Nebius Group N.V. NBIS

Key data

The business

Nebius rents artificial-intelligence computing. It builds data centers, fills them with Nvidia GPUs wired into clusters, wraps them in its own servers and software, and rents the capacity to companies training and running AI models. The unit that matters is the megawatt: a landmark deal now yields more than $20 million of annual revenue per megawatt, and the payback on the hardware runs under two years. The core AI cloud sits alongside three smaller arms, the data-labeling shop Toloka, the self-driving unit Avride, and the ed-tech business TripleTen, none of which is the reason anyone owns this.

The company is what survived Yandex. Founder Arkady Volozh built Russia's Google, sold the Russian operations in 2024 for cash, and pointed the remaining shell and the proceeds at AI infrastructure in Europe, the US and the Middle East. The edge, if there is one, is scale and secured power: owned sites in Finland, leased capacity across New Jersey, Israel, Iceland and beyond, and in-house design that lifts utilization the labs cannot match alone.

The numbers

The reset is the whole story. Revenue collapsed to near zero when Russia left, then reignited: $91.5 million in 2024 became $529.8 million in 2025, and a single quarter, Q2 2026, printed $582.3 million of group revenue, more than the entire prior year. Operating losses widened the whole way down, because scaling this business means spending first.

The per-share loss tells the second-derivative story: heavy in late 2025 as capacity came on, then narrowing fast through 2026 as revenue caught the cost base. AI cloud revenue for the first half of 2026 reached $964.6 million.

The operational picture the last look on this name could only flag as unverifiable is now in the filings, and it reads strong.

Here is the tension the multiple hides. Capex hit $4.07 billion in 2025 against $384.8 million of operating cash flow, and the more the business grows the more cash it swallows. The company is closing that gap with borrowed money, a first $775 million secured facility at SOFR plus 2.50%, backed by GPU hardware and one investment-grade customer's contracts, plus fresh convertibles, with 50% to 60% of capex meant to be self-financed by customer prepayments. What the market does not price is that the disclosed backlog is a framework for more than $40 billion of customer commitments against a $54 billion market cap; the print that settles the argument is energized megawatts converting into recognized revenue.

Management

Volozh has built one $30-billion company and is spending fast to build a second, so judge him on capital allocation, not comp. There he has been resourceful: a repeatable financing framework, prepayments that shift buildout risk onto customers, converts that limit near-term cash cost. The tell that cuts the other way is insider selling. Over the last twelve months insiders sold about $173 million across seventy transactions and bought nothing, led by Andrey Korolenko at $101.6 million in May and $8.0 million in July, with Volozh himself selling $11.0 million; plan status is not disclosed on any of them, so read the size, not the intent.

How it fails or surprises you

The backlog converts (right tail). More than $40 billion of contracted commitments sit against a $54 billion cap, with 5 GW guided for year-end. Energize that capacity at $20-million-plus per megawatt and the $22.2 billion 2028 consensus stops looking fanciful. The market discounts it for execution and financing risk; energized-MW disclosures each quarter reveal it first.

Growth keeps burning more cash. The operating loss widened to $611.7 million as capex reached $4.07 billion on $384.8 million of operating cash flow. This works only while capital markets stay open. If convert or debt terms tighten hard, the buildout slows and the backlog cannot convert; the next financing's pricing is the tell.

Rental prices crack. Today prices rise, up more than 30% on older GPUs and doubling on new, which is the entire case for owning it. A supply flood that pushes revenue per megawatt below the depreciation the debt assumes breaks it. Watch revenue per MW on new deals, not renewals.

Closing thoughts

No single quarter resolves this. What matters is whether Nebius can keep funding a buildout that costs more than four billion dollars a year until the contracted backlog turns into collected cash, and that comes down to continued access to capital on terms cheaper than the yield each megawatt throws off. The right tail is unusually fat here, a backlog larger than most of the market cap, but the left tail is permanent, not a drawdown: if financing closes or prices crack mid-buildout, a company burning this much cash does not get a second chance. On today's evidence the odds favor the upside, with the honest caveat that the stock already sits at the top of its twelve-year valuation range and prices a good deal of the conversion in.

The bet is still that AI companies keep needing more computers to train and run their models, and keep renting them from Nebius instead of buying their own, and that those customers actually pay. It breaks the day revenue per megawatt on new deals falls faster than the interest bill on the debt that bought the machines. Watch those two numbers against each other; nothing else about this name matters as much.