Company Deep Dive #2 - NEBIUS
Selling vertically integrated AI infra to compute-hungry customers
In today’s edition, we’ll dive deep into the publicly traded company Nebius.
1. Why a little-known Dutch cloud company is suddenly on everyone’s radar
Nebius Group began trading on Nasdaq only 9 months ago, yet a major investment bank has already told clients the shares could climb to $68 within a year. Roughly 26% higher than today’s price of about $54.
The excitement comes from one straightforward fact: Nebius is one of very few publicly-listed firms whose entire business is renting high-end GPUs to anyone building AI software. Those chips are the workhorses behind ChatGPT-style models. And the world does not have enough of them.
During the March-to-June quarter, Nebius reported revenue of $55.3 million. Almost 5x over the previous 12 months. The company’s decision last December to raise an extra $700 million (led by Nvidia itself) means Nebius now sits on a cash pile of roughly $2 billion with no debt.
It is ploughing that money straight into hardware. An expansion in Finland alone will add room for up to 60,000 GPUs. That’s enough computing power to train several of today’s LLMs in parallel.
2. What Nebius actually sells
Nebius sells vertically integrated AI infrastructure. Its data centers are full of Nvidia chips wired together with super-fast networking. And they sell compute power to software developers.
These software developers book that compute by the minute, train their models, and shut the lease the moment the job finishes. Nebius also offers storage for the oceans of training data and a basic toolkit for running finished models in production, but the raw GPU time is the draw.
Demand exists because very few companies can afford (or even know how) to run a modern AI cluster themselves. A single H100 card costs tens of thousands of dollars. A useful cluster needs hundreds of them, plus specialized engineers to keep everything humming. Renting is easier, provided someone has spare capacity.
Over the past 18 months that “someone” has usually been a hyperscale cloud such as AWS. The hitch is that AWS and its peers are constantly sold out of top-tier chips. Nebius launched with a simple promise: we will buy the hardware ahead of time so you do not have to wait.
3. Why the whole market is expanding so fast
Two developments sit behind the numbers.
AI adoption has moved from tech giants to everybody else. Stanford’s latest AI Index shows that by the end of 2024 nearly 80% of all organizations surveyed were already using some form of AI in daily work.
Model sizes are exploding. Each new generation contains more parameters, needs more training passes, and therefore consumes more GPU hours. Analysts at Goldman Sachs put the current global market for outsourced GPU compute at roughly $100 billion. And they expect it to more than double by the end of the decade.
Nebius is hardly alone in chasing that pot. CoreWeave, Lambda Labs, and a growing field of “GPU clouds” have appeared. A recent industry roundup counted 30 such providers, up from fewer than 10 just two years ago. Yet the same survey notes that only two (CoreWeave and Nebius) are meaningful public-market companies here. This makes Nebius the only European-headquartered option.
4. Paying the bills, in simple terms
Because Nebius charges strictly for usage, the business behaves more like an electricity utility than a subscription software firm. When its servers sit idle, revenue stops but depreciation of the hardware continues. The safest way to avoid that mismatch is to fill the data center as fast as possible.
Revenue has so far kept pace. The company’s quarterly sales jumped 385% year on year, while operating costs rose only about one-third. Nebius says the cash it already holds is enough to cover all planned building work through 2026.
Investors still worry about spending outstripping income. The largest single cost is buying the GPUs themselves. A back-of-the-envelope sum suggests Nebius is currently valued by the market at roughly $320,000 per installed GPU. Slightly above where CoreWeave trades on the same metric.
If Nebius buys chips faster than it can rent them out, that ratio will climb. It’s bad for profitability. The flip side is just as dramatic. Once a new rack is full, every extra dollar of rent has relatively little additional cost. So the margins widen quickly.
5. The bumps in the road
Three threats stand out here:
Supply risk. Nebius relies on one supplier (Nvidia) for almost all of its chips. Any production hiccup or new export rule could delay hardware deliveries and push customers elsewhere. The strategic investment from Nvidia lowers this risk but does not remove it, as the entire industry competes for the same silicon.
Price wars. Big clouds may decide to slash GPU rental rates once their own stock catches up. A recent Forbes piece on CoreWeave highlighted fears that over-capacity could spark exactly such a war as early as 2026 . Nebius believes its in-house server design gives it operating costs 20–25% lower than rivals’, which would let it match cuts without destroying gross margin. But that claim will only be proven after a real contest.
Customer concentration. Early filings show that at one point, a single tenant represented half of Nebius’s accounts receivable. Management says the mix is broadening, yet the risk remains that one fast-growing AI startup could suddenly vanish or move its workloads back to a hyperscaler.
6. Why infra startups should care
Hundreds of venture-backed companies building developer tooling, data-processing pipes, or AI application layers rely on ready access to GPUs. The Stanford AI Index counts around 70,000 private AI-focused firms worldwide.
