Nebius Group (NASDAQ: NBIS) is my largest position, and I want to be precise about why. This is not a bet on the word “AI.” It is a bet on a founder-led operator building the physical layer that every model — frontier or open-source, training or inference — has to run on. When demand for compute is structurally short of supply, the scarce asset is not the algorithm. It is the GPU cluster, the power contract, and the team that can stand both up on time. Nebius is a pure-play expression of that scarcity, and this is the full work behind the conviction.
Figures throughout are illustrative and drawn from company guidance and my dashboard snapshot. Nothing here is advice.
The business, in one paragraph
Nebius emerged from the restructuring of Yandex — the team kept the AI infrastructure DNA, a meaningful cash pile from the divestiture, and a clean, Western-domiciled listing. The core business is straightforward to explain, which is exactly why I own it: Nebius designs, builds, and operates high-performance GPU clusters and rents that capacity — plus the software layer to actually use it — to companies that need training and inference power they can’t or won’t build themselves. Around that core sit a handful of optionality assets: Toloka (data labeling for AI), TripleTen (ed-tech), Avride (autonomous vehicles), and a stake profile that includes deep hyperscaler and chipmaker relationships. The core is the thesis. The rest is a free call option.
AI data-center capex is the fastest-growing line item in tech
Illustrative industry AI-infrastructure spend. The point isn’t the exact number — it’s the shape: a multi-year, capital-intensive buildout where capacity is the bottleneck. Hatched bars are estimates.
1 — Operators I’d back as owners
I start with people, not the wave. The continuity that matters at Nebius is operator continuity: a team that already ran hyperscale-adjacent infrastructure inside Yandex under real constraints, now running a focused, well-capitalized, pure-play compute company. The tell I look for is capital allocation — are they spending on capacity that’s pre-committed and returns-accretive, or building speculative racks to chase a headline? So far the posture reads like ownership: staged buildout, disciplined regional expansion (Finland, then the US and broader Europe), and a willingness to monetize non-core assets rather than subsidize them forever. This pillar is what lets me size NBIS as a conviction hold instead of a rented theme.
2 — Primary research, not a borrowed thesis
I can defend this position from the source documents and product reality, not a newsletter. The mechanics: Nebius buys or reserves the scarce inputs (accelerators, power, data-center shells), assembles them into clusters, and sells GPU-hours plus an orchestration and MLOps layer on top. The economic question is not “is AI big” — it’s whether Nebius can keep its fleet highly utilized at prices that cover a punishing depreciation schedule and still expand returns. If I couldn’t explain that from filings and guidance, I would not own it at this weight. Anti-hype isn’t a pose here; it’s the reason the position is defensible.
Annualized run-rate revenue (ARR): the ramp is the whole story
Illustrative ARR trajectory framed against management’s stated ambition to scale toward multi-billion run-rate. ARR, utilization, and contracted backlog are the three numbers I track every quarter. Hatched bars are estimates.
3 — A real wave — with room to win inside it
The wave is not in doubt: multi-year, structural demand for training and inference capacity. But a tailwind attracts competition, so the wave alone is not an edge. Nebius is racing hyperscalers (who serve their own priorities first), a crowd of “neocloud” GPU renters, and the chipmakers’ own cloud ambitions. My question is where the defensible seat is. I think it comes from three places: (1) being a neutral, full-stack provider for customers who don’t want to hand their workloads to a competitor’s cloud; (2) speed and cost of standing up clusters, which is an operational moat more than a technological one; and (3) privileged access to the newest accelerators through deep vendor relationships. None of these are permanent. All of them are real today.
My subjective read of moat durability by source. The software layer is the swing factor — if orchestration and tooling create real switching costs, GPU-hours stop being a commodity.
