AI Cloud / Neocloud
The former Yandex international business, rebuilt as a GPU cloud by the founder who built Yandex. It's growing faster than any company on this list, has the cleanest balance sheet of the neoclouds, and the valuation already assumes it keeps doing both.
Personal notes published for reference. Figures are sourced from company filings, releases, and press coverage as of the date above — the interpretation is mine and may be wrong.
CEO Arkady Volozh co-founded Yandex and ran it for 25 years, building Russia's search, maps, ride-hailing, and cloud businesses against Google. After the 2024 split from Yandex, he took the international assets and ~1,000 engineers and rebuilt them as an AI cloud in under two years. That's one of the strongest operator track records on this watchlist. The caveat: the Yandex history still carries a geopolitical overhang for some investors. Volozh sold ~$8.9M of stock on Oct 1, but it was a tax-covering sale on vested shares, not a signal.
AI compute capacity, which is supply-constrained rather than demand-constrained right now. Pricing is moving up: Nebius raised AI cloud prices ~21% this year, and contract value per megawatt reached $20–25M a year. Goldman projects US data-center power demand growing 38% a year in both 2026 and 2027. The ceiling isn't the question. The question is how much of that demand flows to independent clouds versus the hyperscalers building their own.
The best of the three neoclouds by a wide margin. $8B of cash at June 30, 2026, adjusted EBITDA margin of 41% (50% in the AI cloud segment), and a business model in which customers pay up front: ~$9B of customer prepayments expected in 2026, with 70% of Q2 deals carrying prepayment terms. Capex is huge ($20–25B in 2026), but much more of it is funded by customers and equity than by debt, and it's on its first asset-backed facility ($775M at SOFR+2.5%) rather than living on credit.
Exceptional so far. Q2 2026 revenue was $582M, up 454% YoY and 46% sequentially, and ARR went from $1.9B to $3.0B in one quarter. Nebius has landed two hyperscaler anchor contracts: Microsoft ($17.4B, up to $19.4B, over five years) and Meta ($12B dedicated plus up to $15B more over five years, on Vera Rubin, starting early 2027). Q2 alone added four more contracts averaging over $1B each. The contracted-power target was raised to 5 GW by year-end 2026.
Almost all of the value is the AI cloud. The rest is a portfolio of options most investors ignore.
98% of revenue. A full-stack GPU cloud, not just rented racks: its own data centers (16 by year-end), its own server designs, and a software layer on top (managed Kubernetes, inference platform). The Oct 1 acquisition of Inferize adds inference optimization to its Token Factory product. That matters as the market shifts from training to inference, where software and efficiency count for more than raw GPU count.
New in Q2: a partnership model where third parties deploy Nebius's platform in their own data centers. It's a way to grow past the capital constraint without carrying all the capex, and it's the piece that could make Nebius look more like a software company than a landlord over time.
Avride: autonomous vehicles and delivery robots, with 200+ vehicles and 60,000+ commercial rides on Uber. TripleTen: an edtech bootcamp. A minority stake in ClickHouse, the database company that raised at a $15B valuation in January. None of these drives the thesis, but together they're plausibly worth several billion dollars that the market barely counts.
NVIDIA is a strategic investor and an early recipient of new platforms. Palantir recently named Nebius a preferred AI infrastructure partner, and Nebius just signed a 12-year lease for 50 MW in the southeastern US.
A high and active institutional base, more conviction money than index weight.
Goldman, BlackRock, Orbis, Accel, and Fred Alger are all in the top ten. The name that stands out is Situational Awareness, Leopold Aschenbrenner's AI-focused fund, with ~4.5%. It's a concentrated, thesis-driven bet on AI compute demand. It's not an index holder; it chose this.
The best-run neocloud, priced like it.
If this plays out: ~$120
At ~$64B, Nebius trades at roughly 7trades at roughly 8x its own year-end ARR targetndash;8x its own year-end ARR target, against ~4–5x for CoreWeave and IREN on an enterprise-value basis. Any slip in bringing capacity online, a GPU price reset when Vera Rubin supply catches up, or a pause in hyperscaler outsourcing compresses that premium fast. The stock already fell from ~$300 to ~$233 without any bad news. A $20–25B capex year also means more equity or convertible issuance if prepayments disappoint.
If this plays out: ~$280
ARR lands in the $7–9B range by year-end and Meta capacity starts delivering in early 2027. The multiple compresses as the growth rate normalizes, but earnings grow into it. That lines up with the ~$284 analyst average.
If this plays out: ~$400
Nebius goes past 1 GW a year of deployments from 2027 as planned and the asset-light partner model scales. Inference software (Token Factory, Inferize) turns it into a platform with real margins rather than a GPU reseller. Avride and ClickHouse get valued separately. The street-high scenario of ~$320 and beyond is what that looks like.
$221.07 at the Oct 9, 2026 close (52-week range $73.52–$299.86, market cap ~$61B). It spiked to about $253 on Oct 6 and gave it all back within three sessions. New coverage is bullish: Rosenblatt initiated at Buy with a $304 target and William Blair at Outperform, and the Buy consensus averages ~$284. Insiders have been selling into strength, though. Beyond the CEO's tax-covering sale, another executive sold ~29K shares at $234 on Oct 1, and the COO reportedly sold 500,000 shares on Oct 6. Q3 results are due between late October and Nov 10; trackers disagree on the date. (The comparison table below still uses Oct 5 prices.) The scenario targets above are my own rough, directional estimates, not formal price targets. They're anchored to the sell-side range rather than a full model.
All three rent NVIDIA GPUs to AI labs and hyperscalers. They differ in how they're funded, who they depend on, and what you pay for each dollar of contracted revenue.
| NBIS · Nebius | CRWV · CoreWeave | IREN | |
|---|---|---|---|
| Price (Oct 5) | ~$233 | ~$87 | ~$40.50 |
| Market cap | ~$64B | ~$48B | ~$16B |
| Debt / cash | Low debt / $8B cash | ~$35B debt | $7.6B debt / $5.9B cash |
| Latest quarter revenue | $582M (+454%) | $2.6B (+112%) | FY26 $707M, mostly bitcoin; AI cloud $129M |
| 2026 exit run rate | $7–9B ARR target | $18.5–19.5B | $4B contracted ARR ($1B operating) |
| EV ÷ exit run rate (rough) | ~7x | ~4.4x | ~4.4x (on contracted) |
| Backlog / anchors | Microsoft $17–19B, Meta $12–27B, 4 more $1B+ deals | ~$129B: Meta ~$35B, OpenAI ~$22B, Anthropic, Jane Street | Microsoft $9.7B, a frontier lab, NVIDIA $3.4B |
| Customer concentration | Moderate (two hyperscaler anchors plus many AI labs) | High: Microsoft 72% of H1 2026 revenue | High near-term (Microsoft) |
| Profitability | Adj. EBITDA 41%, roughly breakeven net | Adj. EBITDA 59%, net loss ~$1.9B TTM (interest) | Net loss $703M (bitcoin impairments) |
| Analyst avg. target | ~$284 (+22%) | ~$142 (+62%) | ~$78 (+92%) |
The biggest and most proven at scale. Revenue doubled to $2.6B in Q2, the backlog is ~$129B, and it has signed every major lab: Meta, OpenAI, Anthropic, and Google. It also just pushed through cumulative price increases of ~37.5%. The problem is the balance sheet. About $35B of debt, $640M a quarter in interest, $35–39B of 2026 capex against ~$13B of revenue, and negative free cash flow of $5.7B in Q2 alone. Credit markets were pricing roughly a 50% five-year default probability before the Q2 print. Microsoft is still 72% of revenue, and there's a securities class action over its IPO-era delivery claims. It's the cheapest of the three on revenue and the most fragile if GPU pricing turns or a big customer slows.