Not all of them rent compute. Some license models, others sell consulting. But the data indicates that roughly one-in-five fall into the infra bucket, meaning they build platforms that need to run or train models themselves. Industry analysts estimate that close to one-third of these infra startups have shifted at least part of their workload from hyperscalers to specialized GPU clouds such as CoreWeave or Nebius since early 2024. That puts somewhere in the region of four to five thousand young companies on Nebius’s potential client list today.
For those startups, Nebius’s fortunes matter in three main ways:
Pricing anchor. Pure-play GPU clouds set the going rate for high-end compute. If Nebius raises or cuts prices, AWS and CoreWeave often follow. Young companies with thin cash balances will see the effect almost immediately in their monthly bills.
Capacity buffer. When Nebius opens a new hall of servers, wait-times for chips fall. That speeds up model training cycles and can shave weeks off a product roadmap. Conversely, any delay in Nebius’s build-out (e.g. if grid power approvals in Finland slip) could stall dozens of customer launches.
Investor sentiment. Because Nebius is one of only two pure plays visible in public markets, venture investors often use its share price as a crude barometer for the health of the GPU-rental niche. When the stock jumped on the Goldman report last month, several term sheets for seed-stage infra startups reportedly closed on better terms. The link is not mechanical. But in a funding environment where valuations are sharply negotiated, a strong Nebius performance can lift the perceived market multiple across the private cohort.
7. Correlations, dependencies and feedback loops
The most obvious dependency is between Nebius’s revenue growth and the pace of AI model innovation. If larger models keep arriving every few months and each one is more hardware-hungry than the last, Nebius’s utilization will stay high. In that scenario the company’s capacity expansions could even prove conservative, pushing share price (and by correlation, private-market sentiment) higher.
A subtler connection involves Nvidia’s product cycle. Nebius has promised to be the first European provider of Nvidia’s forthcoming “Blackwell” chip family. Investors expect that milestone to attract a fresh wave of customers and keep rental prices from sliding. If Nvidia slips on delivery, Nebius’s own guidance may fall short. This would in turn dent the implied growth valuations of startups that depend on reselling Nebius capacity to their own clients.
There is also a capital-spending feedback loop. When Nebius pours billions into new racks, its suppliers (builders of cooling gear, power-distribution systems, and telecom fiber) experience revenue surges. Many of those suppliers are themselves early-stage companies focused on next-generation data-center gear. A slowdown at Nebius would ripple into their order books and vice versa.
Also financing risk is interlinked. Venture investors prefer startups that can shift workloads between multiple clouds. If Nebius executes its plan and builds out in both Europe and North America, it becomes a credible second source for GPU capacity behind AWS. That diversification reduces technical risk for startups, which can help them secure funding. If Nebius stumbled, the perceived concentration risk in relying mainly on CoreWeave or a hyperscaler could dampen new investments in the whole infra-startup category.
8. Putting the figures in human terms
Consider these headline numbers:
In May, Nebius earned roughly as much in a single quarter as it had in the whole of the previous calendar year.
Its latest fundraise gives it money to buy about 40,000 current-generation GPUs at wholesale prices. At typical utilization and pricing, that hardware can generate annual revenue well north of its purchase cost.
Goldman’s $68 target assumes Nebius captures only a sliver of a global GPU-rental market already measured in the high tens of billions. Even a modest out-performance could move the stock dramatically because nearly every dollar of extra sales will carry a high gross margin once the fixed infrastructure is installed.
Venture capital continues to pour into the AI stack. Crunchbase recorded $91 billion dollars of fresh startup investment in the second quarter, with AI infrastructure taking a double-digit slice. A significant share of that funding flows straight back out as GPU-rental costs, meaning Nebius’s top line is indirectly supported by the VC cycle.
9. Bottom line and things to watch
Nebius is a bet that demand for computing power will out-run chip supply for several more years. So far the results suggest that bet is paying off. If the company keeps scaling on schedule (Finland expansion by this winter and Kansas City coming online next spring), it will likely hit the low end of its goal of roughly $750 million in yearly recurring revenue by December. Meeting that milestone would strengthen the valuation anchor for dozens of smaller infrastructure startups and could even widen the venture fundraising window again.
Watch three indicators:
Capacity utilization. Nebius needs those new GPUs earning rent as soon as they hit the racks. Management does not publish exact utilization but quarterly revenue versus cumulative capex is a useful proxy.
GPU spot-pricing. If hourly prices on competing clouds fall faster than Nebius’s own costs, the firm may either accept thinner margins or risk losing customers.
Venture funding for infra startups. A sharp pull-back in early-stage funding would signal that end-user demand for Nebius’s services could soften 18 months later, once current startup cohorts run out of cash.
On balance, Nebius still offers one of the cleanest ways for public-market investors to benefit from the AI infrastructure wave. And its success or failure will echo across the private sector. When the share price moves, it does more than change a line in a portfolio. It sends a message about the availability, price and perceived future of the computing power that thousands of young infra companies need to thrive.
If you are getting value from this newsletter, consider subscribing for free and sharing it with 1 infra-curious friend: