: $11,000M depreciation / $55,000M revenue = 20%

That 3-point improvement in depreciation intensity drops straight through to margins. Add a modest operating leverage benefit - SGA and R&D don’t double when revenue doubles - and you get a 2.5-point net margin premium: 17.5% vs 15%.

17.5% net margins by 2030 is not an act of optimism. It’s the mathematical consequence of generating $11M per MW instead of $10M, while depreciating the same $16M in capex. The higher the revenue quality, the faster depreciation fades as a drag on profitability.

The sensitivity is clear: if the incremental 2.5GW goes to more hyperscaler deals at $10M/MW instead of cloud/enterprise at $12–13M/MW, revenue stays at $50B, margins compress to roughly 15.5%, and PV falls from $628 to about $506. Still a 4.3x return. Both paths work - the question is whether Nebius can attract premium customers at scale.

The Dilution Reality

The convertible notes add 25.3M shares by 2030. That’s the trade-off for raising $4.6B at 1.25–2.625% interest rates instead of diluting equity today at 9% cost. It’s smart capital allocation, but it still dilutes you.

If the bull case plays out and we actually hit $55B revenue in 2030, we got there by deploying 5,000 MW of capacity. That requires $80B in capex. The $4.6B in convertibles is just the start. More debt, potentially more equity, and more shares will be needed to fund the buildout.

The bull case share count rises to 333M by 2030 - 32% dilution from 253M today. The question is whether per-share value creation outpaces the dilution. At current prices, it still does - comfortably. But let’s not pretend we’re getting there without giving up equity.

The Revised Maths

February bull case:

• 2030: $7,800M NI, 308M shares, $25.32 EPS

• Exit: $25.32 × 42.5x = $1,076

• Fair value today: $668

March bull case (5 GW, $11M/MW, corrected capex):

• 2030: $9,625M NI, 333M shares, $28.90 EPS

• Exit: $28.90 × 35x = $1,012

• Fair value today: $628

Impact: -$40 per share (-6%)

What’s remarkable is how close the headline numbers are despite completely different mechanics underneath. The old model got to $668 through undercooked capex and an inflated multiple. The new model gets to $628 through more net income ($9,625M vs $7,800M) offset by more shares (333M vs 308M) and a lower multiple (35x vs 42.5x).

Three forces push the valuation down:

• Multiple compression: -$207/share from 35x vs 42.5x

• Dilution: -$47/share from 333M vs 308M shares

Two forces push it up:

• Scale: +$133/share from 5GW vs 4GW

• Revenue quality: +$81/share from $11M/MW vs $10M/MW

Net result: -$40. The model is smaller in implied share price, but far more defensible. Every assumption can be traced to a specific mechanism.

What This Means

Current price: $117.62 Base case fair value: $194 (+65% upside) Bull case fair value: $628 (+434% upside)

The bull case still offers a 5.3x return over five years. The path is clearer than before: deploy 5GW (NVIDIA partnership), fill it with premium customers (management’s stated strategy), and let the revenue quality drive margin expansion through lower depreciation intensity.

This is not “AWS-level margins at scale.” It’s “slightly better revenue per MW from customer mix, with depreciation math that works in your favour.” The platform layer - Token Factory, Tavily, Aether - is the mechanism that makes cloud customers pay $12–13M/MW instead of the $8–9M hyperscalers pay. If that mechanism fails, you still get a 4.3x return at $10M/MW. If it works, you get 5.3x.

If you need heroic margin assumptions and undercooked capex to justify buying at $117, you’re not buying a bull case - you’re buying a prayer. At $11M/MW with honest capex and a 35x exit, the bull case still offers 5.3x. That’s compelling without being delusional.

Portfolio Impact

Portfolio and Stock Screener

While the valuation base case has decreased, my conviction in the company has increased. It is currently the top stock in my screener on rootfolio.ai

I still have huge belief in the asymmetric upside opportunity here. If Nebius climbs very fast very quickly, I may be forced to rebalance, but we’ll cross that bridge when we get there. For now I’m holding my shares.

Analysis Returns

Despite a tumultuous start to the year from the markets, my published analysis is averaging a return of nearly 9%.

To support my research, be informed of my Nebius updates, or get special access to rootfolio.ai, consider subscribing below.

The information provided in this article is for informational and educational purposes only and does not constitute financial, investment, or professional advice. The author is not a licensed financial advisor, and the content is based on publicly available data, personal analysis, and opinions.

1. Thesis Reminder

2. Introduction
   Whats happened since our last analysis

3. The Four Horseman of the Nebius Apocalypse

4. Updating The Base Case (2026–2030)
   4.1 Revenue Growth
   4.2 Depreciation as a Percentage of Revenue
   4.3 Net Income Margin
   4.4 Shares Outstanding
   4.5 Valuation
   4.6 Sum of Parts

5. The Bull Case
   What happens if Nebius executes at the upper end of its potential

6. Risk–Reward Summary

1. Thesis Reminder

For those new to here, here’s the thesis for Nebius in 60 Seconds…

• AI compute demand will continue to increase - and someone has to run the infrastructure. Inference workloads are scaling faster than training ever did, and the market is moving from one-time model builds to continuous, always-on compute.

• Nebius is positioned to capture a huge portion of this demand - they’re already sold out through Q1 2026, have contracted >2GW of power, and are raising that target to >3GW by year-end. Everything they build, they sell.

• The GPU infrastructure is the entry point, but the full-stack software layer is where the real economics live - 100% of AI cloud customers use Nebius’s software stack. That’s not just compute revenue, that’s platform revenue with better margins and genuine stickiness. Customers come for the GPUs, stay for Aether, Token Factory, and now Tavily. This is how you build a moat.

• Management has the experience to execute at hyperscale - Arkady Volozh and team spent decades building Yandex’s data centre operations across Russia and Europe. They know how to deploy infrastructure at speed, and they’ve done it before.

• Among the neo-clouds, Nebius is best-in-class - and I believe it’s the best option on the market to capture this opportunity. At $90, the market is pricing in execution risk and GPU obsolescence. I’m pricing in a platform business with 20-30% EBIT margins by 2030.

And no, I don’t think AI is a bubble… (see below)

2. Introduction: What’s Happened Since November

In November, I published my Four Horsemen framework for Nebius - a threat-model approach to understanding whether this business could actually scale. At the time, Nebius was trading around $94, the stock had taken a beating after announcing a 25m share ATM programme, and the market seemed sceptical that GPU infrastructure could ever become a real business.

Three months later, quite a bit has happened.

• Q4 earnings validated key elements of the thesis. Revenue came in at $530m for FY 2025 (ahead of the $525m target), active power hit 170 MW (70% above the 100 MW milestone), and - most importantly - Nebius turned Adj EBITDA positive in Q4 at $15m. The company also raised 2026 guidance to $3.0-3.4B revenue with a $7-9B ARR exit. All four of my horsemen passed.

• The market has turned decidedly sour on AI infrastructure. Most stocks in the space are sitting well below their highs. Software stocks sold off on fears that AI would commoditise everything, which makes the AI infrastructure sell-off even more peculiar. If AI is going to destroy software margins, someone still has to run the compute. In my opinion that will be Nebius, but not many others seemingly agree with me…

• Nebius acquired Tavily - an agentic search provider. This is a strong signal and affirmation that Nebius isn’t just renting GPUs, they’re building the full-stack platform for agentic AI. Every AI cloud customer uses their software stack (100% attach rate), and they’re now layering in the enablement tools that turn raw compute into usable infrastructure.

• Old GPUs are going up in price, not down. H100 lease rates are holding or increasing. Even Hopper-generation GPUs are renewing at 12+ months with pricing “nudging up” (management’s words). The bear case on GPU depreciation - that older hardware becomes worthless - simply hasn’t materialised. Inference demand is real, and it tolerates older hardware far better than training.

• Finally, the market is pricing Nebius as if the AI business is worth just $18B - about 6x 2026 revenue. After accounting for ClickHouse, and other assets, the market is valuing the core AI infrastructure business at a massive discount on price to sales metrics.

That’s the setup - it has been an incredibly bumpy ride, but I’m sure many have lost faith due to the emotional toil - but I’m here to cut through the noise with some data updates following Q4 results.

3. The Four Horsemen: A Quick Recap

In November, I identified four variables that would determine whether Nebius could scale profitably:

1. Revenue Growth - Can they bring 2.5GW+ online and convert it to $10m per MW?

2. Depreciation as % of Revenue - Will GPU depreciation crush margins or fall as revenue scales?

3. Net Income Margin - Can they turn profitable by 2027 and reach 12-15% margins by 2030?

4. Shares Outstanding - Will dilution destroy returns or remain controlled?

If any one of these broke, the valuation would collapse. If all four held, the upside was significant.

Three months later, and after Q4 earnings, here’s my scorecard:

All four horsemen passed. Depreciation we will need to talk about; but revenue beat, EBITDA margins are looking great, and dilution was 24 million shares better than I’d modelled.

I believe the earnings for Nebius were stellar, but thats not a particularly interesting take - what’s more interesting is an objective data driven approach that you can also critique, to see if the numbers align with the narrative.

These four horseman are in my opinion the most fundamental metrics to monitor for a valuation of Nebius - now let’s update the framework and see how the company is performing.

4. Updating The Base Case (2026-2030)

The Base Case is not a prediction. It’s a benchmark for reasonable execution - what Nebius can deliver if they hit the majority of their milestones and avoid self-inflicted wounds.

In November, I modelled:

• $525M → $25B by 2030

• 12.5% net margins

• $203 PV (discounted at 10%)

After Q4, the numbers have changed. Not dramatically, but meaningfully. Let me walk through each horseman.

4.1 Base Case: Revenue Growth

November Assumption:
Revenue scales from $525M (2025) to $6B (2026) to $25B (2030), driven by 2.5GW of active power at $10m per MW.

What Changed:
Nebius delivered $530M in 2025. But 2026 guidance came in at $3.0-3.4B revenue - not the $4-6B I’d modelled. At first glance, this looks like a miss.

But… I think the guidance reflects timing, not demand. Management explicitly stated that capacity deployment is heavily back-loaded to H2 2026. They’re exiting the year at $7-9B ARR (midpoint $8B), but full-year revenue is only $3.2B because most of that capacity comes online in the second half.

This is actually bullish for 2027. If they exit 2026 at $8B ARR, that means they’re generating roughly $700m per month in December. A full year at that run rate is $8-9B+ in 2027 revenue. The revenue didn’t disappear - it shifted right by six months.

The revenue curve is identical to November - just front-loaded by one year due to 2026 capacity timing. The $10m per MW thesis remains intact.

• Current state: 170 MW generating $1.25B ARR = $7.35m per MW.

• As utilisation matures, this ramps toward the $10m target by 2028-2029.

Why 2,500 MW and not 3,000 MW? Because I’m being conservative. Management has secured >2GW of contracted power (February 2026) and raised the target to >3GW by year-end 2026. But I’m assuming they don’t get all 3GW online by 2030. They hit 2.5GW. If they exceed this, the bull case captures it.

The Tavily acquisition reinforces this trajectory. Nebius isn’t just building capacity - they’re building the platform layer that makes that capacity stickier and more valuable. Every customer using Tavily for search, Token Factory for inference, and Aether for orchestration is a customer that’s harder to displace. This is how you maintain $10m per MW pricing even as competitors bring capacity online.

The software attach rate is 100%. Yes, they literally confirmed this on the call - despite many suggesting that nobody would use Nebius’s software.

Mark Boroditsky (Chief Revenue Officer):

"At this stage, 100% of our AI cloud customers are utilizing our AI cloud software, obviously. So we have a 100% attach rate."

Every single AI cloud customer uses the Nebius stack. Remember, this platform revenue has much better economics via strong margins.

4.2 Base Case: Depreciation as a Percentage of Revenue

November Assumption:
Depreciation falls from 66% (2025) to 35% (2026) to 25% (2030) as revenue scales faster than depreciation expense.

This is achieved despite management choosing a conservative 4 year depreciation cycle rather than 5 or 6 years.

What Changed:
FY 2025 depreciation came in at 76% - higher than my 66% assumption. That’s the bad news, but there are few caveats.

First, depreciation improved sequentially every quarter: Q2 at 95%, Q3 at 79%, Q4 at 76%. The trend is moving in the right direction.

Second, and importantly, management extended the depreciation schedule from four years to five.

• This is a bit of financial engineering in my opinion, and it makes profit margins look better in the near term. I am personally not a fan of this move.

• But, It reflects the fact that GPUs are continuing to generate revenue beyond year four, particularly in inference workloads. The economic life of the asset genuinely extends beyond the four-year window so I can see where they are coming from, I just prefer the conservative approach.

While I prefer a more conservative approach, this change does validate the entire GPU pricing thesis. If GPUs were becoming worthless after year four, management wouldn’t extend the schedule - they’d be forced to write them down faster. The fact that they’re moving to five years does reflect that old GPUs are holding value.

In terms of how I’m modelling depreciation: if I modelled this change mechanically, depreciation would fall and margins would increase artificially. Instead, I’m haircutting the benefit by normalising depreciation down an extra 5 percentage points alongside margins but reducing exit P/E multiples by 5x.

Why? Because I don’t want the valuation to rely on an accounting change. The margin improvement is real, but I’m building in conservatism.

I’m forecasting 23% depreciation as a percentage of revenue for 2030 in my base case now - down from 28% last November.

4.3 Base Case: Net Income Margin

November Assumption:
Net margins go from -70% (2025) to 5% (2027) to 12.5% (2030).

What Changed:
Nebius turned Adj EBITDA positive in Q4 2025 at $15M. Core AI EBITDA margins hit 24% in Q4 (up from 10% in Q2 and 19% in Q3). Profitability arrived a year ahead of schedule.

This doesn’t mean 2026 will be profitable - it won’t be. Management guided to 40% Adj EBITDA margins in 2026, but EBIT remains negative due to depreciation and R&D. However, the trajectory is clear: revenue is scaling faster than costs, and operating leverage is kicking in.

I’m keeping the November margin path but adjusting slightly upward to reflect the five-year depreciation benefit (balanced with a haircut on my exit P/E for conservatism).

Why 15% in 2030 and not higher? Because I want the base case to be defensible. AWS operates at higher margins, but they’ve been at scale for a decade. Nebius reaching 15% by 2030 would already represent exceptional execution for a five-year-old infrastructure business.

4.4 Base Case: Shares Outstanding

November Assumption:
Share count moves from 277M (2025) to 335M (2030), assuming modest dilution through 2027 and then operating cash flow covers capex alongside more debt.

What Changed:
This is where the biggest positive surprise hit. Actual shares outstanding in Q4 2025: 253M. I’d assumed 277M. That’s 24 million fewer shares - about 9% better than expected.

Even better: zero shares were issued from the ATM programme. The market panicked in November when Nebius announced a 25M share ATM. I modelled conservatively and assumed they’d use it. They didn’t. Q4 operating cash flow came in at +$834M, and the company has $3.7B of cash on the balance sheet. They don’t need equity right now, and they have access to debt.

This is a massive change. Dilution was the silent killer in my November framework. If Nebius kept issuing equity through 2028-2030, returns would be destroyed. Instead, they’ve demonstrated capital discipline. They’re funding the $16-20B 2026 capex plan with cash, operating cash flow, deferred revenue ($1.58B), and structured financing - not equity. We’ll need to hope the financing is on palatable terms.

On the earnings call, management was clear that they will need to raise funds to structure this CapEx, but they communicated that issuing shares is not their preference.

While I think this is an overall good thing - I do think this will result in elevated debt levels for Nebius, and to model that risk I am going to drop my end-state P/E by 2.5 for both scenarios. The capital has to come from somewhere, and this multiple can expand in time if the debt is structured well - issued shares are more difficult to reel back.

The threat of serial dilution has been defused, for now. Albeit I’ll be keeping a very close eye, and we should still expect some.

4.5 Base Case: Valuation

Let me put the updated framework together:

2030 Implied Price: $338
PV @ 10% WACC: $210
Upside vs $89.73: +134%

The market is pricing Nebius at $89.73. My base case suggests fair value is $210. That’s a 134% gap.

What changed vs November?

In November, my base case PV was $203. Today it’s $210 - only a 3% improvement. That might seem surprising given all the positive Q4 surprises, but there’s a trade-off happening:

The good:

• Better starting share count - 253M actual vs 277M assumed. That’s 24M fewer shares (9% better). Dilution was the silent killer in November. It’s now contained.

• Better margin trajectory - Q4 turned Adj EBITDA positive a year early. The five-year depreciation schedule validates GPU pricing durability. Net margins reach 15% by 2030, up from 12.5% in November.

• Lower dilution path - 305M ending shares vs 335M assumed. Zero ATM usage despite $16-20B capex needs. Management is funding with cash, OCF, deferred revenue, and debt - not equity.

The offset:

• Lower exit multiple - I’ve reduced the 2030 P/E from 35x to 27.5x (down 7.5x total, including the 5x haircut for depreciation benefit). Why? Because capturing this revenue opportunity requires massive leverage. The $16-20B capex in 2026 alone, plus continued buildout through 2030, means Nebius will be carrying substantial debt. Thats the exit multiple even as margins improve.

The November model assumed 30x exit P/E on 12.5% margins with 335M shares. The updated model assumes 27.5x exit P/E on 15% margins with 305M shares. Better operations, lower multiple, fewer shares. Net result: roughly flat valuation, but with much higher confidence.

Key model inputs:

• Revenue: $530M (2025) → $3.2B (2026) → $25B (2030)

• Capacity: 2,500 MW at $10M per MW (conservative - management guiding >3GW by YE 2026)

• Margins: 15% net by 2030 (vs AWS at 30%+, but Nebius is only 5 years old)

• Depreciation: Falls from 76% to 23% as revenue scales while depreciation falls

• Shares: 253M → 305M (21% dilution, down from 277M → 335M assumption)

• Exit P/E: 27.5x (reflects debt burden from $16-20B (and the rest) annual capex through buildout phase)

Why this matters:

If you believe in the $10m per MW thesis and 2.5GW deployment, the base case could offer 134% upside. If you don’t believe in it, you shouldn’t own the stock at any price.

4.6 Sum of Parts: What Is The Market Pricing In?

Here’s the most interesting part: the market is still mispricing Nebius’s subsidiaries - and for me this is our margin of safety, which I purposefully do not include in the valuations.

I don’t want to spend too much time on the subsidiaries or valuing them - but I’ve used management guidance and here alongside some rough workings to come up with basic valuations for all 4. (maybe will be 5 through Tavilly next time)

The market is valuing the Nebius AI business at around $18bn if you assume the subsidiaries do carry value. This is a chunky margin of safety that adds conviction to my base and bull models.

5. The Bull Case: What Happens If Nebius Executes at the Upper End?

Read more

Here’s my take on the fear that set into the market over the last couple of weeks.

1. Market volatility

2. CapEx spending

3. Software is dead

4. Nebius

5. Amazon

6. Valuation

7. Analysis Updates

Market volatility

Within the last week or two, we’ve begun to see sharp volatility in some areas of the stock market - and there is a lot to talk about.

• Gold and Silver suffered one their worst drop in decades, sparked by fears that interest rates will not ease quite as quickly following the nomination of new Fed Chair Kevin Warsh.

• The NASDAQ dropped ~6.5%

• Profitable, high moat software stocks took a beating, as fears that AI could make them redundant took hold.

CapEx Spending

Amazon just guided for $200bn in CapEx in 2026. Thats almost double all cash from operations for Amazon in 2024.

To be frank, all of the hyperscalers are spending dramatically more - other than Apple, who is happy to wait and see.

If Amazon can not provide a solid return on investment for this $200bn spend, we could be entering dark days for Amazon - and the bubble fears start to become cemented. When we see spending commitments far outweigh the liquidity of companies committing to them - that is classic bubble territory, and investors are rightly worried.

OpenAI has been committing to huge CapEx, off a very small unprofitable revenue base, and this is worrisome.

This shocking increase in spend for Amazon has rightly spooked investors, who will need some time to digest this figure. Amazon stock was down around 10% after earnings, and currently sits at around 19% off its all time high.

It’s a stark difference to Google, who also guided for a huge CapEx spend, but the entire spend was covered by free cash flow. If Googles 2026 CapEx investments don’t produce any return, it will not hurt the future bottom line.

Software Stocks

The S&P dropped modestly over the past week or so, but software dropped aggressively.

The idea is that AI is becoming so powerful, that software will become completely redundant, just access to an LLM will be enough to automate everything that software does.

And, to be fair, there is an element of truth here.

Claude released a plug-in for legal professionals, which caused a meltdown in stocks related to legal software. Probably for good reason. Having used some of these tools in the past - they have access to licensed data, which is great, but they are mostly a UI for that data. The ability to search, filter, and query these data sets will be vastly improved by an LLM.

Especially if that LLM has access to other tooling, maybe it can draft contracts, advise on negotiations, or perform legal research - much like a legal junior might do - completely autonomously in the background.

https://legaltechnology.com/2026/02/03/anthropic-unveils-claude-legal-plugin-and-causes-market-meltdown/

So yes, some of these companies should be worried, but not all.

The likes of Palantir, Engineering software, Platform software (AirBnB, Uber), Social media (Meta), and many more - all software with incredibly high moats, I think will be largely unaffected.

In fact, all of these companies above, and even potentially the likes of Salesforce, ServiceNow, and other SaaS companies whom are great businesses - will probably benefit from AI not cause their downfall. So I do think this is all a bit overdone, but the market will figure out over time what the true impact is.

Nebius

The interesting thing I noted last week was that Nebius fell with these software stock fears - peculiar to me, because in a world where AI models replace all software, we are going to need a drastically increased amount of compute than what we are envisioning right now.

I’ll be providing an updated look at Nebius after its earnings results, but as somebody who thought their outlook was looking so strong - a drop in price due to fears that ‘AI might get even more important and powerful’ has felt like a fantastic opportunity.

Amazon

First of all lets digest the CapEx spend.

$200bn is a big number, and it’s a roughly 55% increase on last year’s capex spend of ~$128bn. But here’s the thing - capex doesn’t hit the bottom line all at once. It gets depreciated over time, meaning the $200bn will weigh on profits gradually over the coming years, not as a single lump sum in 2026.

What will take an immediate hit is free cash flow. Amazon generated about $140bn in operating cash flow last year - so if that holds steady, a $200bn capex bill leaves them roughly $60bn in the red on a free cash flow basis. That’s a big number, but Amazon is sitting on ~$80bn in cash and has no trouble raising debt, so they can absorb it.

The real question isn’t whether Amazon can afford to spend $200bn - they can. It’s whether that $200bn earns a strong return on invested capital over time.

I noted the bubble risks of AI spending earlier, and OpenAI is rightly the posterboy for such bubbalicious fears, but Amazon should not be lumped in with Mr Altman’s antics.

Amazon believes this investment needs to happen in order to meet AI demand, and is investing as much as they prudently can to capitalise on the opportunity.

Make no mistake, there is some risk here, and Amazon’s net income has been helped by investing returns over the past year which may make their liquidity seem stronger than it is, but remember they have ~80bn cash in hand. They can afford to take a little risk, and I am happy for them to do so if they believe they will secure a strong return on invested capital - in fact, I think it is absolutely necessary.

So will Amazon continue to grow? I think there are a few catalysts for Amazon which make the outlook incredibly promising.

Amazon’s Outlook

Claude and Anthropic

Estimates suggest Amazon owns 15-20% in Anthropic, creator of Claude. Anthropic’s largest external investor.

https://www.businessinsider.com/amazon-ai-bet-anthropic-soars-61-billion-valuation-2026-2

• This is just generally a great investment, I am incredibly jealous

• Claude is becoming the chosen tool for enterprise AI

• So much so, it is leading to software stocks dropping immensely

Claude work, which connects directly to your local machine, and can do whatever you want it to do on your computer, is showing how powerful AI can be. It seems Claude’s rise has directly contributed to software selling off.

As I mentioned earlier, I don’t believe software is dead or will become redundant, but Claude will certainly eat some of its breakfast.

First and foremost, I believe this is an excellent investment, and owning a sliver of Anthropic is a plus point for owning Amazon even if it’s a very small piece of the pie for now.

Anthropic and AWS

The more salient point regarding the Anthropic investment is that Anthropic runs primarily on AWS.

In fact, Anthropic just last October committed to using 1 million of Amazon’s bespoke ‘trainium’ chips.

https://finance.yahoo.com/news/amazon-says-anthropic-will-use-1-million-of-its-custom-ai-chips-141829311.html

So far it looks like this partnership, set amongst the backdrop of increased cloud demand, is leading to a reaccelerating for AWS growth. Expanding to $35.6bn a quarter, and growing at over 20% - pretty staggering growth.

Tranium and Graviton

Amazon’s custom chip business - Trainium (for AI workloads) and Graviton (for general compute) - is no longer a side quest. In their recent earnings, Amazon disclosed the combined chips business is now operating at a run rate of over $10 billion, growing at triple-digit percentages year over year. Trainium alone is a multibillion-dollar run rate business, and is fully subscribed.

This is a huge deal for Amazon and I believe will significantly boost the bottom line over years to come. Why?

• Better Margins: When a customer uses an NVIDIA GPU, Amazon has to pay NVIDIA for the hardware, which eats into their profit. By using Trainium, Amazon “recaptures” that margin because they own the entire stack - from the silicon design to the data centre cooling.

• Customer Savings: Trainium instances are typically priced 30-50% cheaper than comparable NVIDIA instances. This lower price point attracts cost-conscious companies that are “priced out” of high-end NVIDIA hardware.

Robotics and Automation

Amazon has invested a huge amount of capital into automating workflows within its warehousing units. So much so that this has begun to result in layoffs for its employees - which while of course bad for employees, and depending on your viewpoint, the overall economy, this is of course good for Amazon shareholders whichever way you think about it.

It reduces a huge operating expense for Amazon. This trend will continue, and investments made now will result in better margins as robotics improves.

Advertising

Amazon’s advertising business has grown rapidly, and advertising is a huge beneficiary of AI.

Just look at Meta’s revenue growth recently, while Amazon isn’t quite there yet - it bodes well for this section of the business.

Amazon Leo

Some don’t know but Amazon does have a space economy arm, primarily focussed on satellites - formerly known as Project Kuiper, now known as Amazon Leo.

The space economy has huge potential, and is a theme I am always looking for exposure in, right now Leo doesn’t move the needle - but this once again is another plus and potential for excellent upside.

Valuations

NOTE: I’ve built a dashboard with AI to help me tailor inputs for valuations, and give a clear view for readers as to how my assumptions impact the overall valuation for the company.

The valuation framework is the same as I have used in the past, looking out 5 years and using an exit multiple - but hopefully this new formatting makes everything a little easier to digest. Please let me know what you think!

Base Case

Read more

I wanted to provide a short note to provide some colour on the company, and give the rationale for why you might want to consider adding Filtronic to your watchlist.

Contents

1. Executive Thesis Summary

2. What Does It Actually Do?

3. Economic Moat

4. Balance Sheet

5. Management

6. Narrative

7. Risks

8. Projections and Valuation

1. Executive Thesis Summary

Filtronic manufactures mission-critical components for high-frequency data communications, primarily used in space and defence applications. As the space economy scales, data transmission in space will scale with it. Filtronic sits directly in that value chain. Easy correlation to make, and thats what got my attention here.

It is a picks-and-shovels investment in the infrastructure required to move data reliably in extreme environments - supported by two decade long megatrends we are in the infancy of:

• Rising European and Western defence spending

• The rapid expansion of the global space economy

The stock is not cheap. However, at a market capitalisation of roughly £420m, it is still starting from a very small base relative to the size of the markets it serves.

Filtronic has partnered with the current leader of the space economy, SpaceX, which has chosen to buy Filtronic’s hardware rather than build it in-house. Given SpaceX’s strong preference for building in house, that decision matters.

The most recent SpaceX contract was almost larger than Filtronic’s entire FY2025 revenue. If additional customers are onboarded and the space economy continues to scale, Filtronic’s revenue - and therefore market capitalisation - could increase rapidly in discrete steps as new contracts are signed.

This is not a company to allocate 20% of a portfolio to. It is one to watch very closely. I currently hold a small position (~2%), which I would look to increase at more attractive levels or reduce if the thesis breaks.

This is not a value stock in the traditional sense. It carries real execution and customer-concentration risk.

2. What Does It Actually Do?

Filtronic is best thought of as high-tech plumbing for the invisible world of radio waves.

While we interact with data via phones and computers, that data ultimately travels through the air as radio signals. Filtronic builds the specialised hardware that ensures those signals are:

• Powerful enough to travel long distances

• Clean enough to avoid interference

• Robust enough to operate in space or hostile environments

What they make

• Power amplifiers – boosting weak signals so they can reach satellites or distant radar systems

• Filters – removing interference in crowded radio environments

• Transceivers – integrated systems that transmit and receive data at extremely high speeds

This hardware operates at very high frequencies, where performance requirements are unforgiving and failure is not an option.

3. Economic Moat

Filtronic’s moat does not come from brand or scale. It comes from operating in an extremely difficult manufacturing niche.

• High-frequency (E-band) RF hardware - which is incredibly hard to manufacture

• Increasing demand for volume production

• In a world where sovereignty and self-reliance is moving higher up the priority list, being a British company that isn’t reliant on US or Chinese chips for defense modules is a massive plus.

Many companies can design RF components. Very few can manufacture them at scale, profitably, and to the reliability standards required by space and defence customers.

A useful analogy is engines: many companies can build a standard engine, but very few can build a Formula 1 engine - and then produce thousands of them reliably.

The fact that SpaceX continues to source from Filtronic rather than internalising production is a strong external validation of this moat. In fact, they actually have warrants for 5% of the company - this is a genuine strategic

Moat rating: 8 / 10

4. Balance Sheet

Filtronic’s balance sheet is sound but not a central part of the thesis.

• No excessive leverage

• Investment is being funded primarily through operating cash flow and prior profitability

• Recent CapEx reflects capacity expansion, not financial stress

Near-term cash flow and margins are being depressed by investment, but this is intentional, and I can absolutely forgive them for investing in their own future cash flows.

There is no balance-sheet fragility that forces dilution or defensive behaviour.

Balance sheet rating: 8 / 10

(Healthy, supportive, but not a hidden asset or backstop)

5. Management

Management has demonstrated a willingness to:

• Invest ahead of demand

• Secure large, strategic contracts

• Accept near-term margin pressure to enable long-term scale

The decision to invest heavily in the Sedgefield facility is particularly telling. This will likely suppress reported profits over the next 12 months, but it clearly signals intent to support materially higher output in the future.

Further proof points will come from:

• customer diversification

• delivery execution

• margin normalisation post-investment

Management rating: 8 / 10

(Strong strategic intent; execution still to be proven at scale)

6. Narrative

Filtronic currently benefits from a powerful but volatile narrative:

• Space is one of the fastest-growing strategic industries globally

• Defence and space are becoming increasingly intertwined

• European and Western countries are beginning to invest more heavily in defence

• FY2025 revenue growth of 121% demonstrates huge demand

• Often discussed as a partial public proxy for SpaceX’s infrastructure spend

This narrative cuts both ways. It can drive rapid re-rating, but it also raises expectations. For now, the market is catching up - the stock is rising meteorically.

There is no doubt however, that if you believe the space economy will scale rapidly - as many do (including me), then the tailwinds for filtronic could continue for many years to come.

Narrative rating: 10 / 10

7. Risks

Key risks are clear and should not be understated:

• Customer concentration, particularly SpaceX

• Lumpy revenue recognition

• Execution risk as output scales

• Valuation risk – this is not a cheap stock

If SpaceX demand slows, or if manufacturing scale-up proves more difficult than expected, the downside is real.

8. Projections and Valuation

Filtronic is difficult to value precisely:

• Earnings are distorted by current investment

• Growth is lumpy rather than linear

• Margins are in transition

The market is currently looking past the 2026 'investment dip' and is pricing Filtronic on around a 50x P/E multiple of FY27 expected earnings. Filtronic is investing heavily at the moment, and given the rate they are growing, it’s difficult to forecast what their end state profit will look like.

The valuation is not cheap, and carries uncertainty - but the strength of the narrative means Filtronic needs watching closely to see how it executes. If Filtronic can keep winning contracts and embedding itself into the space economy at the rate it has done, I have no doubt the current valuation will cheap in hindsight.

The investment case ultimately rests on one question:

Can Filtronic retain and grow its moat as the space economy scales?

• If yes, today’s valuation will look reasonable in hindsight in years to come.

• If not, it will not.

I have begun a starter position, and intend to watch this one very closely. I added this stock as one of my strongest ideas of 2026, with an official buy at the price of £1.86 - not because its the safest idea I have, but because it poses immense potential upside yet is largely unknown.

For subscribers - here’s Filtronic ranked in my UAV model after my initial research phase has completed. Not quite a screaming buy but certainly one to watch.

The information provided in this article is for informational and educational purposes only and does not constitute financial, investment, or professional advice. The author is not a licensed financial advisor, and the content is based on publicly available data, personal analysis, and opinions.

1. Performance in 2025

2. Strongest ideas for 2026

3. Updated UAV & stocks to watch (paid subscribers only)

Performance in 2025

One of the core principles of Root Logic is transparency.

To track the progress of this blog - and to hold myself accountable - I record the share price at the time of publication for every idea I put forward. I track how those ideas perform over time - whether they grow 100% in 6 months, or go to 0.

All of the selections below were published during 2025.

If an investor had allocated equal capital to each idea at the time of publication, the portfolio would have returned +17.9% over ~7 months.

Over the same period, the S&P 500 returned ~16.5% (as of 2 February 2026).

