The AI infrastructure race is on fire, and Nebius Group (NASDAQ: NBIS) sits right at the heart of it. Once the international arm of Yandex, the company relaunched in 2024 as a pure-play AI cloud platform and has since become one of the fastest-growing names in the sector. Triple-digit revenue growth, multi-billion-dollar contracts with Microsoft and Meta, and a stock that has swung wildly between euphoria and panic — Nebius has it all. But there is one phrase that keeps coming up in every earnings call, every investor presentation, and every analyst note: capacity bottleneck.
The company is literally selling everything it can switch on — and still leaving hundreds of millions, possibly billions, of dollars on the table every quarter because it simply does not have enough GPUs, power, or floor space to meet demand.
So the million-dollar question (or rather, the multi-billion-dollar question) is this: Can Nebius break through its capacity constraints fast enough to turn today’s explosive demand into tomorrow’s explosive profits?
The Scale of the Problem
In its most recent quarter (Q3 2025), Nebius posted revenue of $146.1 million — a jaw-dropping 355% increase year-over-year. Almost 90% of that came from its AI infrastructure business, which grew more than 400%. Any normal company would be popping champagne. Instead, the stock dropped over 10% after the report.
Why? Because analysts were expecting $155–160 million, and management openly admitted the miss wasn’t due to soft demand — it was because they ran out of capacity halfway through the quarter. CEO Arkady Volozh said it plainly on the earnings call: “Every single watt and GPU we bring online gets sold immediately. If we had more, we would have sold more.”
The company currently has roughly 1 gigawatt of contracted power capacity either live or in late-stage development. That sounds like a lot — until you realize the biggest hyperscalers are building individual campuses that will consume 1–2 GW each. Nebius already has a visible pipeline of committed customer spend that could push annual recurring revenue toward $7–9 billion once capacity catches up. The gap between “what we have” and “what customers want right now” is enormous.
The Expansion Roadmap
Nebius is not sitting still. Management has laid out an aggressive build-out plan:
- Reach 2.5 GW of total power capacity by the end of 2026
- Bring 800 MW to 1 GW of new capacity online during 2026 alone
- Multiple new data center regions under construction or negotiation in Europe and North America
- Continued heavy investment in its flagship Finnish campuses, which benefit from cheap hydroelectric power and some of the best PUE ratios in the industry
The Meta deal alone — a three-year, $3 billion commitment — is expected to ramp aggressively in 2026, with some clusters coming online in as little as three months once the hardware arrives. Microsoft’s multi-year cloud agreement will follow a slower curve but still adds hundreds of millions in high-margin revenue as new pods light up.
The Real Hurdles
Throwing money at the problem sounds easy, but AI data centers are not built like traditional warehouses. The bottlenecks are real and multilayered:
- GPU Supply Nvidia’s latest chips (H100, H200, Blackwell) remain allocation-constrained. Even preferred partners wait months. Nebius has secured solid placement in Nvidia’s queue, but it can’t magically jump ahead of Microsoft, Google, or Meta’s own builds.
- Power Availability Getting gigawatts of new grid connections approved and built takes years in most countries. Nebius is attacking this with long-term power purchase agreements, renewable energy deals, and sites in power-rich regions (Nordics, parts of the U.S.), but the timeline is still measured in quarters and years — not weeks.
- Construction & Cooling Modern AI clusters generate insane amounts of heat. Liquid cooling, custom racking, and high-density designs add complexity and lead time. Nebius has proven it can deploy fast (some clusters in under 90 days), but scaling that speed across dozens of projects simultaneously is a logistical marathon.
- Capital Intensity The company is burning cash on CapEx at a furious pace. Adjusted EBITDA is still negative, and net losses are widening in absolute terms (though shrinking dramatically as a percentage of revenue). More capacity = more short-term losses before the revenue tsunami hits.
The Bull Case
If Nebius executes even reasonably well, the upside is staggering:
- Locked-in hyperscaler contracts provide revenue visibility through 2027–2028
- Gross margins in the AI infrastructure segment already exceed 60% and should climb higher as utilization stays near 100%
- A massive backlog of unsold but committed demand acts like a coiled spring
- Geographic and power diversification reduces single-point risk
- The company’s software stack and operational playbook are battle-tested from its Yandex days
Many analysts believe that once the 2026 capacity wave comes online, revenue could triple year-over-year again — this time without the “we ran out of GPUs” excuse.
The Bear Case
Execution risk is real. Delays in power connections, GPU deliveries, or construction permits could push meaningful revenue acceleration into 2027 or beyond. Competition is ferocious — CoreWeave, Crusoe, Lambda, and the hyperscalers’ own builds are all fighting for the same scarce chips and megawatts. And the stock is not cheap on a near-term basis; any hint of slippage tends to get punished hard.
The Bottom Line
Nebius is not a company struggling to find customers. It is a company struggling to build fast enough to serve the customers who are already waving billion-dollar checks.
The capacity bottleneck is painful today, but it is also the clearest possible proof of demand. If management delivers on even 70–80% of its 2026 build-out targets — and early evidence suggests they are hitting milestones — the revenue and profit explosion that follows could be one of the great growth stories of this AI cycle.
For long-term investors willing to stomach the volatility and the cash burn, Nebius looks less like a company with a problem and more like a company with a temporary inconvenience on the road to something much, much bigger.