Nebius fell about 15% when Bloomberg reported Meta was exploring its own cloud business. It fell 13% again a couple of weeks later. Then on Monday it jumped 16% after Nvidia disclosed a 9.3% stake, dragging CoreWeave up 8% in sympathy.
None of those moves reflected a change in how much AI compute the world needs, which is the thing these companies supposedly sell. They reflected changes in the market's confidence about three things that get far less attention than the contract announcements: who is carrying the capital risk, whether the power arrives on schedule, and whether the machines run hot enough to make money.
Those are the variables. The headline numbers mostly are not.
Start with the math that decides everything
GPU cloud economics are not a gentle curve where more utilization means proportionally more profit. They are a threshold.
By one analysis, a 1,024-GPU H100 cluster at competitive market rates needs around 75% utilization to cover costs. At 70% it loses roughly $80,000 to $130,000 a month. At 90% it generates $170,000 to $220,000 a month in profit.
Twenty points of utilization is the difference between bleeding six figures monthly and earning six figures monthly, on the same hardware. The specific numbers move with rental rates, power and colocation costs, but the shape holds across the sector.
This is why a $27 billion contract announcement is less informative than it sounds. Nebius has a Meta agreement worth up to about $27 billion over five years, and CoreWeave has one worth roughly $21 billion. Those are real commitments. They also say nothing about whether the clusters serving them run at 70% or 90%, and that gap determines whether the contract is profitable.
Then look at what is actually scarce
The constraint moved, and the sector's own language has caught up: this stopped being a scramble for GPUs and became a fight over electricity.
CoreWeave and Nebius have each secured 3.5 gigawatts of contracted power capacity, with the vast majority not yet online. CoreWeave is targeting 1.7 GW of active power by the end of 2026, while Nebius targets 800 megawatts to 1 gigawatt of connected power.
Read that as a conversion problem. Both companies have contracted roughly the same power footprint, and both expect less than half of it to be energized by year end. Backlog cannot convert into revenue faster than substations, transformers and interconnection agreements permit, and those run on utility timelines rather than software timelines.
So the honest scorecard for these companies is not contract value. It is megawatts energized per quarter, and then utilization on those megawatts. Everything else is a claim about the future.
The part almost nobody frames correctly
Here is the structural feature that explains more about this sector than any technology argument, and it is fundamentally about accounting.
Meta is expected to generate roughly $136 billion in cash from operations this year against capital expenditure guidance of $125 billion to $145 billion, which could leave it free cash flow negative. It also holds up to $62.2 billion in neocloud agreements. If Meta built the equivalent capacity itself, that spending would be balance sheet capex. Instead, neocloud agreements add nothing to capex, because the costs are recognized as operating expenses over the life of the contracts, which extend into 2031 and 2032.
Same compute. Same electricity. Radically different financial statements.
That means part of what neoclouds sell is not speed or GPU access at all. It is balance sheet absorption. A hyperscaler under pressure to show free cash flow can move a multi-billion-dollar capital program off its own books and onto someone else's, converting a lump of capex into a smooth stream of opex.
This is a genuine service, and it is priced accordingly. It also means the neocloud is holding the risk the hyperscaler declined to hold: the asset, the depreciation, the utilization exposure, and the residual value question of what a 2026-vintage GPU cluster is worth in 2031.
And it explains why the Meta Compute report hit so hard. If Meta is willing to build and sell capacity itself, the arbitrage it was paying for is worth less than the market assumed.
Where Nvidia sits in this
Monday's 16% jump came from Nvidia disclosing a 9.3% stake, and the market read that as validation. It is worth being precise about what the relationship actually is.
Nvidia provides capital or guaranteed purchase commitments to neocloud providers in exchange for a share of cloud revenue, meaning it both sells the GPUs and participates in the revenue generated by renting them. It has taken equity stakes rather than merely providing demand floors, and in some structures collects a percentage of cloud revenue as a return on financing, beyond the hardware sale.
That arrangement is defensible. Nvidia has a legitimate interest in seeing its chips deployed rather than sitting in warehouses, and financing capacity buildout accelerates its own market. Analysts covering it are direct that these are not arm's-length supplier deals.
What it means for the neocloud is that a single counterparty is simultaneously its supplier, its shareholder, and sometimes its revenue-share partner. Equity from your supplier is a vote of confidence and a concentration, and the market has so far priced only the first.
The scale of the surrounding commitment is worth noting too: Nvidia is guiding to $190 billion of capital expenditure in calendar 2026 against roughly $200 billion in forecast cash from operations, which would consume about 95% of operating cash flow. This is an ecosystem in which nearly everyone is spending nearly everything.
Not all of these companies are the same trade
The sector gets traded as a basket, and it should not be.
The clearest distinction is financing structure. Nebius's asset-light pivot contrasts with the capital-intensive, debt- and dilution-heavy, balance-sheet-owned models used by CoreWeave and the miner-pivot names, which fund GPU buildouts on their own books.
That difference determines what happens if utilization disappoints. A company that owns the assets outright and financed them with debt has fixed obligations against variable revenue. A company that has structured around less balance sheet ownership has more flexibility and, typically, less margin capture when things go well.
The miner-to-AI cohort illustrates how little contract news protects you. IREN signed a five-year, $3.4 billion Nvidia cloud contract plus up to $2.1 billion in Nvidia investment, and the stock still fell roughly 37% over a month. Applied Digital posted a 61% revenue beat and slid with the group. When the market repriced financing risk, announcements did not help.
What would actually change the picture
Three measurable things, in rough order of importance.
Megawatts energized against plan, because that is the binding constraint and it is verifiable quarter to quarter.
Utilization rates on live capacity, which determine whether contracts are profitable and which companies disclose inconsistently, so the disclosure itself is informative.
And financing terms on new capacity, particularly the cost of debt and the extent of dilution, because that is what separated the winners from the losers in the recent selloff far more cleanly than any contract did. The demand for AI compute is not seriously in question, and none of the volatility of the past month suggested otherwise. What is in question is who ends up owning depreciating hardware, whether it runs above the profitability threshold, and how much of the eventual margin flows to the company that bought the chips versus the company that sold and financed them. Those questions get settled slowly, on utility timelines and utilization reports, not in the reaction to a single headline.