That comparison isn’t perfect:

• In reality, position sizing is never equal

• A couple of names (notably Hims and Novo Nordisk) were exited earlier

But the goal here isn’t to gloat about how I may or may not manage my portfolio in private - or make trades prior to posting publicly. The goal is to share ideas publicly and track their progress transparently - something fairly rare in this space.

If you’re a paid subscriber - you have full access to my proprietary upside adjusted valuation stock screener, complete with company appraisals and valuations. If you have questions about the screener, please let me know.

So whats next?

Strongest Ideas for 2026

It’s been a while since I shared a consolidated update on the stocks I’m watching. The last company I highlighted as materially undervalued was Bloomsbury, late last year.

My strongest ideas heading into 2026 are:

• Nebius Group

• Games Workshop

• Bloomsbury

• Filtronic

I’m adding all four as my first picks of 2026.

Each combines what I care about most:

• A strong and durable narrative

• Clear medium‑term catalysts

• A valuation that still implies meaningful upside

Nebius Group

I’ve written fairly extensively about Nebius, and I continue to believe it is one of the strongest AI investment opportunities available today.

What stands out:

• A diversified AI portfolio through its subsidiaries, all showing strong momentum

• One of the most experienced data‑centre and infrastructure teams in the neocloud space

Nebius recently onboarded Revolut as a customer. Executive Director Pavel Nesterov highlighted why Nebius was chosen - and a few points are particularly telling:

Key takeaways:

• Hyperscalers did not meet Revolut’s AI requirements - Nebius did

• Nebius delivered the required solution faster than hyperscaler competitors

• Customer satisfaction continues to translate into larger, more sophisticated clients

My next Nebius update will follow the upcoming earnings release. I am still accumulating shares.

Percentage undervalued according to my valuation: 54%

Games Workshop

Another carry-over from 2025 - and one where the investment case has quietly continued to strengthen.

I walked past numerous Games Workshop retail stores over the last couple of months - they are always full of enthusiasts.

Since my original write-up, there has been no single headline-grabbing announcement, but the underlying business has continued to deliver consistently strong results, reinforcing why Games Workshop remains one of the highest-quality IP businesses in the UK market.

What makes Games Workshop exceptional:

• A near-unassailable moat built around Warhammer IP

• Direct-to-consumer distribution with pricing power

• A passionate, self-reinforcing global community

• High-margin licensing optionality layered on top of an already profitable core business

The upcoming Amazon TV adaptation represents a genuine step-change opportunity. If executed well, it materially expands the addressable audience for Warhammer without requiring incremental capital from Games Workshop itself.

In my view, any weakness in the share price ahead of meaningful updates on the Amazon project continues to represent opportunity rather than risk.

Percentage undervalued according to my valuation: 32%

Bloomsbury

Bloomsbury is, in many ways, the publishing equivalent of Games Workshop.

It is a business whose value is structurally underestimated because the market tends to treat publishing as cyclical, rather than recognising the durability of elite intellectual property.

Bloomsbury owns the rights to one of the most enduring IP franchises ever created: Harry Potter.

A Harry Potter TV series is now gaining momentum, with a likely release window around 2027. While still some way off, this represents a multi-year catalyst that could meaningfully re-rate market expectations for Bloomsbury’s long-term earnings power.

More immediately, Sarah J. Maas is expected to release a new book in the ACOTAR series (affectionately known as the fairy porn series). Historically, major releases in this franchise have driven significant revenue and profit spikes, and I expect the next instalment to be no different.

Importantly, Bloomsbury has shown discipline in capital allocation, maintaining strong margins and a robust balance sheet while monetising its back catalogue across formats.

As with Games Workshop, I see short-term volatility ahead of long-dated catalysts as a fantastic entry point.

Percentage undervalued according to my valuation: 42%

Filtronic

A scarcely known UK industrial stock - Filtronic produces advanced micro electronics for the defence and space economy.

At its core, Filtronic designs and manufactures highly specialised RF and microwave components used in defence, space, and satellite communications. These are not commoditised products - they are mission-critical systems with long qualification cycles and extremely high switching costs.

The clearest validation of Filtronic’s moat is its relationship with SpaceX.

Filtronic components are installed in every Starlink satellite produced.

This matters more than it might first appear. SpaceX is famously vertically integrated and will aggressively bring production in-house whenever it makes sense. The fact that Filtronic continues to supply these components strongly suggests:

• Deep technical expertise that is difficult to replicate

• High reliability in mission-critical environments

• Trust earned through long-term delivery performance

Financially, growth has been exceptional:

• Revenue grew 121% in 2025

Despite this, Filtronic trades on an earnings multiple of roughly 30×, which looks modest when set against:

• The visibility of SpaceX-related demand

• Expanding European defence and space budgets

• The company’s improving operating leverage

As production volumes scale, incremental margins should continue to improve, which I believe the market has yet to fully price in.

Filtronic is not a household name but this business is supplying the picks and shovels of the space and defence economy.

More to come on this one in due course.

Percentage undervalued according to my valuation: 37%

The information provided in this article is for informational and educational purposes only and does not constitute financial, investment, or professional advice. The author is not a licensed financial advisor, and the content is based on publicly available data, personal analysis, and opinions.

Read more

When you think of Bloomsbury, most people jump to them being the Harry Potter publisher. If you’re thinking about Bloomsbury the stock, you might think the best days are well behind it.

I believe the market agrees with this sentiment, which right now is reflected in the stock price - currently 38% from its most recent highs.

But, behind the obvious Potter association sits:

• A growing, high-margin digital and academic publishing platform

• The most commercially powerful author in modern fantasy via Sarah J. Maas

• A multi-year Harry Potter TV catalyst expected for 2027

• A recently announced AI partnership with Google Cloud

As is often the case with stocks I end up interested in, the share price has not been kind.

Bloomsbury’s May results showed profits down despite higher revenue, and since then the stock has struggled.

In this article, I’ll make the case that the current pause in the share price is a potential opportunity to capitalise on the Harry Potter supercycle, a global Sarah J Maas craze, and a new business line which has more upside than most imagine right now.

Why the Stock is Down

To understand Bloomsbury today, you need to start with how unusual FY2023/24 really was.

That year benefited from a perfect storm:

• a global “romantasy” wave led by Sarah J. Maas

• TikTok (BookTok) pushing her backlist into the mainstream

• ongoing Harry Potter strength, amplified by the 25th-anniversary cycle

It was, by any reasonable definition, an exceptional year; that set up a brutal comparison for 2024/25. It was always going to look weaker optically.

And sure enough, the headline numbers gave the market what it wanted:

• revenue rose from £342.7m to £361m, around 5% growth

• profit before tax fell to roughly £32.5m

• adjusted profit came in around £42m, implying margins in the low-teens

From there, the narrative wrote itself:

“Profits down. Margins normalised. The Maas boom is over. Sell.”

But if you actually read the annual report and subsequent updates, a different picture starts to emerge.

The profit dip was not driven by demand collapsing. It was driven by:

• the absence of a new Sarah J. Maas frontlist release

• normalisation of Consumer margins after an extraordinary prior year

• ongoing investment and integration effects

Underneath that noise:

• core revenue is still growing

• the Academic & Professional division continues to scale (up 20% in H1 this 2025/2026)

• Bloomsbury Digital Resources is expanding, with management targeting roughly £41m of revenue by 2027/28 versus around £27m today

So the market feels like the company has peaked, but I think we’re somewhere else entirely.

This looks more like a good business that has just absorbed a demand surge, is reinvesting into long-duration digital annuities, and still has multiple IP-driven catalysts ahead of it.

Why this rhymes with Games Workshop

If you read my Games Workshop piece, you’ll recognise the shape of this idea.

Both businesses share:

• exceptional IP with multi-decade endurance

• loyal, global fanbases

• optionality through TV and film adaptations

• backlists that monetise for decades, not quarters

Bloomsbury just expresses this in books rather than miniatures.

At the centre of the thesis sit a few key franchises and formats:

• Harry Potter: the original engine

• Sarah J. Maas: the romantasy phenomenon

• Digital academic resources: the quiet, SaaS-like core

• Other IP: from The Three-Body Problem to Samantha Shannon and beyond

Each on its own is interesting. Together, they form something the market still struggles to categorise.

Harry Potter: the next wave of hysteria is brewing

There are two moving parts here: Bloomsbury’s publishing levers, and Warner Bros’ TV machine.

In the nearer term, Bloomsbury has launched a “Pocket Potters” range - compact editions designed for younger readers, collectors, and subscription formats. This is good to see, but the bigger lever is further out.

The HBO / Max Harry Potter TV series is expected to debut around 2027, with one season per book.

History matters when doing the analysis here - during the original film cycles, Potter book sales rose materially - estimates range from 20 to 40%. Even the 25th-anniversary cycle delivered a roughly 35% uplift.

And unlike films, a long-running series is a multi-year fan-acquisition engine, not a one-off spike. When the series releases, some viewers may be consuming Potter for the first time, and it may encourage them to delve into the book universe.

This is a couple of years away at least, but I would expect the hype to begin building late next year. The stock will almost certainly price in the new Harry Potter cycle before the series releases.

Sarah J. Maas: the romantasy engine (with a gap)

Sarah J. Maas was the best-selling author in the US in 2024. Forbes reported she sold 3.1 million copies in the first four months alone, and by mid-year had sold nearly 5 million print copies. She has captured the ‘romantasy’ genre, particularly through her A Court of Thorns and Roses (ACOTAR) series - which many are referring to as ‘fairy porn’.

Her backlist continues to perform strongly even without a new release, and I believe this is the biggest driver for the Bloomsbury share price drop in 2025.

That changes again soon. The next major Maas release, ACOTAR #6, has completed its first draft and is expected in late 2026 or early 2027.

Historically, new Maas releases don’t just sell well in isolation. They re-ignite demand across the entire catalogue - backlist volumes, box sets, and adjacent series - creating a step-change in revenue rather than a one-off spike. The pattern is similar to what we see in other fandom-driven ecosystems (and expect to see in Harry Potter).

There is also adaptation optionality. I don’t underwrite this in the base case, but it remains a free option. Previous attempts haven’t landed, but that doesn’t invalidate the long-term appeal of the IP - it simply means the timing hasn’t been right yet.

Digital and academic: the quiet compounding engine

Bloomsbury Digital Resources (BDR) is the least visible part of the business and arguably the part with the clearest compounding economics.

BDR is sold to universities, libraries, and academic institutions on a subscription basis. Once embedded in curricula and research workflows, churn is low and pricing power is stronger than in consumer publishing.

The Numbers

• Current BDR revenue: approximately £27m

• Management target for FY2027/28: approximately £41m

• Academic & Professional division revenue grew 20% in H1 2025/26 to £46.1m

• Profit margin for the division reached 24% in H1 2025/26 (up from 16%)

While Bloomsbury does not disclose standalone margins for BDR, academic digital publishing economics are well understood: gross margins typically exceed 70%, and operating margins can reach 30-40% at scale. This is classic operating leverage: content is already owned, platforms are built, and distribution is digital.

The AI Angle

Bloomsbury owns substantial proprietary academic content across the humanities and social sciences. This is structured, curated, peer-reviewed material, precisely the kind of data that is scarce and legally clean in an AI-driven world.

Two recent developments validate this thesis:

1. AI Licensing Agreement (July 2025): Bloomsbury signed its first non-exclusive AI licensing agreement, demonstrating the monetisation potential of its academic content.

2. Google Cloud Partnership (December 2025): A strategic collaboration to integrate Google’s AI technologies (NotebookLM, Vertex AI, Gemini Enterprise) for improved content discovery, learning outcomes, and inventory management.

I do not explicitly underwrite a large AI-driven revenue uplift in my valuation, as that would be premature. However, AI meaningfully raises the ceiling on what BDR can become. If BDR evolves into a £40m+ revenue business with 30%+ operating margins, supported by proprietary data and AI-enhanced workflows, it becomes one of the most valuable assets in the group.

Management and Key Risks

Management: Founder-Led Since 1986

Nigel Newton founded Bloomsbury in 1986 and remains CEO. This is genuinely rare. Founder-led businesses with multi-decade horizons tend to think differently: they invest through cycles, protect brand equity, and make capital allocation decisions with longevity in mind.

Newton famously gave the Harry Potter manuscript to his eight-year-old daughter to review before publishing. Bloomsbury’s track record reflects a culture that values quality IP and long-term thinking. The company was named Publisher of the Year at the British Book Awards in May 2025.

Risk Factors

Any honest investment analysis must address the risks so here are the key concerns:

1. Key Person/IP Risk: A significant portion of consumer revenue depends on Sarah J. Maas. Any change in her publishing relationship, writing output, or personal brand could materially impact results. Similarly, the Harry Potter franchise depends on continued engagement with J.K. Rowling’s IP, and JK Rowling herself - who is a polarising figure.

2. BookTok Dependency: The Maas phenomenon has been heavily amplified by TikTok virality. Social media trends can shift quickly, and there’s no guarantee this dynamic persists.

3. Retail Concentration: In January 2025, Amazon UK temporarily stopped selling Bloomsbury titles in a contract dispute. While resolved, this highlights channel concentration risk. Major retailers have negotiating leverage.

4. Publishing Cyclicality: Publishing has historically been cyclical, with demand tied to economic conditions and consumer spending patterns.

5. Currency Exposure: With significant US operations (approximately 54% of revenue), Bloomsbury has material FX exposure. BDR revenue was flat in constant currency in H1 but declined due to FX headwinds.

6. Integration Risk: The Rowman & Littlefield acquisition ($83m) is ‘substantially complete’ but ongoing digitisation and integration costs could pressure near-term margins.

7. Execution Risk on Digital: The BDR growth targets are ambitious. Failure to scale digital subscriptions would undermine a key part of the thesis.

While these risks warrant monitoring, my base case assumptions are conservative enough to provide a margin of safety. Even if some of these headwinds materialise, the thesis remains intact given the current valuation discount.

Valuation: A 5-year framework

Assume Bloomsbury simply carries on doing what it already does reasonably well.

No blockbuster surprises. No heroic assumptions.

• Consumer publishing grows steadily but unspectacularly

• Academic and digital continue to compound in the background

• Margins improve modestly as digital mix increases

• No transformational M&A

• The Harry Potter HBO series attracts attention, but does not yet trigger a full supercycle

In other words: competent execution, nothing more.

Under this scenario, Bloomsbury compounds revenue at 6% annually over the next five years, taking group revenue from ~£361m to ~£483m. Net margins expand modestly toward ~12.5% as higher-margin digital revenue becomes a larger part of the mix.

That results in Year-5 net income of roughly £60m.

Using a 5-year DCF with:

• 6% revenue CAGR

• Terminal net margin of ~12.5%

• 9% discount rate

• Terminal P/E of 14× applied to Year-5 earnings

This implies a terminal value of ~£10.30 per share in Year 5. Discounted back to today, and combined with interim cash flows over the period, this results in an implied fair value of ~£8.70 per share.

At around 475p today, that represents roughly 80% upside from a scenario where relatively little goes right.

The assumptions here are deliberately conservative. The revenue growth rate is well below Bloomsbury’s five-year historical average (inflated by exceptional Maas-driven years), and the terminal multiple assumes the business is recognised as a high-quality publisher - not a media or IP platform.

Given the durability of the business and the continued dividend (currently ~3%), this base case alone already looks attractive.

Bull Case (Potter Supercycle, a Maas Hit, and Digital Scale)

Now assume Bloomsbury’s optionality actually shows up.

• One major Sarah J. Maas release lands during the period and is received well

• Digital Academic (BDR) hits its stated targets

• The Harry Potter TV series is a genuine hit

• Operating leverage kicks in as higher-margin digital revenue grows

None of this requires heroic assumptions. It is simply Bloomsbury doing what it has done historically, with scale.

In this scenario, revenue compounds at 9% annually, taking group revenue to ~£555m by Year 5. Net margins expand toward ~13.5% - producing ~£75m of net income.

Using a 5-year DCF with:

• 9% revenue CAGR

• Terminal net margin of ~13.5%

• 9% discount rate

• Terminal P/E of 16× applied to Year-5 earnings

This implies a terminal value of ~£14.70 per share in Year 5. Discounted back to today, this results in an implied fair value of ~£11.60 per share.

That represents 140%+ upside from today’s price.

Conclusion

At ~479p, the current Bloomsbury price only makes sense if you believe:

• Harry Potter never meaningfully re-accelerates

• Sarah J. Maas is permanently past peak relevance

• Digital endeavours remains immaterial

I don’t believe any of those assumptions will hold. Bloomsbury is sitting well below it’s 5 year average P/E and has suffered from a near 40% drop - I believe all the risks and worries are now priced in. I have begun purchasing Bloomsbury and believe the company is significantly undervalued.

The information provided in this article is for informational and educational purposes only and does not constitute financial, investment, or professional advice. The author is not a licensed financial advisor, and the content is based on publicly available data, personal analysis, and opinions.