A bitcoin miner turning into an AI cloud, with one real advantage: it owns its power and land, including multi-GW sites at Childress and Sweetwater in Texas, and power is the industry's binding constraint. It has a $9.7B Microsoft contract, $4B of contracted ARR, and NVIDIA "Exemplar Cloud" status on GB300. But only $1B of ARR is actually operating, FY26 AI revenue was just $129M, and the transition was funded with $4.7B of new equity and $6.3B of convertibles in a single year. That's ONDS-style dilution at a larger scale. It's the cheapest way to own contracted AI capacity, and also the earliest and least proven.
| Management | TAM | Cash / position | Execution | |
|---|---|---|---|---|
| NBIS | Strong | Strong | Strong | Strong |
| CRWV | Good | Strong | Weak (debt) | Strong |
| IREN | Good | Strong | Mixed (cash, but heavy dilution) | Early |
| OKLO | Good | Strong | Mixed ($3B cash, but serial ATMs) | Pre-revenue |
On the framework this page uses, Nebius is the only one of the three that passes all four. My rule is that weakness on more than one criterion is usually disqualifying. CoreWeave fails clearly on one (balance sheet) and carries concentration risk. IREN is borderline on two (dilution and unproven execution). The trade-off is price: Nebius costs roughly 7x its year-end run rate, while the other two are near 4–5x. You pay a premium for quality.
That's how they line up against my criteria, not a recommendation to buy any of them. I'm not a licensed advisor. Position size, timing, and whether a 7x multiple is acceptable are personal calls. All three are high-beta: they move together on AI-capex sentiment, and in a GPU-pricing downturn they'd all fall at once.
Oklo isn't a fourth neocloud. It sits one layer down. The neoclouds sell compute, and Oklo wants to sell the power that compute runs on. It's the same AI-buildout theme with a completely different risk profile, so it doesn't belong in the revenue-multiple table above.
Small fast-reactor "Aurora" powerhouses, 15–75 MW each, built to sit next to data centers and sell power under long-term contracts. Co-founders Jacob DeWitte and Caroline Cochran run it; Sam Altman was chairman until 2025. The customer pipeline is real on paper: a 12 GW master agreement with Switch, and a Jan 2026 prepayment deal with Meta for a 1.2 GW campus in Ohio. Meta and Equinix have already paid money in. There are also LOIs with Equinix, Diamondback, and Prometheus Hyperscale, about 14 GW of pipeline in total.
Its Groves isotope test reactor in Texas reached first criticality less than 11 months after groundbreaking, which validates the fast-reactor physics. The DOE approved the preliminary safety analysis for the first commercial unit, Aurora-INL, which still targets a 2028 startup. A separate radioisotope business should bring in first revenue in early 2027. That's still a long way from a plant selling power. The NRC rejected Oklo's first license application in 2022, and fuel (HALEU) is scarce: the Centrus supply LOI starts in 2029.
Strong on TAM, since power is the real bottleneck for every company on this page. Cash looks fine at $3B against roughly $120–150M of operating burn plus $400–500M of 2026 capex. But that cash came from $1.9B of at-the-market stock sales this year, and a fresh $1B ATM started Sept 11. That's the ONDS/IREN pattern of funding by dilution. On execution it's pre-revenue by definition: no Aurora plant produces electricity before 2028 at the earliest. So it's weak on two criteria, cash funded by dilution and unproven execution, the same as IREN, except IREN at least has $1B of ARR running today.
Down ~81% from its $194 high, trading near its 52-week low after the new ATM and a UBS cut to $41 (Neutral). Canaccord still says $100 (Buy), and the average target is ~$76. You can't put it on a revenue multiple. The ~$6.7B market cap is an option price on 2028+ power sales actually happening.
Nebius is the most proven and most expensive. Oklo is the least proven and, after an 81% drawdown, the most beaten-up, so it's a long-dated option rather than a business with numbers to underwrite yet. The one real argument for owning it alongside a neocloud: its risks are regulatory, fuel, and construction, not GPU pricing. A crash in GPU rental rates wouldn't hit Oklo the way it would hit NBIS, CRWV, and IREN together, though a broad pullback in AI capex would still hit all four.
The richest multiple of the neoclouds. The stock has already swung from $74 to $300 and back to $233 within a year.
Today's price increases reflect scarcity. When Vera Rubin supply ramps and older Hopper/Blackwell fleets age, rental rates could fall while the depreciation schedules stay fixed.
The Microsoft and Meta contracts are anchors, but both companies are also building massive in-house capacity. Neoclouds are overflow capacity until proven otherwise.
$20–25B of capex in 2026 and 1 GW+ a year after that. Even with prepayments, expect more converts or equity over time.
The Yandex origin no longer affects operations (it's Netherlands-headquartered, with data centers in Finland, France, Iceland, Israel, and the US). It still shows up in some investors' risk screens and in headlines.