4 — The economics underneath the narrative
This is where the story lives or dies, and where I spend most of my time. GPU cloud is brutally capital-intensive: you buy accelerators that depreciate fast, wrap them in power and cooling that cost real money every hour, and hope utilization and pricing stay high enough to earn a return before the hardware is a generation behind. The bull case requires three things to be true at once: utilization stays high (contracted, not spot), pricing holds (differentiation, not a race to the bottom), and each new cluster earns a better incremental return than the last as the platform layer scales. If those hold, revenue growth converts into improving margins and eventually real cash generation. If they don’t, this becomes a capital furnace. I model the unit economics, not the pitch deck — physics and power bills don’t care about a narrative.
Capex leads revenue — the gap is the risk and the opportunity
Illustrative. Capex running ahead of revenue is normal and necessary for a capacity build — but it’s exactly why the balance sheet and financing terms matter as much as the demand.
Valuation: how I frame a pre-profit compounder
You cannot value NBIS on a trailing earnings multiple — it’s deliberately spending ahead of profitability. I anchor on revenue/ARR multiples, the trajectory of gross margin, and the eventual free-cash-flow profile at scale, then sanity-check against the comparable set. The honest read: NBIS is priced as a high-growth infrastructure story, so the multiple is a promise about the future, not a reflection of the present. My job is to keep asking whether the ramp is tracking the promise.
| Comp | Model | Growth profile | Margin today | My read |
|---|---|---|---|---|
| NBIS | Pure-play AI cloud | Hyper-growth | Negative (investing) | Own it |
| Hyperscalers | Cloud + everything | Steady | Strongly profitable | Different bet |
| Neocloud peers | GPU rental | Fast | Thin / negative | Less full-stack |
| Chip vendors | Silicon | Fast | Very high | Own the picks, not the mine |
Illustrative positioning, not a screen. The purest way to express “compute is the bottleneck” without buying a mega-cap conglomerate is a focused operator — which is why NBIS earns the top slot in the book.
5 — Asymmetry & stance
Stance: Bullish. Conviction: High. The upside case is that Nebius becomes a durable, neutral pillar of AI infrastructure — a business the market eventually values on cash flows, not just growth — while the optionality assets (Toloka, Avride, TripleTen) surprise to the upside. The downside is equally real and I refuse to pretend otherwise: valuation can compress violently on any growth wobble, GPU-hour pricing can reprice as supply catches demand, and a capacity build is precisely the kind of thing that slips. That’s why this is a conviction position, not a “set-and-forget” one — I expect volatility and size so I can hold through it.
Bull case
- Compute stays supply-constrained for years; Nebius sells everything it can build.
- Software/MLOps layer creates switching costs — GPU-hours stop being a pure commodity.
- Margins inflect as utilization and scale improve; the market re-rates to cash flows.
- Optionality assets (Avride, Toloka, TripleTen) get monetized as free upside.
Bear case
- Supply catches demand; GPU-hour pricing compresses and utilization dips.
- Capex outruns financing; dilution or debt on worse terms.
- Hyperscalers use scale to undercut neutral providers.
- Hardware cycles depreciate the fleet faster than it earns its return.
The Ovatek Lens scorecard
Green = clears cleanly. Amber = the pillar I’m actively monitoring — here, unit economics. That’s the number that decides whether this is a great business or an expensive one.
Catalysts I’m watching
- ARR milestones: each print that confirms the ramp is tracking guidance de-risks the multiple.
- Gross-margin inflection: the first clear sign that scale is improving unit economics, not just top line.
- Marquee contracted backlog: long-dated commitments from serious customers convert “demand” into “revenue you can underwrite.”
- Optionality monetization: any value-crystallizing event at Avride, Toloka, or TripleTen.
Thesis-break criteria
I write these before I’m emotional so a falling price can’t rationalize me into or out of the position. I’m out, or materially trimming, if:
- Sustained evidence that GPU-cloud pricing and utilization are structurally impaired — a trend, not a noisy quarter.
- Capital allocation or governance stops looking like ownership thinking (value-destructive M&A, sloppy dilution, empire-building).
- The path to improving economics as the fleet scales clearly closes — margins go the wrong way as revenue grows.
Until one of those hits, I treat drawdowns as information about price, not automatically about thesis. Process over predictions — and this is the position I’ve done the most work to earn the right to hold.