Read more

1. Introduction
   Why Nebius deserves a proper valuation framework after Q3

2. The Four Horseman of the Nebius Apocalypse

3. The Base Case (2025–2030)
   3.1 Revenue Growth
   3.2 Depreciation as a Percentage of Revenue
   3.3 Net Income Margin
   3.4 Shares Outstanding
   3.5 Valuation

4. The Bull Case
   What happens if Nebius executes at the upper end of its potential

5. Risk–Reward Summary
   How the scenarios compare, and a view on the opportunity

1. Introduction

The market has been weak over the past month, and Nebius has taken a real beating alongside it. That isn’t especially surprising for a company still in its infancy – when you’re priced for hypergrowth, you live with volatility.

Up to now, my valuation work on Nebius has been deliberately rough. We simply didn’t have a clear enough view into the unit economics to understand where this business might realistically end up. But after the Q3 results, I finally feel confident enough to move from broad strokes to something approaching a proper framework.

In this article, I’m going to lay out my three scenarios – the bear, base and bull case – and explain exactly how I’ve modelled them.

To structure the analysis, I’ve taken a threat-modelling approach: identify the variables that genuinely move the valuation, then examine how changes in those variables reshape the outcome. These are the four threats that matter:

• Revenue growth

• Depreciation as a percentage of revenue

• Net income margin

• Shares outstanding

These are, in my view, the most sensitive levers for Nebius’s valuation over the next five years. If any one of them breaks, the valuation breaks. If all four fall into place, the upside is significant.

I’ll provide a quick summary of the importance for all 4 before diving into my valuation frameworks.

2. Nebius Threat Model: The Four Horseman

Maybe in time these will become the four horseman of the Nebius apocalypse, but for now these are just the four main threats to business model feasibility.

Revenue growth

Growth is non-negotiable. For Nebius to justify any sort of valuation premium, it needs to bring at least 2.5GW of contracted power online over the next few years and convert that into active, revenue-producing capacity. Everything else flows from this. If the growth slips, every scenario collapses.

Depreciation

Depreciation is one of the key worries here. Michael Burry recently raised concerns about overly long asset lives in data-centre businesses, and he’s right. A six-year depreciation schedule simply doesn’t reflect economic reality for AI hardware. Nebius uses four years, which I think is financially prudent.

The more interesting question is whether GPUs continue earning money after year four. For training, absolutely not. But Nebius isn’t building a training cloud; it’s positioning itself as an inference cloud. Inference workloads tolerate older hardware far better, which means you can still monetise GPUs long after they fall out of the “premium” window. That’s why depreciation is a threat: not because the assets become worthless, but because the timing of the write-down affects when the business looks profitable on paper.

Net income margin

I have no interest in owning a business that plans to remain unprofitable for a decade. I’m all for heavy reinvestment – Nebius should be pouring capital into GPUs, land and data-centres for the next few years – but by the time we approach 2030, I expect actual profitability.

Net income matters because it reflects two things: depreciation growing slower than revenue, and capex falling as a percentage of revenue. Once those two flip in your favour, hyperscaler-style businesses can turn profitable remarkably quickly. If they don’t, none of the modelled outcomes mean anything.

Shares outstanding

This is the silent killer. The share price drop after earnings was, in my view, largely due to the announcement of a potential 10 per cent dilution via the 25m share ATM programme. Early dilution is fine – it’s part of hypergrowth – but dilution past 2027 is not.

If by that point Nebius is still issuing equity simply to keep expanding capacity, you run into the classic capital-intensity death spiral: more dilution leads to weaker returns, weaker returns demand more dilution, and so on. The Base Case assumes the opposite – that by 2027, capacity is producing enough cash to fund capex without leaning on shareholders.

All of these threats are interlinked, and I’ll be watching them closely each quarter to update the valuation framework.

With that in mind, let’s move on to the Nebius Base Case.

3. The Base Case:

The Base Case is not my prediction. It is my benchmark for “reasonable execution”. It reflects what I believe Nebius can deliver if the company does what it claims it will do, hits the majority of its milestones, and avoids self-inflicted wounds.

But numbers without context are meaningless. So before I give the scenario summary, I want to walk through the rationale for each of the four threats and show how they shape the Base Case.

3.1 Base Case: Revenue Growth

Revenue growth in the Base Case is not an act of optimism - it is the logical consequence of what Nebius has already committed to. If we don’t assume strong revenue growth, there really is no point going any further.

The company is targeting:

• $7–9bn ARR in 2026

IMPORTANT: ARR is the annualised run-rate measured at year-end, not the same as calendar FY revenue. Under fast growth, ARR runs ahead of revenue.

• 2.5GW of contracted power by 2026

• 800MW–1GW of active power in the same timeframe (GPU’s actually brought online ready to sell to customers)

• Multi-billion-dollar customer contracts with Microsoft and Meta

None of these allow for a “slow and steady” revenue trajectory. This is a company building ahead of demand, not chasing it.

In the Base Case, revenue ramps as follows:

• $525m in 2025 (guided)

• $6bn in 2026 (FY revenue, not ARR)

• $9bn in 2027

• $15bn in 2028

• $20bn in 2029

• $25bn in 2030

Why this curve?

Because Nebius is capacity-limited, not demand-limited. Every cluster they power on is immediately sold. The real threat, therefore, is not “do they have enough customers?” but “can they bring power and GPUs online fast enough?” If the power plan slips, everything else slips with it.

The Base Case assumes:

• Build-out continues roughly to plan

• GPU supply remains tight but navigable

• Core customers continue scaling workloads

• Enterprise adoption grows in parallel

This is not a smooth process — data centres never come online perfectly — but in aggregate, the company hits enough of its milestones to justify 25bn revenue by 2030.

I’ve chosen 25bn because it directly correlates with their 2.5GW of contracted power - at around $10m per MW.

I’m hoping you haven’t got this far and are wondering if I think AI demand will continue, if that’s you, please see my earlier article where I share my views on AI demand and how AGI will be here faster than analysts expect:

3.2 Base Case: Depreciation

Depreciation is the most misunderstood part of the entire Nebius model, and also the most important. A GPU cloud is a depreciation machine. It buys assets aggressively, writes them down over four years and either repurposes or replaces them as technology evolves.

The bear argument is that depreciation will overwhelm revenue growth and prevent profitability. The bull argument is that revenue will scale far faster than depreciation and the company will turn profitable rapidly. The truth sits somewhere in between.

In the Base Case, depreciation as a percentage of revenue falls from:

• 66 per cent in 2025 (yes 66 percent)

• to 35 per cent in 2026

• to 25 per cent by 2030

Why this pattern?

First, Nebius is deliberately front-loading capex. 2025’s $5bn capex increase all but guarantees high depreciation in 2026–2027. This is unavoidable and correct for a fast-growing infra business.

Second, revenue ramps disproportionately. If Nebius really does hit $7–9bn ARR in 2026, then FY revenue of $6bn is basically a mid-year snapshot of demand catching up with capacity. This is why depreciation collapses as a percentage of revenue - not because depreciation shrinks, but because revenue accelerates faster.

If capex continues to grow faster than revenue, depreciation will not fall - and net income will remain suppressed for longer. The Base Case assumes they strike the balance: heavy capex, but increasingly swallowed by far larger revenue, driven by a return on invested capital internally.

If this balance fails, profitability does not happen.

3.3 Base Case: Net Income Margin

Nebius is not allowed to be unprofitable forever. At some point, this has to become a real business with real earnings.

But profitability in AI infrastructure is not created by cost-cutting, it is created by three structural forces:

• higher utilisation

• stabilising capex

• falling depreciation intensity

Once those three turn in your favour, margins expand quickly.

In the Base Case, net margins go from deeply negative to:

• 5 per cent in 2027

• 10 per cent in 2028

• 11.5 per cent in 2029

• 12.5 per cent in 2030

I’ve modelled net margin in this way, because it mirrors the early economics of hyperscalers. AWS, Google Cloud and Azure all went through a similar curve: capex-heavy early years, followed by operating leverage as asset utilisation improves and software/inference layers begin to contribute more revenue.

I think Nebius is following the same playbook, but may even have the optionality for higher margins.

Of course, this assumes that Nebius actually earns the right to scale margins. If utilisation falters, if pricing collapses, or if hyperscalers undercut the market, these margins never materialise. The business is not in a good way in this scenario. I have spoken about all of these risks in previous articles, and while I watch them carefully I don’t believe they will materialise right now.

The Base Case assumes Nebius maintains high utilisation simply because everything they build is already pre-sold. If that stops being true, this margin path is too high.

3.4 Base Case: Shares Outstanding

Nebius is raising billions in capex over the next two years. Some of that is financed by debt and structured capital, but a clear portion will come from equity. The ATM programme signalled this, and the market reacted accordingly.

In the Base Case, share count moves from:

• 277m in 2025

• to 335m in 2030

This is modest. It assumes:

• no major equity raises after 2027

• capex begins to be covered by operating cash flow

• the company avoids serial dilution to keep pace with hyperscaler-scale investment

The threat here is existential. If Nebius is still reliant on equity raises in 2028–2030, the valuation collapses under its own weight. Not because the business is bad, but because returns cannot outpace dilution.

The Base Case is, therefore, a claim about capital discipline: Nebius uses equity early, then leans on operating cash flow and debt once scale arrives.

If this assumption is wrong, then once again entire Base Case is wrong.

3.5 Base Case: Valuation

Here is the full valuation framework, consistent with Nebius’s Q3 disclosures and the threat framework above. I’ve given you my reasoning for the narrative, and how they relate to the numbers, now it’s time to put everything together and give a present day value.

If you’re enjoying this analysis, consider becoming a paid supporter for just $5 a month to unlock the full valuation frameworks, analysis, and the latest view of my Upside Adjusted Valuation model.

Read more

As an investor trying to buy fantastic companies at below intrinsic value, this should be thrilling. And it is, but I’ve been moving carefully. Adjusting my valuation has not been straightforward. In fact, I need more time.

Up to now I’ve used a few simple frameworks, price-to-sales anchors, sum-of-the-parts, broad comps. But with Q3, we’re finally getting a clearer window into Nebius’s unit economics and long-term trajectory. The new data points mean I can build a more comprehensive valuation without feeling like I’m firing shots in the dark.

Valuation is an art, not a science, so yes, I’m still shooting from the hip, but the difference now is that the Q3 have allowed me to fire with more accuracy.

But, before I publish the full valuation model, I want to step back and share a qualitative take on the business, the reaction, and the broader question every investor won’t shut up about: are we in a bubble?

Here’s what I’ll cover:

1. Q3 results and the 22 percent reaction

2. Qualitative update on the business

3. Are we in a bubble?

Q3 results and the 22% reaction

If you haven’t looked at the Q3 numbers, the headline is simple: Nebius is in rude health. With a few caveats, which we’ll get to, but the underlying business is doing exactly what any AI cloud investor should want to see.

Business Development

• Contracted power target raised from 1GW to 2.5GW by end of 2026, with 800MW to 1GW expected to be physically connected.

• Nebius launched AI Cloud 3.0 and Token Factory, a clear pivot from GPU rental to a fully integrated, enterprise-ready AI cloud, very much taking a leaf from AWS’s playbook.

• AI coding startup Cursor was added as a customer.

• A 3 billion dollar deal with Meta was announced, and this was capped only by Nebius’s available active power. Meta wanted more.

Revenue Growth

• ARR expected to hit 7 to 9 billion dollars by end of 2026.

• That translates to roughly 6 billion dollars in FY2026 revenue (my estimate, not guidance).

• FY2025 revenue looks to land around 525 million dollars, meaning revenue is likely to increase over 1,000 percent year-on-year.

Profitability

• Core AI infrastructure EBITDA margin reached 19 percent

• Gross margins were around 71 percent

• Depreciation and CapEx overwhelm profitability for now, but that’s precisely what you’d expect in the build-out phase of a hyperscaler-like company.

Capital Requirements

This is where the financial prudes start clutching their pearls.

• Nebius filed to issue 25 million new shares, roughly 10 percent dilution

• CapEx for 2026 expected to exceed 5 billion dollars

• Total convertible notes plus equity raised in 2025: 4.3 billion dollars

These moves fund the data center and GPU expansion needed to go from roughly 500 million dollars in revenue today to 7 to 9 billion in ARR within two years. But they also explain why the stock fell. Rapid expansion requires capital, and some of that capital came at shareholders’ expense.

But you cannot have your cake and eat it. This is a capital-intensive business, and the return on invested capital is expected to dwarf the outlay.

So how can we be confident that these huge investments will generate outsized returns?

Because Nebius pre-sells every watt of active power it brings online.

Capacity doesn’t sit there waiting to be sold, it’s spoken for before it’s even switched on.

As Nebius put it on the Q3 earnings call:

“Capacity remains our main bottleneck. Everything we deploy, we sell. We see the demand that continues to significantly outstrip our supply each time we add new clusters.”

To summarise Q3

Nebius is growing faster than anyone expected. To keep up with demand, an enormous amount of capital is required. I view this as incredibly bullish, and I’m perfectly comfortable with the dilution.

Nebius has minimal debt. If dilution is the trade-off required to fund infrastructure that delivers exponential returns, I’m happy to pay it.

Critically, Nebius scales as demand grows, not the other way around. They’re not building giant empty data centers and praying. They’re building because demand is overwhelming. That is the ideal dynamic for any growth investor.

I’ll go deeper into the unit economics in the valuation article, but for now, let’s zoom out.

Qualitative Update on the Business

As Warren Buffett once said (I think), a narrative without numbers is a fairytale. Numbers without a narrative is a spreadsheet.

I want to talk about how I’m feeling about Nebius’s positioning and execution, but I won’t rehash everything:

• Nebius is now seen as a gold standard AI cloud provider, ranked ahead of AWS and Google Cloud in many AI workload tests, and competitive with Oracle and Azure.

• Nebius is also drastically cheaper than all of them.

A leading product in a hypergrowth market at an extremely competitive price point? That’s a recipe for scaling fast.

But what really gets me excited is the power build-out.

Nebius is on track for 2.5GW of contracted power by FY2026. At roughly 10 million dollars per MW, this is 25 billion dollars of ARR potential once online.

The previous estimate was 1GW. This is a massive step-change in ambition. And frankly, I don’t think 4GW of active power by 2030 is out of reach, which would imply an eye-watering level of revenue.

Every time they secure more power, it instantly expands their future earnings capacity.

Do I think we’re in a bubble?

In short, no. I don’t believe we are.

And one part of the bubble conversation I rarely see discussed is just how close we are to reaching AGI, Artificial General Intelligence, which I define simply as:

• Intelligence that surpasses humans across all cognitive tasks

• The ability to pass a strong Turing Test

• The capacity to solve problems humans can’t even conceptualise

ChatGPT in 2022 was rudimentary. Today, AI is embedded across enterprises globally, and is following the steepest adoption curve of any technology in human history.

By 2030, I fully expect AGI to be a reality. And when that happens:

• Companies will hire compute instead of hiring developers

• Pharma companies will need massive AI cloud access to simulate drugs

• Robotics will consume enormous inference workloads

• Autonomous vehicles will become the norm

• Entire industries will be upended

And AI factories, Nebius’s core business, will be the essential infrastructure enabling all of it.

Expert estimates for AGI vary:

• 2026 to 2027: Elon Musk, Sam Altman, Jensen Huang

• 2030 plus: Yann LeCun, Noam Chomsky

• Midpoint: Ray Kurzweil’s long-held prediction of 2029

I’ve just finished Kurzweil’s The Singularity is Nearer, and he makes an incredibly compelling argument:

• Compute scales exponentially

• The curve holds across 100 plus years

• By 2029, commodity compute will surpass the roughly 10 to the power of 16 FLOPS needed to match the human brain’s functional capability

Kurzweil has correctly predicted:

• the rise of smartphones

• the internet

• machine vision

• speech recognition

• human-competitive AI

• LLM-like systems decades before they arrived

So until proven otherwise, I’m sticking with his 2029 forecast. It’s a clean midpoint and the most coherent argument I’ve heard.

If AGI is truly arriving this decade, then we are not in a bubble. In fact, the returns from the AI build-out might be far greater than any analyst is modelling today.

Trying to forecast the AI TAM in 2030 today feels like trying to forecast the internet in 2000. Would anyone have predicted trillion-dollar app stores? Or a trillion-dollar digital advertising market? Or social networks bigger than the GDP of many countries?

Of course not.

So no, I don’t think we’re in an AI bubble. I’m comfortable not knowing the exact TAM in five years. What I do know is that the one input everything requires is power and compute. Nebius sits directly at the heart of that demand.

Some companies may be richly valued. Some may never reach their trillion-dollar dreams. But AI itself? It is nowhere near done. And Nebius, in my view, is positioned exactly where I want to be, as long as they keep executing.

Conclusion

To wrap things up, Nebius’s Q3 results tell a very clear story: this is a company experiencing demand that most businesses could only dream of. Every watt of active power is sold before it even comes online. The product is strong, the pricing is aggressive, the growth is exponential, and the long-term tailwinds from AI’s trajectory are powerful enough to carry Nebius for years to come.

Yes, it’s capital intensive. Yes, the dilution isn’t pretty. And yes, the headline numbers can be jarring if you’re only looking at this quarter in isolation. But when you take a step back and frame Nebius as a foundational AI infrastructure provider, potentially one of the defining players in the AGI era, the picture looks very different.

I’ll be going into much more detail on this in the next article, where I’ll share the full valuation frameworks I’ve been building. I’ll walk through the unit economics, the power-to-revenue mechanics, the long-term earning potential, and the scenarios that justify both the upside and the risks. There’s a lot to unpack, and the Q3 results finally give us enough clarity to do it properly.

As a small spoiler alert: yes, despite the volatility, dilution and market reaction, I still believe Nebius is undervalued, and I think it remains an excellent opportunity for long-term investors who understand what’s being built here.

More to come soon.

The information provided in this article is for informational and educational purposes only and does not constitute financial, investment, or professional advice. The author is not a licensed financial advisor, and the content is based on publicly available data, personal analysis, and opinions.

Since launching this blog, I’ve had some success hunting for companies the market has given up on but shouldn’t have. Companies that under the hood seem to be doing just great, but at the time appear to be universally hated and sometimes even written off as ‘dead’ companies.

Performance so far is below, which includes a few of those: Uber, Google, Novo Nordisk (which is taking a little longer than expected to prove the doubters wrong) and AMD.

PayPal caught my eye precisely because sentiment has been so dire

The stock’s down 77% from its 2021 peak of $310. Over the last three years, it’s slid roughly 65%. The narrative? PayPal’s a pandemic winner turned loser, getting eaten alive by Apple Pay, Stripe, and every new fintech startup with a slick app.

But here’s what made me look closer: PayPal has continued to show it can innovate, it’s core metrics are accelerating, and new CEO Alex Chriss has been impressing me as each quarter passes.

Today it was announced that PayPal is becoming the first payment rails for ChatGPT. The one with 800 million weekly users. The stock is only up 4% on the news, at a price of $73. Yet it’s still trading at a forward P/E of around 12.

This is exactly the kind of disconnect that gets my attention. Is PayPal really doing that badly?

My thesis is that PayPal is not just alive and well, but is about to start retaking market share, continue to accelerate its core metrics, and importantly - capitalise on AI.

The Funeral for PayPal was Premature

Let’s address the bear case head-on. Critics point to slowing growth post-pandemic, fierce competition, and questions about relevance in a mobile-first world. Fair points. PayPal’s active accounts barely budged (up 1% to 438 million), and the glory days of 20%+ growth are gone.

But Q3 2025 results tell a different story than the obituaries suggest:

• Revenue grew 7% to $8.4 billion

• Total payment volume hit $458 billion, up 8%

• Free cash flow reached $1.7 billion ($2.3 billion adjusted)

• Non-GAAP EPS guidance raised to $5.35-$5.39 (15-16% growth)

These are not the numbers of a dying business. They’re the numbers of a mature company finding its footing after a rough patch.

More interesting is what’s happening under the hood. Venmo revenue surged 20%. Buy Now, Pay Later volume jumped 20% with 80% higher average order values. The Braintree processing business stabilised after client renegotiations.

I’ve been pay close attention each quarter since Alex Chriss took over as CEO, and I am finally confident he is executing on the turnaround he promised.

Another echo of PayPal’s death rattle came through the launch of an ads business. In October 2025, they rolled out PayPal Ads Manager, letting small businesses create targeted ads using PayPal’s transaction data.

Think about what PayPal knows: What 400+ million people buy, when they buy it, how much they spend. That’s a pretty good start. Amazon built a $50 billion ad business on similar data. PayPal’s just getting started, but it’s already live in the UK with U.S. expansion underway.

This isn’t core to my investment thesis, but it’s the kind of optionality I am interested - high-margin revenue streams that could surprise to the upside and are completely untapped. It’s essentially an additional margin of safety - much like the two themes I’ll speak about next.

The Real Story: Becoming the Payment Layer for AI

I’m not buying PayPal for medicore revenue growth, or just new advertising revenue. I’m buying it because it’s positioning itself at the intersection of two massive shifts: AI-driven commerce and the death of the traditional checkout.

Here’s the future PayPal is betting on, and why I think they’re right:

Today, buying something works ok. You search Google, compare products across sites, read reviews, add to cart, enter payment details, wait for confirmation. It takes 10 minutes to hours.

Tomorrow? You’ll tell your AI assistant what you want, and it will go buy it for you.

PayPal just became the first integrated payment method in ChatGPT. When OpenAI’s users want to buy something an AI recommends, PayPal handles it. They’ve also partnered with Google’s Gemini (450 million monthly users) and Perplexity AI. Combined, we’re talking about 1.25 billion potential users who could discover PayPal through AI interactions.

CEO Alex Chriss put it perfectly: “PayPal will power payments that help people go from chat to checkout in just a few taps.”

Running the Numbers on AI Commerce

Let’s be conservative. Assume just 1% of those 1.25 billion AI users make one $100 purchase via PayPal in Year 1. That’s $1.25 billion in additional payment volume. At a 1% take rate, it’s only $12.5 million in revenue - which is barely a rounding error.

But that misses the point entirely.

The real value isn’t the initial transaction. It’s the flywheel effect. Users discovering PayPal through AI chats start using it elsewhere - for Venmo transfers, branded checkouts, peer-to-peer payments. One touchpoint becomes many. New users become active users. Active users become power users. Power users tell every man and his dog to use PayPal because its the best service for online payments.

If this scales to 5% adoption with multiple transactions, we are talking $100-500 million in additional annual revenue as AI commerce matures. But that’s just in the near term - the total addressable market for “agentic commerce” could reach $1-5 trillion by 2030 according to some projections.

I have no idea how high PayPal’s share will be, and forecasting at this stage is almost futile, but sometimes a qualitative approach is helpful to forecast the future state:

• PayPal has the first mover advantage for AI payments

• PayPal is creating the standards for AI payments

• PayPal has a CEO who appears laser focussed on capturing this opportunity

• PayPal isn’t just participating in this shift. They’re helping define it through the Agentic Commerce Protocol (a new standard for AI payments) - essentially becoming the payment standard for how AIs transact.

Given the above, if PayPal’s core business is fairly valued, the AI optionality makes the investment worthwhile.

My Valuation: Conservative Base, Explosive Upside

Base Case (The Boring Scenario)

Assume PayPal muddles along with 5% annual growth for the next 5 years, steady margins around 15%, and continues its aggressive buybacks at the same rate. No real AI opportunity is captured and adverts don’t really take off.

Using a DCF with:

• 5% CAGR

• 10% WACC

• 3% terminal growth rate

• Terminal P/E of 18

Fair value = $100 per share

That’s 37% upside from today’s closing price - given that PayPal has announced a dividend, and I believe its a low risk stock, this starting point is appealing.

Bull Case (AI and Advertising Capitalisation)

Now assume a return to double-digit growth driven by ads and AI services. Buybacks continue alongside a small dividend. Margin growth expands to 16.5%.

Using a DCF with:

• 12.5% CAGR

• 10% WACC

• 3% terminal growth rate

• Terminal P/E of 20

Fair value = $148 per share

That’s an upside of 103%.

Why I’ve Finally Pulled the Trigger

I’ve watched PayPal for over a year, waiting for the right entry point and clearer signs of turnaround. The Q3 earnings and OpenAI partnership give me confidence that we are about to re-enter calmer waters for PayPal.

My upside adjusted valuation model (UAV) ranks PayPal as a top pick after Q3 earnings, which warranted me upgrading the business fundamental metrics and amending the valuation.

At a forward P/E of 12, the market’s pricing PayPal like a melting ice cube. But between the core business stabilisation, AI commerce positioning, advertising potential, and their PYUSD stablecoin (not touched on today, but I’ll be sure to follow up on this in more detail), there are multiple ways to win.

The downside feels limited - the business generates serious cash, buys back stock aggressively, and just initiated a dividend.

The upside - If AI commerce takes off and PayPal is at the forefront, we are currently undervalued by at least 50%.

The market thinks PayPal is dead money. I think it’s healthier than it has been in a long time, and has a decent probability of becoming a leader in the technologies which are going to shape online payments over the next 5 years.

The information provided in this article is for informational and educational purposes only and does not constitute financial, investment, or professional advice. The author is not a licensed financial advisor, and the content is based on publicly available data, personal analysis, and opinions.

Since launching this blog in June of this year, my analysis has continued outperforming the broader market. The S&P 500 has returned 13.6% over this period, while my published stocks have delivered an average return of 40.2% in just 4 months.

In this article I will review the following:

1. Quick update on all the companies I’ve covered so far

2. An update on my UAV watchlist, and any stocks of note towards the top of the table

3. General thoughts on the market and my approach over the next year

Company Updates

Nebius

Lets start with the elephant in the room.

My initial analysis for Nebius on June 12th included the following breakdown, which looked at all of Nebius’s sub units, cash, investments and core business offering.

For the core ‘Nebius AI’ business offering I used a price to sale multiple of 12 against Revenue of $1 billion for FY 2025.

Revenue is now likely to be quadruple that in FY 2026 - and looks likely to continue to accelerate. The stock price has increased by 145%.

I revalued the stock based on accelerating revenue off a new annual run rate base of $5 billion for FY2026. Getting a per share price of $180 - I’m happy with this figure as my starting point.

I am no longer including the sub units or investments in my valuation, but this helps to ground the valuation and adds in a margin of safety by default if my revenue assumptions are too optimistic. When a company accelerates this quickly, it pays to use more flexible valuations with a large margin of safety.

For now there is plenty of juice in the price, and I am happy holding for the next year or so without trimming to find out how the story plays out and if my valuation of $180 can be realised.

Novo Nordisk

Novo Nordisk has had a tumultuous time since entering my blog, suffering from profit warnings, revenue drops, a change in CEO and a breakup with Hims & Hers.

But - the price is only 11% down from the initial coverage date. I think this reflects how oversold Novo Nordisk has become, and represents exhaustion of sellers.

My valuation and projections for Novo Nordisk are incredibly conservative, yet still offer plenty of upside.

There are a number of catalysts that could reduce the negativity of this business:

• Expanding GLP-1 Market Demand: The global market for GLP-1 drugs is booming due to rising obesity and diabetes rates. Novo Nordisk’s drugs like semaglutide (the active ingredient in Ozempic and Wegovy), could see sales surge as more people seek these treatments. Research suggests this market could grow to over $100 billion by 2030, giving Novo plenty of room to capture more share.

• New Drug Approvals and Pipeline Progress: Novo is advancing oral versions of GLP-1 drugs (pills instead of injections) and combinations for other conditions like heart disease or Alzheimer’s. Recent approvals, such as expanded uses for Wegovy in cardiovascular health, could open new revenue streams if trials continue to show positive results.

Ultimately this is a great value stock which is decoupled from the euphoria and mania of the AI market, and I think GLP-1 as a market is simply having a breather before its next run up. Competition in this market is strong, but so is Novo’s pipeline and I fully expect them to punch back.

Hims & Hers

Hims & Hers has experienced significant volatility since the blog’s launch in June, marked by a sharp stock plunge following the breakup with Novo Nordisk over compounded GLP-1 drugs, a brief revenue miss in Q2, and ongoing shifts in its weight loss offerings.

However, the price is only down about 4% from mid-June levels around $57, having recovered strongly from the post-breakup low near $42. This rebound suggests the market has digested the negatives and is pricing in robust growth, with shares now trading around $55.

My initial valuation was fairly conservative at $78 however I have reduced this to $70 following a breakup with one of the leading GLP-1 manufacturers (NVO).

Hims operates in a potentially grey regulatory environment with it’s use of modified GLP-1 drugs, and this regulatory risk weighs down the valuation and prospects of the business.

Ultimately, I mostly sold Hims at around breakeven purely because I felt there were better opportunities in the market. Which is a philosophy I try and maintain.

That’s not to say I won’t be buying Hims, but I’d like to see either the price dip, or the business outlook improve before I want to think about buying again.

AMD

Yesterday, OpenAI penned a deal with AMD to spend tens of billions in inferencing chips. The deal involves some potential share dilution, and there are worries about where this money is coming from - but ultimately I believe this is the start of AMD having its moment at the AI party.

My initial analysis of AMD was purposefully bullish to consider the art of the possible, with very little margin of safety built in.

For the purposes of my model, my updated slightly more conservative valuations are below:

1 year base case: $215

2 year upside: $450

I’m happy to continue holding and see if AMD has a chance at capturing its upside valuation.

Google

Since posting about Google in June, the stock went from being a ‘dead’ company that was about to be eaten by AI, to having the best AI model and proving their AI roadmap is ready to penetrate global enterprise at scale.

My initial valuation was $190, with an upside of $275 - as of today we have almost claimed the upside target, showing a dramatic turnaround in a short space of time.

I’ve since re-rated Google based on stellar earnings and proven ability to lead in the AI race, and there is plenty room to grow.

While I trimmed Google recently following it’s sharp rise, I still think Google is well positioned to beat expectations over the long term.

Uber

Uber continues to execute, and I continue to not be scared of Tesla’s world domination.

As expected back in June, it wasn’t a case of Tesla turning on Full Self Driving, and them taking over the taxi market by nightfall.

I expect Tesla will continue to move slowly and carefully in this world, whilst of course making some ground. Meanwhile, I expect Uber to continue executing it’s core offerings, venturing further into new travel markets such as flights and trains, and continue to invest in the autonomous vehicle revolution.

Uber is still well positioned to succeed and become the agnostic ride hailing platform in the world of driverless cars. This is another company I am looking forward to holding for a very long time.

No changes to the thesis just yet and Ubers stock remains undervalued in my opinion.

Games Workshop

Newest addition to the fold.

No update from the initial analysis just yet, with the next catalyst likely the winter trading update, and keeping a close ear to the ground for updates on the Warhammer TV adaption.

Current UAV model

One of the key drivers for making my investment decisions is my Upside Adjusted Valuation Model.

It contains a watchlist of around ~80 stocks, all of which I have shown some interest in, with a rough valuation - and importantly, ratings of the business metrics on the right hand side.

The high level rationale for this approach is designed to avoid two specific bad behaviours:

1. Selling too soon

   • By making business metrics (Moat, Balance Sheet, Management, Narrative) part of my valuation criteria, it means high quality companies still get a good UAV score even if they are ‘overvalued’.

   • Valuation is an art not a precise science, and this model is designed to include more qualitative data in the decision making process, and not get shaken out of great companies because your valuation figure has been exceeded.

2. Ensure I’m always buying quality companies

   • Often, the companies that seem like the best value are the companies nobody wants to buy. Sometimes, this is for good reason.

   • By not combining metrics that denote the quality of the business into an investment decision, you can fall into the trap of buying a cheap company that is not a great business. I believe this is a high risk trap to fall into when it comes to longer term investing.

If those two principles are adhered to, along with a good sense of base case valuation and an eye to what the upside looks like, I believe this forms a great framework for picking winners.

Four of my analysis picks are still in the ‘Green’ category which is generally my green light to buy if I so wish to.

AirBnB and Paypal are two companies I have been following closely. Paypal in particular is starting to show signs of life in its business turnaround, with clear avenues for restoring growth and some fundamentals already improving.

Mercado Libre and London Stock Exchange are notably high up the list, but they would need to reach a UAV of 80+ before I could consider buying them, as they are not part of my top 20 stock-list. More on that below…

Top 20 Stocks Ordered by UAV

As mentioned, I have 2 core principles of not selling great companies too soon and not buying poor companies just because they are cheap.

To bolster those principles further, the following stocks gain a slight weighting above all other stocks in my watchlist.

These companies all have a combination of a strong narrative (one or more strong catalysts for growth with high likelihood over the next 2+ years) and strong business fundamentals.

The idea is that I’m less likely to want to impulse sell or be shaken out of any of these investments, as long as I pay the right price. These companies form the basis of the companies I follow closely, and I think they’re all worth adding to your watchlist for their own reasons.

State of the Market and Approach over the Next Year

My Upside Adjusted Valuation Model will continue to form the basis for my decisions moving forward.

If there are companies with great upside and very strong fundamentals, I will continue to buy them.

I will not sell good companies just because they have run up in price. I will however look to trim or exit if the price has little upside left in it over a 2 year period, or the business fundamentals start decreasing. Both of these criteria will result in a stock dropping down my watchlist, and turning amber. (see Palantir below)

Aswath Damodran and the UAV Indicator

Aswath Damodoran (one of my favourite investing minds) posted the following Substack article in the week on the state of the market which I thoroughly enjoyed reading. I would highly recommend spending some time digesting it too.

He shows the following illustration to highlight poor and strong arguments for suggesting the market is overvalued:

• Just because stocks have gone up a lot in a short space of time does not mean its time to sell

• They might be going up for a reason

• The true test of whether stocks are overvalued is when they rise drastically faster than their cashflows, that is when euphoria and speculation is too heavily baked into the stock market.

The thinking is that by utilising conservative valuations, and using my UAV tracker as a tool to shortlist companies for further research, I should have a ready made ‘overvalued’ tracker which I can share as part of this blog.

When very few companies are undervalued on my watchlist, or when much of my watchlist turns amber, I believe that will be a strong indicator to suggest the market is ready to turn.

For now we have:

• 9 Amber Stocks

• 8 Red stocks

As part of more continuous appraisal of the general market, I’ll keep checking in on the number of overvalued stocks in my UAV watchlist. I expect it to rise significantly over the next couple of years but lets see.

The information provided in this article is for informational and educational purposes only and does not constitute financial, investment, or professional advice. The author is not a licensed financial advisor, and the content is based on publicly available data, personal analysis, and opinions.

Maximum upside so far is 145%, courtesy of Nebius and maximum downside of just 11.4% is thanks to Novo Nordisk. I am still incredibly happy with Nebius and Novo Nordisk’s valuation - while I am less confident in Hims going forward.

I’m treading carefully here, but I still see opportunity. Nebius continues to have significant runway, and the picks so far span different industries with natural diversification built in. I’m holding all positions while evaluating new names to build positions in if I trim winners as part of my risk management strategy.

Soon I’ll post my updated analysis and UAV model on all of the companies I track, but this article is about my comfort blanket stock - completely unrelated to AI or any other hot industry - which I’m slowly rotating profits into.

The UK Comfort Blanket Stock: A Business Built to Last

Games Workshop (GAW.L) has quietly delivered one of the most impressive long-term track records in UK market history. Since 2003, the stock has returned over 15,000%, turning £10,000 into £1.5 million.

The business model is deceptively simple: create miniature figurines for tabletop wargaming, sell paint and modelling supplies, publish rulebooks, and operate retail stores.

The Warhammer universe - spanning Warhammer 40,000’s grim future and Warhammer Age of Sigmar’s fantasy setting -has cultivated one of gaming’s most devoted fanbases. Players don’t just buy pre-painted figures and play. They spend hundreds of hours assembling, painting, and customising their armies.

Yes thats right, Games Workshop sell the miniatures and get consumers to paint them themselves.

Recent financials prove the model’s strength. Revenue hit £617.5 million in the latest fiscal year, up 17.4% year-over-year. But here’s what matters: this wasn’t driven by discounting or margin compression. Operating profit margins remain north of 30%, and the company generated £197 million in profit.

Core tabletop revenue grew 14.2%, proving the traditional business remains healthy. But licensing revenue exploded 81%, primarily from Space Marine 2 - a console game that became one of 2024’s surprise hits. This is where the story gets interesting.

Current valuation sits at a P/E of roughly 24 with a 3.5% dividend yield. For a company with complete IP ownership, proven supply chain management, and gross margins most businesses dream about, this is entirely reasonable as a starting point. My fundamental thesis is that the Warhammer universe has untapped potential, that could very soon start to be realised.

The Licensing Trojan Horse

Most analysts fixate on Games Workshop’s licensing revenue growth. They see Space Marine 2’s success (a recent console video game set in the Warhammer universe) translating into royalty payments and call it a day. This misses the bigger picture entirely.

Here’s how licensing works: Games Workshop owns the Warhammer IP outright. When external studios want to make video games, films, or TV shows using Warhammer characters and settings, they pay Games Workshop. This comes in two forms - upfront licensing fees and ongoing royalties based on sales or revenue. Space Marine 2 generated substantial royalty payments as the game sold millions of copies, which is why licensing revenue jumped 81% last year.

But licensing revenue isn’t just about the direct payments. When Warhammer IP reaches new audiences through video games, TV shows, or films, it functions as a massive advertisement for the core tabletop business. New players discover the universe exists. Lapsed hobbyists remember why they loved it. The community expands, and with it, tabletop sales.

The real prize is what happens when millions of people encounter Warhammer through a screen and a fraction of them decide to explore the tabletop hobby. That’s where the economics get genuinely interesting - these become high-lifetime-value customers who spend thousands over years on miniatures, paint, books, and hobby supplies.

The evidence across similar franchises is interesting to say the least:

The Witcher:

Saw a 554% surge in U.S. physical game sales in December 2019 when Netflix’s show premiered. This was years after the game’s original release—the show introduced an entirely new audience to a franchise they’d never heard of. CD Projekt Red reported The Witcher 3 reached 50 million copies sold by 2020, with a significant portion attributed to the Netflix effect. The show completely resurrected sales of a five-year-old game.

The Last of Us:

Experienced similar dynamics. The HBO show drove The Last of Us Remastered to 1.8 million monthly active users in January 2023, representing over 1,000% month-over-month growth. Sony reported the franchise had sold 37 million copies across all titles by March 2023, up from 32 million pre-show. Again, this was for games that had been out for years - the adaptation brought dormant IP back to life.

Fallout:

Might be the most relevant comparison. Amazon’s TV adaptation generated an estimated $80 million in incremental revenue for Bethesda from April to September 2024, including roughly $30 million in licensing fees. But here’s what matters: week-on-week sales of Fallout 4 spiked by up to 7,500% in Europe during the premiere week. Fallout 76, a game that had been widely criticised and largely abandoned, saw player counts surge 600%. The show didn’t just make money from royalties - it resurrected an entire franchise and drove sales across multiple titles spanning a decade of releases.

Lord of the Rings and Potter:

Even book-to-film adaptations follow this pattern. Lord of the Rings trilogy drove book sales from roughly 50 million copies pre-films to over 150 million by 2007. Harry Potter saw similar spikes, with book sales jumping 700% in the year following the first film’s release despite the books already being bestsellers.

So what does this mean for Games Workshop?

The common thread? Rich fictional universes with deep lore, interesting characters, and world-building that invites obsession. Warhammer has this in spades - arguably more than any franchise I’ve listed. The 40K universe alone spans 30,000 years of fictional history across thousands of books, with a darkness and complexity that’s genuinely unique in popular fiction.

The question becomes: what happens to core revenue when Warhammer gets its mainstream moment through Amazon’s series? If even 1% of viewers who’ve never encountered Warhammer before decide to try the hobby, we’re talking about hundreds of thousands of potential new customers entering a high-lifetime-value business. And that’s before considering lapsed hobbyists who get pulled back in, or existing players who decide to start new armies.

The licensing revenue growth grabs headlines. The core revenue growth that follows is where the real money gets made.

The Obvious Risks

The single biggest risk is Amazon botching the adaptation. If the show disrespects the source material, review bombing and fan backlash could turn this catalyst into a disaster. Games Workshop’s relationship with its community could sour, and the boycotts could be real.

I’m less worried about this than most would be. The Rings of Power failure provides a clear template of what not to do, and Cavill’s involvement provides genuine insurance. He has too much to lose personally and professionally to attach his name to something the community will reject.

Beyond the show, Games Workshop faces the standard risks of any consumer discretionary business. Economic downturns hit hobbies first. Supply chain disruptions could pressure margins, though the company navigated COVID-era challenges well. Competition exists, though no competitor has IP remotely comparable to Warhammer’s depth.

Currency exposure matters - Games Workshop reports in pounds but operates globally. Strengthening sterling could pressure reported results, though this is a temporary accounting issue rather than a fundamental problem.

The valuation risk is real but manageable. At a P/E of 24, the market is pricing in continued growth. If the TV show flops and growth stalls, downside exists. But the business quality and existing moat provide significant protection - this isn’t an AI infrastructure stock with a risk of not being around in 5 years time.

Valuation: Calculating the Upside When Warhammer Goes Mainstream

If you go looking at Games Workshops balance sheet, you’re not going to find too many issues to save you a job - it’s a very well run company with minimal dilution and enough cash to weather a storm.

Current revenue sits at £617.5 million with profit margins of 31.8%. The stock trades at a P/E of roughly 24. Below are my two scenarios and valuations.

To keep this simple, and as conservative as I can, I have assumed no dilution but also have excluded dividends. Todays dividend yield is 3.5% - how this would evolve if we hit our revenue targets below, I don’t know. To ensure there is a sufficient margin of safety and to understand if this company is undervalued based on my scenarios, I’m seeing them as a nice optional bonus rather than a reason to buy.

Base Case: The Show is Good, Not Great

Assume the Amazon series succeeds but doesn’t become a cultural phenomenon. It attracts viewers, reviews are positive, and the fanbase expands modestly.

• Revenue grows 15% annually for five years.

• Net profit margins expand slightly to 35% as higher-margin licensing revenue increases (note: this is a slight expansion vs. current levels).

• Terminal P/E of 25, reflecting the proven track record and expanded distribution.

Using a 10% discount rate, this values Games Workshop at approximately £205 per share—roughly 43% upside from current levels around £143.50.

Bull Case: Warhammer’s Mainstream Moment

Assume the show becomes a genuine hit on par with The Witcher or Fallout. It drives sustained interest in the franchise, multiple seasons get greenlit, and the community expands dramatically.

This scenario assumes the Amazon show drives not just licensing revenue but substantial conversion of new customers to the high-margin core tabletop business

• Revenue grows 30% annually for five years (most of this revenue arrives in years 3 - 5, but it is a strong new cycle of growth from new and existing Warhammer enthusiasts).

• Net profit margins expand to 35% from licensing growth.

• Terminal P/E of 30, still below the company’s historical average but reflecting the new growth trajectory.

Under these assumptions, valuation reaches approximately £453 per share—216% upside from current levels.

The base case alone justifies my position. The bull case represents stellar, yet admittedly optimistic, returns if Warhammer achieves the strong penetration its IP quality deserves.

30% is not impossible if we look at our comparison examples, which could suggest years where revenue hugely increases. But it’s important to note how tough this would be - it’s a true bull case where the TV show is a huge hit.

So What Now?

Games Workshop sits at an inflection point. The business is already excellent—strong economics, loyal customers, pristine balance sheet. The upcoming Amazon series represents optionality that the market hasn’t fully priced in.

The best thing about this opportunity is that it is some way off. We have absolutely no announcements on the timings for when this TV show is coming. We could get an update any month now, and when we do the excitement about this proposition is likely to start baking itself into the stock valuation.

The status of the project is kept tightly secret. But speculation suggests the earliest we might see the project on Amazon is late 2026 (I think this is likely aggressive, and some time early to mid 2027 is more likely.

At a P/E of 24, the valuation is reasonable even without the show. With the show, it’s potentially cheap. The risk-reward strikes me as compelling, particularly with Henry Cavill’s involvement reducing the single biggest risk factor.

I’m a long term holder in Games Workshop but will be continuing to buy shares over the next year and today’s closing price is an attractive valuation. In a world of AI euphoria, this is a sleeping giant that adds fantastic diversification.

I will be adding this stock to my list of undervalued picks to begin formally tracking as part of this blog.

The information provided in this article is for informational and educational purposes only and does not constitute financial, investment, or professional advice. The author is not a licensed financial advisor, and the content is based on publicly available data, personal analysis, and opinions.

Since the inception of this blog, the S&P500 has returned 9.8%.

Since 2nd June, the analysis I have published, and their indicative purchases have outperformed the market:

• Average return of 19.6%

• Maximum downside of 19.4%

• Maximum upside of 87.7%

I’ve published analysis when I have been purchasing shares, and believe a company is sufficiently undervalued. The last few months have been quiet in terms of my output, but in all honesty I haven’t made too many moves during the summer months, instead opting to stick with my high conviction analysis.

But for now… my main focus is on the recent incredible Nebius news and what it means for the company. Lets go through:

1. Deal summary

2. The Importance of Microsoft

3. State of play for Nebius

4. An updated valuation

So What Has Happened?

On September 8, 2025, Nebius, through its U.S. subsidiary Nebius, Inc., locked in a five-year commercial agreement with Microsoft to deliver dedicated GPU-powered capacity from a spanking-new data center in Vineland, New Jersey. T

his isa $17.4 billion commitment over 5 years, with the potential to climb to $19.4 billion if Microsoft requires extra services (I think incredibly likely they will). The rollout kicks off later this year, pumping out an average of $3.48 billion annually. Some of this revenue is likely already included in the $1bn ARR already forecasted, but I expect the vast majority of it is not.

This deal is Nebius handing Microsoft the keys to a high-tech engine room, purpose-built for AI workloads like Azure’s cloud services and OpenAI’s ChatGPT. The company has $2.4 billion in cash and zero debt, but it’s eyeing debt financing secured by Microsoft’s rock-solid credit to juice up the buildout. Expect more financing details soon as they aim to scale faster than originally planned.

Nuts and Bolts of the Deal

Nebius will deploy this GPU capacity in stages, and the deal hinges on meeting deployment timelines. Miss a deadline after a grace period, and Microsoft can ditch that tranche unless Nebius offers a workaround. Both sides can bail if the other breaches (with a 60-day fix window) or goes bust. It seem like this is all fairly standard stuff here: service commitments, late-delivery penalties, and legal safeguards - I don’t think these terms should be flagged as specific risks.

The cash flow -$3.48 billion a year on the base deal -will fund the hefty capex, estimated at over $2 billion annually.

At this stage, Nebius has not achieved the economies of scale needed for us to start thinking too hard about profit margins, but that may start changing quickly. If $3.5 billion is now our base revenue figure for the next 5 years, $5 billion will come quickly.

Approaching $1 billion in free cash flow is not out of the question before 2030.

Taking On The Final Boss

There are customers, and then there's Microsoft. I had to double take when reading this news.

Landing the world's largest software company as your anchor tenant isn't just a commercial win - it's a credibility transformation that opens every other door in the industry.

Microsoft's AI infrastructure demands are enormous. They're powering OpenAI's ChatGPT, Copilot across Office 365, and Azure's expanding AI services. These are mission-critical applications serving hundreds of millions of users. Microsoft choosing Nebius means their specialised GPU clusters aren't just competitive with hyperscaler alternatives, but they're actually preferred.

The risk profile changes fundamentally when your largest customer is also the most sophisticated buyer in the market. Microsoft's due diligence process is exhaustive - they're not betting $17.4 billion on unproven technology or questionable execution capability. This is validation at enterprise scale.

What Changed:

• Revenue visibility extended 5+ years

• Technical capability validated by industry leader

• Competitive positioning strengthened through hyperscaler partnership

• Risk profile improved through enterprise customer diversification (albeit now concentration becomes a new risk to monitor)

What Remains:

• Exceptional growth trajectory

• Specialised experience and market position in AI infrastructure

• Primed to capitalise on AI inference, the theme I am most bullish on going into 2030

• European regulatory advantages

• Disciplined capital allocation with $2.4B cash, zero debt

Summary:

• Possibly the single biggest vote of confidence I’ve ever seen for a stock I own

• A huge step forward in the right direction

Updated Valuation

My current valuation is as follows.

• Current ARR: ~$3.5 billion (Microsoft deal baseline).

• Projected ARR (2025): $5 billion, assuming modest additional contracts + initial projections

• P/S Ratio: 10 (aligned conservatively with high-growth AI infrastructure peers like CoreWeave).

• Market Cap: $5 billion × 10 = $50 billion

• Per share ~$180

The sum of parts approach I adopted previously is almost no longer worth doing, the extra ‘parts’ will soon be completely dwarfed by the main AI business.

My strategy here is to keep refining this valuation as we get more evidence. A company that is growing so quickly, with such high uncertainty on the financials does not warrant a breakdown like I do with other companies. For now, my plan is to watch closely.

• As revenue keeps growing, there will be more and more eyes on profitability. Right now that is not the focus, but we should keep an eye on gross margins and what the net income margin might be in 2030.

• For now my thesis continues to be the $1bn in free cash flow target, how quickly that happens I don’t know but when it’s possible - the $100 billion valuation is well in sight.

For those interested, my UAV model right now is as follows:

Recent changes:

• Moat, Management and a re-rating on Nebius’s valuation means it is the number 1 pick on my watchlist right now.

• I’ll be covering the other names close by in future articles.

The information provided in this article is for informational and educational purposes only and does not constitute financial, investment, or professional advice. The author is not a licensed financial advisor, and the content is based on publicly available data, personal analysis, and opinions.

This article breaks down my $295 base case and a $500 bull case for AMD’s stock, fueled by skyrocketing revenue, inc…

Read more

Key Points

• The Hims-Novo Nordisk partnership collapse in June 2025 caused a 35% drop in Hims stock, signalling risks in its GLP-1 strategy.

• I have cut Hims’s fair value from $100 to $60, reflecting a shortfall in 2025 GLP-1 revenue, litigation risks, loss of branded drug access, and reputational damage.

• The GLP-1 market could reach $471.1 billion by 2032, driven by obesity’s $173 billion annual U.S. cost and growing demand for effective treatments.

• Novo Nordisk and Eli Lilly, with 83% market share, are the dominant players; Novo offers 63.33% upside, Lilly 26.21%, per my valuation framework.

• Both NVO and LLY are highlighted as buys in my Upside Adjusted Valuation model

• Novo’s CagriSema trials, though initially disappointing, could still enhance its portfolio with further development. Eli Lilly has not won the war yet.

Why GLP-1 Is a Must-Watch

GLP-1 drugs like semaglutide (Novo’s Wegovy) and tirzepatide (Lilly’s Mounjaro) are transforming obesity treatment, a market with massive potential in the early innings of its disruption.

Obesity affects 42% of U.S. adults, costing $173 billion yearly, and GLP-1s deliver 15-20% weight loss, reducing risks for diabetes and heart disease. Many analysts expect this market to exceed $100 billion by 2030, with some estimates as high as $471.1 billion by 2032.

There is no doubt that GLP-1s have the potential to be a multi-decade trend with durable growth, understanding which companies are positioned the best, and which are currently the cheapest is the aim of this article.

The Hims-Novo Fallout Explained

In April 2025, Hims & Hers partnered with Novo Nordisk to sell Wegovy, a leading GLP-1 drug, through its telehealth platform at $599/month, aiming to broaden access. This was a key driver for me purchasing the stock.

Earlier this week, the deal collapsed. Novo accused Hims of illegally selling compounded semaglutide—(custom-made versions of the drug). Compounding is allowed during shortages but becomes illegal for drugs mimicking approved ones like Wegovy once shortages end, unless specific exemptions apply. Novo claimed Hims used unapproved ingredients, labeling them “knockoffs.”

Hims’s CEO countered, accusing Novo of stifling competition and defending their compounded drugs as clinically necessary.

The market reaction was harsh: Hims’s stock fell 35%, from $65.67 to $42.95, erasing $2.6 billion in market cap. Novo’s shares dropped 5.49%, while Lilly gained 1.04%, suggesting market share shifts.

Hims: A Riskier Bet

Hims’s telehealth platform, spanning weight loss, mental health, and dermatology, thrives on disruption, but the Novo fallout exposes its weaknesses. My Upside Adjusted Valuation (UAV) framework initially valued Hims at $100, driven by a $725 million 2025 weight-loss revenue target and 111% Q1 2025 growth ($586 million total revenue).

Post-fallout, I’ve cut the fair value to $60, reflecting a shortfall in GLP-1 revenue. At $41.41, Hims offers 44.89% upside, but the margin of safety is narrower - and its business looks more fragile.

Why Revenue Projections Dropped

Hims’s GLP-1 revenue, was a key growth driver. The fallout changes this:

• Litigation Risks: Novo’s accusations of illegal compounding could lead to lawsuits, fines, or operational restrictions, draining resources and capping growth for 18-24 months.- Novo Nordisk would not make these claims without sufficient evidence. I expect this will go to court, and I expect Hims will not come out on top versus this goliath of pharmaceuticals.

• Loss of Wegovy Access: Without Novo’s partnership, Hims lacks access to branded Wegovy, a top-tier GLP-1 drug. Alternatives like compounded drugs or generics (e.g., liraglutide in 2025) I believe are less competitive long term, as customers will prefer branded options for perceived safety. Especially if the accusations of Hims turn out to be completely true.

• Reputational Damage: The “knockoff” label hurts Hims’s credibility, potentially driving customers to competitors like Ro or LifeMD, who maintain compliant partnerships with Novo and Lilly.

• Competitive Pressure: Other telehealth platforms with superior access to branded GLP-1s can offer better pricing or reliability, eroding Hims’s market share.

I estimate Hims’s poorer outlook for GLP-1 segments will hurt revenue. Non-GLP-1 segments (mental health, dermatology) should continue to grow at 30%, but they can’t fully offset the impact.

Adjusted Valuation for Hims after the fallout

Using a 10-year DCF model (10% WACC, 5% terminal growth, net profit-based):

• Revenue Growth: 20% annually, reflecting GLP-1 headwinds but continued growth in other verticals.

• Net Profit Margin: 10% (down from 12%), due to legal and regulatory costs.

• Terminal P/E: 20x

• Fair Value: $60/share, implying 44.89% upside from $41.41 - still good value

Hims’s 2.4 million subscribers and new verticals (e.g., testosterone therapy) keep it attractive, but its low moat (6/10 in my UAV) reflects dependence on manufacturers. Customers will likely favour branded drugs as supply stabilises, and legal risks add uncertainty. This is definitely not the time to completely panic, but it’s long term outlook has taken a significant hit for now - management will have to work hard to get things back on track.

Novo Nordisk vs. Eli Lilly: The GLP-1 Titans

The outlook for Hims may take a while to recover, and things may get worse before they get better. I want to assess the positioning of Novo Nordisk and Eli Lilly to understand the opportunity they have in the GLP-1 megatrend.

Novo Nordisk and Eli Lilly control 83% of the $53.46 billion GLP-1 market, making them the core of this investment theme. Below, I compare their strengths, narratives, and valuations.

The Players

• Novo Nordisk: This Danish pharma leader pioneered GLP-1s with Ozempic and Wegovy, generating $35.4 billion and $14.3 billion annually, respectively. With 72% global market share and a $9 billion cash pile, Novo is a stable giant but faces supply and competitive challenges.

• Eli Lilly: A U.S. powerhouse, Lilly has surged with Mounjaro and Zepbound, capturing 53% of the U.S. market. Its $24 billion manufacturing investment and diversified portfolio (oncology, Alzheimer’s) make it a formidable challenger.

Narrative Over the Past Year

• Last year, Novo dominated GLP-1s, but supply shortages for Wegovy allowed Lilly to gain ground.

• Lilly’s Mounjaro, with 20.2% weight loss vs. Wegovy’s 13.7%, shifted U.S. market share.

• Lilly’s Zepbound launch and oral GLP-1 candidate (orforglipron) fueled optimism, while Novo’s stock took hits from compounding competition and the recent Hims fallout.

• Novo’s amycretin trials, showing 24% weight loss, signal a potential rebound.

Current Sentiment

I see Lilly as the market darling, with strong enthusiasm for its clinical edge and pipeline. Novo’s sentiment is more cautious—its fundamentals are solid, but Lilly’s gains and drug efficacy make it tough to split the two.

Outlook: Potential Catalysts

Both stocks have a narrative rating of 10 in my Upside Adjusted Valuation model

Novo Nordisk:

• CagriSema Development: Despite the initial 15.7% weight loss in recent trials, Novo is tweaking dosing and extending trial durations, aiming for 20-25% weight loss to rival Lilly’s best, potentially reversing market perception by late 2025 or 2026.

• Amycretin Approval: This oral drug, with 24% weight loss in early trials, could appeal to patients avoiding injections, boosting market share in 2026-2027.

• Wegovy Supply Stabilization: Resolving supply issues could push sales higher, building on Q1 2025’s $2.6 billion, with significant upside by 2026.

• Asian Market Expansion: Targeting rising obesity rates in Asia, where Wegovy and Ozempic could see triple-digit growth over the next two years.

• New Use Cases: Trials for addiction and heart disease could tap into GLP-1’s cardiometabolic benefits, adding billions in revenue by 2027-2028.

Eli Lilly:

• Orforglipron Approval: This oral GLP-1, matching injectable efficacy in Phase 3, could launch by mid-2026, capturing a large needle-averse segment and driving 20%+ revenue growth.

• Mounjaro/Zepbound Growth: With current quarterly growth rates of 140% and 111%, these drugs could double U.S. market share by 2027 if supply keeps pace.

• Donanemab Approval: Approval for Alzheimer’s treatment could add $5-10 billion annually by 2027, diversifying beyond GLP-1s.

• Eloralintide Development: Early 11.5% weight loss at 12 weeks suggests a new contender, potentially adding $2-3 billion by 2028 if trials succeed.

Outlook: Risks

Novo Nordisk:

• Lilly’s competitive pressure eroding market share.

• Persistent compounding pharmacy challenges.

• Supply chain disruptions delaying growth.

• Patent cliffs post-2040, though distant.

Eli Lilly:

• Pricing pressures from payers on high-cost GLP-1s.

• Safety concerns, like gastrointestinal side effects.

• Regulatory delays for orforglipron or donanemab.

• Emerging competition from new GLP-1 candidates.

Valuation Comparison

Using a 10-year DCF model (10% WACC, 5% terminal growth, net profit-based)

• Novo Nordisk:

  • Revenue Growth: 12.5% annually

  • Net Profit Margin: 34%, reflecting high profitability.

  • Terminal P/E: 20x, for stable cash flows.

  • Fair Value: $110/share, implying 63.33% upside from $67.35 ((110 - 67.35) / 67.35).

• Eli Lilly:

  • Revenue Growth: 20% annually

  • Net Profit Margin: 25%, slightly higher than historical averages.

  • Terminal P/E: 25x, reflecting a growth premium.

  • Fair Value: $1000/share, implying 26.21% upside from $792.3

Lilly’s clinical superiority and pipeline are impressive, but it is not cheap. It may pay to buy the market leader, but this is a fickle industry and Novo Nordisk could become more competitive very quickly.

Conclusion: Novo Nordisk is attractive, Lilly is the standout performer but not at this price - Hims’ situation could quickly get worse.

Hims’s innovation and platform is compelling, but its low moat and new legal uncertainties make it a hold - furthermore, I expect the fallout with Novo Nordisk to get worse before it gets better, and legal costs may drastically impact the bottom line.

• Both Novo Nordisk and Lilly are undervalued, with potentially more upside for Novo Nordisk at the current prices.

I personally own Novo Nordisk and will continue adding at the current price of $67.35.

Eli Lilly is hard to ignore, but offers a strong risk/reward at lower prices. I will be observing my Upside Adjusted Valuation tracker to spot when Lilly becomes a more obvious investment.

This is my analysis, not financial advice. I hold positions in Hims and Novo Nordisk. Do your own research.

Many years ago I wrote a dissertation on autonomous vehicle liability—I was annoyingly ahead of my time and nobody was too interested in it. Since then, the obvious AI connection has meant I've been tracking both spaces closely. GLP-1s are the newest theme on my radar, and it's only recently that I'm hearing about their profound results.

What I love about these three themes is they all share the same DNA as every transformative technology that analysts completely missed in the past.

The internet was "a fad." Mobile phones were "too expensive for mass adoption." Social media was "just for students." Every single time, the experts underestimated not just adoption, but the entire ripple effect of what these technologies would enable.

I believe we're seeing the exact same pattern with AI, GLP-1s, and autonomous vehicles. All three have the potential to transform everyday lives globally, and more importantly, I suspect analysts are still unable to project their true impact. The market is pricing in the obvious first-order effects while missing the wealth-creating second and third-order opportunities.

This article won't be a deep dive into each theme—if you need background, I've linked my detailed pieces below. Instead, I'm sharing my high-level assessment of company positioning and valuation using my UAV framework. Think of it as a roadmap and watchlist for some of the best risk-adjusted plays in each space, not a detailed business valuation exercise.

The companies scoring highest on my UAV tool are the ones that suggest the most compelling risk-adjusted returns as these themes play out over the next decade.

• UAV >80: Exceptional undervaluation, high strength business

• UAV 70-79: Strong opportunities with solid margin of safety

• UAV 60-69: Fair value and no significance margin of safety

I’m going to cover stocks from each theme, identifying which are of major significance and which are potentially undervalued.

I will only be using stocks that I track in my Upside Adjusted Valuation tool (current top 12 below). All prices as of 15th July 2025. As at time of writing, I expect the middle-east conflict to have increased the scores of all stocks when this is published.

Transformative Investment Themes: High-Conviction Plays in AI, GLP-1, and Autonomous Vehicles

Three key megatrends are emerging —autonomous vehicles, GLP-1 medications, and AI inferencing. With the potential to reshape industries , these aren't just buzzwords; they're the kind of game-changers that may go on to rival the internet or smartphones.

Using my Upside Adjusted Valuation (UAV) tool, I’ve zeroed in on ten stocks from these themes, with eight scoring above 60 and four above 70—signalling serious undervaluation and upside potential. The analysis below will consider these businesses further and assess how much they can impact from technological disruption over the next 5-10 years.

Autonomous Vehicles and Robotaxis: The Road to Riches

The robotaxi market is revving up, projected to rocket from $789 million in 2024 to a staggering $29.3 billion by 2030 (P&S Intelligence).

This is not everyday innovation, it has a chance at redesigning transport - if that happens, owning a piece of the new designers is likely to pay well. Here’s who’s leading the pack.

• Uber (UBER) | UAV: 86.5

  • Why?: Uber is the ultimate "go-to-market" player for robotaxis, albeit they abandoned building their own car some time ago. With 171 million monthly active users, it’s platform is ready to tap into safer and cheaper autonomous rides. Partnerships with 18 AV providers, including an exclusive Waymo deal in Austin, let Uber ride the wave without bleeding R&D cash.

  • If the likes of Waymo and Tesla want to go global with their autonomous vehicle fleets, they will need to adhere to thousands of local regulations, Uber can fast track that process.

  • Financial Firepower: Q1 2025 delivered $11.53 billion in revenue and $1.78 billion in net income—compare that to a $654 million loss last year.

  • As drivers get replaced by robots, Uber’s biggest cost vanishes.

  • Why Not?: Tesla and Waymo have the autonomous driving technology - Uber does not. If they decide to go alone and succeed, Uber could be toast.

• Alphabet (GOOGL) | UAV: 77.7

  • Why?: As it stands, Waymo is the tech leader, running 250,000 paid trips weekly across four cities with an 81% crash reduction versus human drivers.

  • A $5.6 billion funding round in October 2024 suggests this is not an insignificant business segment for Alphabet, and could begin to give revenue a boost in the near future.

  • The Catch: Expensive sensors and low-volume production are hurdles, and Waymo’s safety-first approach might slow market share grabs.

• Tesla (TSLA) | UAV: 68.2

  • Why?: Tesla’s betting big on vertical integration, with a robotaxi launch planned for Austin in June 2025. Its 6+ billion miles of driving data and manufacturing scale are unmatched.

  • Tesla’s sub $40,000 cars have the same autonomous technology, and the company has built their strategy with scale in mind.

  • The Risk: NHTSA probes and a camera-only approach keep it high-risk. Tesla’s head of self-driving admitted they’re “a couple years” behind Waymo, though cost advantages could help them scale faster.

• Nebius (NBIS) | UAV: 80.3

  • Why It’s Hot: Nebius’s subsidiary, Avride, is an emerging player in autonomous vehicles, with plans to have 100 robotaxis operational in Austin by year-end, positioning it as a direct competitor to Waymo. Avride’s recent partnership with Hyundai strengthens this credibility in the AV space.

  • Why It Matters: While Avride is a smaller part of Nebius’s AI-focused business, its AV ambitions add upside potential, especially with Nebius’s 100% ownership.

  • Why It Might Not Matter: This is a tiny part of Nebius’s business to date - is almost pre-revenue and likely to be bleeding cash. Placing any sort of meaningful valuation on this business segment is a risky game.

• NVIDIA (NVDA) | UAV: 71.2

  • Why It’s Hot: NVIDIA’s DRIVE platform is a beast, with automotive revenue hitting $1.69 billion in FY25 (+103% YoY) (CNBC). Its end-to-end solution locks in customers (NVIDIA).

  • Perspective Check: AVs are just 1.45% of NVIDIA’s Q1 FY26 Data Center revenue ($39.1B vs. $567M). It’s a side hustle, not its bread and butter.

GLP-1 Medications: Healthcare’s Next Big Thing

GLP-1 drugs are transforming diabetes and obesity treatment, with the market expected to balloon from $53.5 billion to $157.5-$268 billion by 2030-2035 (IQVIA). Here’s some main players to follow:

• HIMS & Hers (HIMS) | UAV: 84.6

  • Why It’s Hot: HIMS is riding the GLP-1 wave via telehealth, offering injections at $199/month versus $1,000+ for branded drugs. It raked in $225 million in GLP-1 revenue in 2024 with 100,000 subscribers, hitting a $100 million run rate in under a year. Total revenue was $1.48 billion (+69% YoY), with EBITDA jumping from $50M to $177M.

  • The Edge: Its 2025 generic liraglutide launch sidesteps FDA shortage risks, and telehealth’s convenience creates sticky customers.

• Novo Nordisk (NVO) | UAV: 70.8

  • Why It’s Hot: Novo owns the original GLP -1 Ozempic, plus a number of innovative improvements which are competing with the top drugs available - It’s a cash-flow machine with dividend appeal.

  • The Challenge: For now, Lilly’s tirzepatide is sharing a big piece of its lunch.

• Eli Lilly (LLY) | UAV: 64.5

  • Why It’s Hot: Lilly’s tirzepatide (Mounjaro, Zepbound) delivers 20.2% weight loss vs. 13.7% for semaglutide (Novo Nordisk), grabbing 53% market share in Q1 2024 (PharmaVoice). 2025 revenue guidance is $58-61 billion (+32%) (Lilly).

  • Why It Matters: Its premium valuation reflects clinical superiority, but Novo Nordisk has other drugs on the way and appears to be a significantly cheaper business at this moment. The tables could turn very quickly.

AI Inferencing: The Brain Behind the Future

AI inferencing, the art of making predictions and solving tasks with AI models, is set to grow rapidly over the next decade. If you think chatGPT generating you a simple website is cool, hang on tight for whats coming next.

Here’s some main players to follow:

• Alphabet (GOOGL) | UAV: 77.7

  • Why?: Alphabet’s Tensor Processing Units offer 4-6x cost efficiency, powering Google Cloud’s $11.4 billion Q3 2024 revenue (+35% YoY) with a 32.3% margin. AI integration protects its search moat and fuels enterprise growth.

  • Why It Matters: It’s a diversified bet on AI’s future, Google is covering a lot of bases with high conviction.

• NVIDIA (NVDA) | UAV: 71.2

  • Why?: NVIDIA’s Blackwell architecture delivers 30x better inference performance, with Q1 FY26 data center revenue at $39.1 billion (+73% YoY). Its $2 billion software run rate adds sticky revenue.

  • Other players such as AMD are emerging but right now Nvidia is still the leader - it may pay to stick with the current winner.

  • The Edge: Custom chip competition (Google, AWS) is under 10% market share, reinforcing NVIDIA’s dominance.

• Nebius (NBIS) | UAV: 80.3

  • Why?: Nebius is a powerhouse in AI inferencing infrastructure, offering GPU clusters and an AI-native cloud platform. Its Q1 2025 revenue surged 385% to $55.3 million, with ARR up 700%.

  • Why It Matters: With a $1 billion funding round and plans for UK expansion with NVIDIA Blackwell Ultra, Nebius is poised to capture a growing share of the AI-as-a-Service market.

• AMD (AMD) | UAV: 65.2

  • Why?: AMD’s MI300 series is gaining traction with Microsoft and Meta, offering 192GB HBM3 memory for inference workloads. It’s the anti-NVIDIA play for cost-conscious enterprises.

  • AMD has missed a lot of gains in the AI space to date, adding confidence that they have bought their time to capitalise on inference.

  • Why It Matters: Buybacks and valuation make it a sleeper hit if inferencing surges, but toppling or eating Nvidia’s cake will not be an easy task.

A Summary of the UAV Outputs

My UAV scores point to a tiered strategy to balance conviction and risk:

• Tier 1 (UAV >80): Uber (86.5), HIMS (84.6), Nebius (80.3)—the heavy hitters with massive upside and catalysts. Risks are present here, but the upside is particularly strong.

• Tier 2 (UAV 70-79): Alphabet (77.7), NVIDIA (71.2), Novo Nordisk (70.8)—safe bets with strong moats and steady growth

• Tier 3 (UAV 60-69): Tesla (68.2), AMD (65.2), Eli Lilly (64.5)— Could be fantastic opportunities, but may require a lower entry price. A 10% drop or more in any of these names in the coming weeks and months would be of major significance.

These themes—autonomous vehicles, GLP-1s, and AI inferencing—are the kind of disruptions that analysts often underestimate, like the internet or mobile phones back in the day.

My UAV framework spots companies trading at discounts to their true potential, giving a common sense nudge towards the business capitalising from the he technologies shaping the next decade.

I own a stake in Uber, Google, HIMS, Nebius, Alphabet, Nvidia, Novo Nordisk and Tesla. I’m hoping to enter AMD and Eli Lilly at a lower price point

Disclaimer: Personal investment analysis, not financial advice. All investments carry risk. I may hold positions in mentioned companies. Do your own research and consult a financial advisor before investing.

Introduction to AI Cloud Infrastructure

The rise of artificial intelligence is reshaping industries, but building AI models requires immense computational power that traditional cloud providers like AWS or Azure aren’t fully optimised to deliver just yet. Unlike these general-purpose platforms, Nebius AI, the flagship division of Nebius Group, offers specialised GPU clusters designed for AI-specific tasks like training, fine-tuning, and inference. This flexibility contrasts with AWS and Google Cloud’s broader, less AI-focused infrastructure, which can be costlier and less tailored for cutting-edge AI workloads.

To those less familiar with AI infrastructure, the first paragraph may have made absolutely no sense, but consider the examples below:

• AI-Powered Coding Revolution: As AI coding tools like Replit evolve, their utility hinges on computational power. While tools like ChatGPT offer solid coding assistance, connecting an AI coding platform to Nvidia’s Blackwell GPUs—available via Nebius’s cloud—delivers astounding results. For instance, xAI’s Grok 3, when paired with high-performance GPUs, can generate complex codebases in seconds, a feat validated by developer communities on X. In a future where software engineers may be augmented or even replaced by AI, Nebius’s ability to provide cloud access to super GPUs positions it at the forefront of a disruptive shift, potentially one of the decade’s most profound changes.

• Just for you to consider, if the higher performing GPUs can perform tasks as well, or better than mid-level software engineers, buying cloud access to a high performing chip may be the financially prudent move rather than hiring a human. No, this is not science fiction, this is playing out right now.

• Empowering Cash-Strapped Startups: Imagine a startup building an AI chatbot to compete with industry leaders. Limited by budget, it can’t afford a $40,000 Blackwell GPU that depreciates rapidly and isn’t needed full-time. Nebius enables this startup to access GPU clusters on a pay-as-you-go basis, allowing it to train and refine its chatbot efficiently. This model reduces upfront costs and accelerates development, giving small players a shot at innovation without the financial burden of owning hardware.

In a world where a single Nvidia Blackwell GPU can cost upwards of $30,000–$40,000, low-cost access to high-performance AI compute is in massive demand. Nebius is democratizing this access, ensuring high-performance AI isn’t reserved for tech giants with deep pockets.

The AI infrastructure market is nascent but explosive, with a ceiling that’s hard to predict. Like mobile phones, the internet, and cryptocurrencies, AI infrastructure could far exceed initial estimates, potentially rivalling transformative markets. Conceptually, this feels like a completely disruptive shift in society. The question I am trying to answer is whether it would be wise to invest some money into capitalising on the change.

Projected to grow from $47–50 billion in 2024 to $200 billion by 2028 (CAGR ~35%), this sector’s potential mirrors the unpredictable growth of past technological revolutions.

Nebius is positioned to capture the opportunity, enabling a wide range of AI applications through its specialised platform. The following article explores its positioning, history, financial health and valuation to assess whether this a player to consider.

History of Nebius

Nebius Group was born from the complex restructuring of Yandex, a Russian tech giant often dubbed “Russia’s Google.” Founded in 1997 by Arkady Volozh, Yandex dominated Russia’s internet ecosystem with services like search, e-commerce, and ride-hailing. However, geopolitical tensions following Russia’s 2022 invasion of Ukraine forced a strategic pivot. Yandex faced potential sanctions risks and operational constraints, prompting a $5.4 billion divestment of its Russian assets in 2024—the largest corporate exit from Russia since the invasion. This transaction, approved by Russian authorities and finalized in July 2024, separated Nebius as a new entity focused on international markets, with no remaining Russian operations or revenue streams.

The restructuring was a monumental feat, navigating regulatory hurdles across the U.S. and EU. Nebius relocated its headquarters to Amsterdam, retained 1,300 employees (mostly engineers), and secured $2.4 billion in cash with zero debt. Listed on the Nasdaq in October 2024, Nebius has since focused on AI infrastructure, leveraging Yandex’s decade-long expertise in AI and cloud computing. This unique history—transforming from a Russian tech conglomerate to a global AI player—positions Nebius as a rare case of successful corporate reinvention amid geopolitical challenges.

It also potentially provides some value. Nebius has snuck under the radar, and many looking at the company may be wary of the governance risks of the Russian ties. This would suggest that the company is discounted due to these risks, but are the risks true?

Impact of Nebius’s Russian Origins

Of course, concerns about Nebius’s Russian origins linger, but after a closer look I was surprised to find minimal risk due to strategic and regulatory measures:

• Complete Asset Separation: The $5.4 billion divestment of Yandex’s Russian businesses was a clean break, transferring all Russia-based operations (e.g., search, ride-hailing) to new ownership. This eliminated Nebius’s exposure to Russian market volatility and sanctions, ensuring no operational or financial ties remain. The scale of this exit underscores Nebius’s commitment to global markets.

• Regulatory Approval: The restructuring received full clearance from U.S., EU, and Nasdaq authorities, confirming compliance with international regulations. This approval mitigates risks of future sanctions or restrictions, as Nebius operates exclusively in jurisdictions like the U.S., Europe, and Israel.

• No Ongoing Exposure: Nebius generates zero revenue from Russian customers or suppliers. Its data centers are located in the U.S., Europe (e.g., Finland, UK), and the Middle East (e.g., Israel), with new facilities planned in New Jersey and Kansas City. This global footprint ensures no reliance on Russian infrastructure or markets.

• Leadership Relocation: CEO Arkady Volozh, was removed from EU sanctions lists in March 2024 following the divestment. The leadership team, now primarily in Amsterdam and the U.S., faces no geopolitical restrictions, further distancing Nebius from Russian influence.

Conclusion: Nebius’s Russian ties pose low risk due to its complete separation from Russian assets, regulatory approvals, zero exposure to Russian markets, and relocated leadership. Geopolitical concerns are not insurmountable and should not result in a deep discount, offering hope that this company may be overlooked.

Nebius - Weirdly Diversified: A Sum-of-the-Parts Valuation

Nebius AI is only one piece of the puzzle - it has a number of business segments operating in different markets. As I have done with Google, I’ve tried to value each segment separately and also conservatively. This provides a rough valuation with a high margin of safety which can be useful to appraise the price.

Nebius AI (AI Infrastructure Segment)

Overview

Nebius AI, the AI cloud provider positioned to provide low cost access to high performance GPUs, is the cornerstone of Nebius Group. It accounts for roughly two-thirds of revenue of the group.

Use Case and Market Growth

In 2024, Nebius AI’s revenue surged 462% year-over-year to $117.5 million, with Q1 2025 showing 385% growth. The broader AI infrastructure market is projected to grow from $47–50 billion in 2024 to $200 billion by 2028 (CAGR ~35%), driven by demand for generative AI and machine learning. Nebius’s annual recurring revenue (ARR) reached $220 million in March 2025, with a target of $750 million to $1 billion by December 2025. This aligns with industry trends, as companies like CoreWeave and Nvidia report similar growth. Over the next five years, AI infrastructure spending could double, fueled by enterprise adoption and startup innovation, positioning Nebius to capture significant market share.

Strategic Positioning

Nebius AI’s strategic focus on Europe and its Nvidia partnership set it apart in the AI infrastructure landscape.

Here are the key pros and cons:

• Pro - Access to $50 billion European cloud market: Nebius capitalises on Europe’s push for “digital sovereignty,” attracting clients seeking alternatives to U.S. based offerings. This aligns with regulatory trends like GDPR, positioning Nebius as a preferred partner for European enterprises and governments.

• Pro - Nvidia partnership enhances technological edge: As a launch partner for Nvidia’s Blackwell Ultra AI Factory Platform, Nebius gains early access to advanced GPUs, enabling superior performance for AI workloads and attracting cutting-edge AI developers.

• Con - Slower adoption in Europe: European clients may adopt AI infrastructure more slowly than U.S. counterparts due to conservative tech budgets and regulatory complexities, potentially limiting near-term growth compared to CoreWeave’s U.S. centric model. This is quickly changing, but it’s worth flagging.

• Con - Nvidia supply chain dependency: Reliance on Nvidia’s GPUs exposes Nebius to supply chain risks, as global chip shortages could delay data center expansions and impact client delivery timelines.

Competitive Edge

At first glance, Nebius AI lacks the scale of hyperscalers like AWS or Google Cloud, which could theoretically dominate the AI infrastructure space with their vast resources. However, Nebius leverages its specialist advantage to compete effectively in the AI cloud niche:

• Specialist Advantage by Proprietary Design: Nebius designs its own GPU racks, tailored specifically for AI workloads, reducing latency and costs compared to hyperscalers’ generic infrastructure. This specialisation enables Nebius to offer high-performance computing at competitive prices, appealing to AI startups and enterprises.

• Engineering Expertise: With 850 AI/ML engineers from Yandex’s legacy, Nebius builds future-proof systems, unlike competitors relying on third-party designs, ensuring adaptability to evolving AI demands.

• Niche Focus: By targeting AI startups and enterprises, Nebius avoids direct competition with hyperscalers’ broader offerings, carving out a $20–30 billion submarket within the broader cloud industry.

Financials and Valuation

Unsuspectedly, Nebius’s financial discipline shines through its zero-debt balance sheet and $2.4 billion cash reserve, enabling aggressive expansion without shareholder dilution. Management’s $2 billion 2025 capex plan (up from $1.5 billion) is demand-driven, avoiding overbuilding risks seen in competitors like CoreWeave ($20–23 billion capex, $7.5 billion debt - quite frankly terrifying). Nebius’s Q1 2025 operating cash burn was $197.8 million, but positive EBITDA is projected for H2 2025, reflecting prudent cost management.

For valuation, I am going to avoid CoreWeave’s inflated 23.22x price-to-sales (P/S) multiple ($71–78 billion market cap on $1.9 billion 2024 revenue) due to its debt and client concentration risks (77% revenue from two customers). It also just looks incredibly overvalued and bubbly.

Instead, to ensure a margin of safety and keep things conservative, I’m going to compare Nebius AI to traditional legacy cloud providers like AWS (Amazon: ~7x P/S) and Google Cloud (Alphabet: ~6x P/S), which grow at 20–30% annually—well below Nebius’s 385%.

Given Nebius’s faster growth (over 100% per year compared with 25%), I’m applying a still reasonable 12x forward P/S to its projected $1 billion 2025 ARR (around half the valuation of CoreWeave):

Valuation: $1 billion × 12 = $12 billion

Other Business Divisions: The other ‘parts’

At first glance, without considering Nebius’s other business segments the valuation seems quite reasonable. To complete the valuation though we need to consider the remaining revenue and assess the business in its entirety.

Toloka AI

Explainer: Toloka AI helps companies like Amazon improve their AI models by organising and labeling data—think of it as cleaning up messy data so AI can learn better. For example, it ensures a chatbot understands questions by providing high-quality training data.
Valuation: Estimated at $300 million (15x P/S on ~$20 million 2024 revenue), based on Scale AI’s 20x multiple. Toloka’s niche in data quality for LLMs offers high margins but faces competition from larger players.

Avride

Explainer: Avride builds tech for self-driving cars and delivery robots, like programming a car to drive itself or a robot to deliver pizza. It’s a futuristic business but needs heavy investment.
Valuation: Estimated at $100 million (20x P/S on ~$5 million 2024 revenue), reflecting its early-stage status and competition from Waymo. Strategic partnerships could unlock value.

TripleTen

Explainer: TripleTen runs online coding bootcamps, teaching people to become programmers. It’s like an online school for tech jobs, helping career-switchers land roles at tech firms.
Valuation: Estimated at $75 million (5x P/S on ~$15 million 2024 revenue), comparable to Coursera. Its low capital needs make it a stable cash generator.

ClickHouse Stake

Explainer: Nebius owns 28% of ClickHouse, a tool that helps companies store and analyze huge datasets, like tracking sales for a big retailer. It’s a passive investment with growth potential.
Valuation: Estimated at $560 million (28% of ClickHouse’s $2 billion valuation), adding value without operational risk.

Sum-of-the-Parts Valuation

Combining the valuations, Nebius’s total enterprise value is presented below, with the cash reserve added to derive the total equity value:

Note: The total equity value of $15.435 billion, above the current market cap of ~$11.905–$12.143 billion ($50–51/share), suggests ~27–30% upside.

Conclusion

Nebius Group is a compelling investment in a space where the ceiling is currently unknown. Replacing software engineers with GPUs could be the first real instance of AI taking our jobs, and while this article may be slightly sensationalist, if it comes to fruition even only slightly, the disruption and change from this trend could be seismic.

The market opportunity is large, and Nebius is well positioned. Nebius almost feels like a small scale European Google that has ingrained AI into its business right from the outset.

This is an early stage company though, and so far we do not know how well they will execute. It has neither the profitability or track record of Google - but is trading at a similar discount. This therefore makes Nebius more of a speculative play, rather than a certified undervalued business. An investment in Nebius requires a strong long term belief on the direction of the AI market, and an assurance that Nebius is positioned well.

I personally am inclined to feel positively on both those points, and have started acquiring a very small amount of shares. The real test will be Nebius’s next results, and whether it looks like profitability is possible. If revenue keeps growing at a high rate, and profitability is around the corner, the valuation could increase significantly and very quickly.

The information provided in this article is for informational and educational purposes only and does not constitute financial, investment, or professional advice. The author is not a licensed financial advisor, and the content is based on publicly available data, personal analysis, and opinions.