Short-Term AI Compute Contracts Signal a Spot-Pricing Regime
Nebius and CoreWeave have turned short-term AI compute contracts into the hottest product in cloud infrastructure, selling three-to-six-month GPU capacity at premiums the market did not expect a year ago. The two Nvidia-backed neoclouds pitched the strategy to investors this week as GPU shortages keep demand ahead of supply, and both stocks moved sharply higher on their second-quarter prints: Nebius gained 34% and CoreWeave 19%.
The economics are the story. At Nebius, short-term capacity now commands about $40 million in annual contract value per megawatt. A year earlier, the market expected that figure to land between $10 million and $15 million. CoreWeave's July 2026 repricing added roughly 25% to the cost of new capacity, and near-term supply remains fully booked. Nebius management says it could sell all of its 2027 capacity today, but the company is withholding part of that supply to keep selling short-duration deals at today's premiums. It is a direct bet on continued scarcity.
Short-Term AI Compute Contracts vs the Five-Year Hyperscaler Model
The move breaks with the template that has defined the AI buildout so far. AWS continues to sign five-year commitments, the kind of long-term agreements that deliver revenue stability and make it possible to finance enormous capacity upfront. The neoclouds are betting the opposite: that scarcity persists, that GPU supply stays tight, and that locking in today's prices means leaving money on the table next quarter. Both companies' earnings calls pointed to a clear shortfall in AI cloud capacity, with demand running well ahead of supply.
Two separate pricing and financing systems for AI infrastructure are emerging from that contrast. CoreWeave still derives more than 98% of its business from multi-year take-or-pay contracts, according to Morgan Stanley, so its pivot is incremental rather than wholesale. Nebius has always carried heavier exposure to the spot market, and CEO Arkady Volozh's decision to reserve 2027 capacity is a deliberate escalation of that stance. Each company is effectively hedging a different view of where prices go from here.
Two pricing regimes now coexist in the same market. Hyperscalers sell certainty: fixed prices, long lock-ins, predictable budgets. The neoclouds are beginning to sell scarcity: premium prices, short windows, and the option to reprice as conditions change. Buyers face a genuine fork, and the choice carries consequences on both sides of the ledger, which is why the two models are now being valued so differently.
Margins, Budgets, and the Cost of Certainty
On the sell side, the numbers explain the enthusiasm. CoreWeave's second-quarter revenue doubled to $2.58 billion, and its backlog now exceeds $100 billion. Nebius posted an adjusted EBITDA margin of roughly 50% for its AI cloud business. A contract that pays $40 million per megawatt per year, against a $10-15 million expectation from a year ago, rewrites the unit economics of every cluster these companies deploy and strengthens the case for the debt-funded expansion.
| Metric | CoreWeave | Nebius |
|---|---|---|
| Share move after Q2 2026 results | +19% | +34% |
| Q2 2026 revenue | $2.58 billion, doubled year over year | n/a |
| Backlog | Above $100 billion | n/a |
| Contract posture | 98% multi-year take-or-pay | Higher spot exposure; 2027 capacity held back |
| Short-term pricing signal | New capacity priced about 25% higher in July 2026 | Around $40M annual contract value per MW vs $10-15M a year earlier |
| AI cloud adjusted EBITDA margin | n/a | About 50% |
On the buy side, the trade-off is sharper. Enterprises that sign three-to-six-month deals trade budget certainty for flexibility and lower exposure to long lock-ins. The re-pricing risk is real: a short contract means the buyer absorbs the next price increase, while a five-year AWS commitment caps it. But neoclouds still undercut the major clouds by up to 66% on comparable GPU capacity, so even premium spot pricing can look cheap next to a hyperscaler invoice. The open question is whether that discount survives the next round of scarcity.
For enterprise buyers, the practical effect lands in the budgeting cycle. Finance teams that priced AI workloads on twelve-month contracts a year ago now face a market where the same capacity can cost materially more at renewal, and where the cheapest option in any given quarter may carry the least protection against the next price move. That favors a two-tier procurement strategy rather than a single bet on either model.
Managed inference is where the stickiness comes from. CoreWeave's contracted annual recurring revenue for managed inference services grew from $1 million to over $100 million in a single quarter. These services carry higher margins than raw GPU rental and raise switching costs, which cushions both companies if the spot market softens. That shift in revenue mix is part of why the market has stopped valuing CoreWeave purely as a GPU rental business.
The Debt-Funded Buildout Hangs on the Useful Life Debate
The pivot to short-term pricing sits on top of an unusually leveraged supply chain. CoreWeave and Nebius fund their data center buildout through a mix of Nvidia equity, hyperscaler commitments, and GPU-backed debt, and the scale of those commitments is enormous relative to current revenue. Microsoft alone has struck roughly $60 billion in deals across CoreWeave, Nebius, and other private providers such as Nscale, and the two public neoclouds together hold about $122 billion in hyperscaler contracts.
The debt math depends on GPUs holding their value. CoreWeave has argued on its earnings call that the useful life of previous-generation chips keeps extending, and the contract book supports it: inference workloads are still absorbing older Nvidia A100s, and contracts for those chips run through 2029. That feeds directly into the bulls-versus-bears debate. For bulls, re-contracting older hardware at strong prices means the assets backing the debt keep earning well past their original depreciation schedule. For bears, a market that prices capacity on three-month windows is the same market that can reprice those assets downward the moment supply catches up.
The tension between the two business models is the real story. If scarcity holds and spot pricing persists, the neoclouds capture a windfall their fixed-price rivals cannot match. If the shortage eases, the same short contracts that generate today's premiums become the fastest route to a revenue cliff, because there is no five-year backlog to smooth the fall. CoreWeave's $100 billion backlog gives it a cushion; the 2027 capacity Nebius is deliberately holding back does not. For hyperscalers, the neoclouds' willingness to take repricing risk keeps the pressure on their own margins, since spot pricing establishes a public reference point for what AI compute is actually worth.
The Verdict
The shift to short-term AI compute contracts is a rational response to a genuinely undersupplied market. The $40 million per megawatt figure is the market's way of saying GPU capacity is priced like a commodity in shortage. CoreWeave's near-term bookings confirm that diagnosis. The risk is symmetrical: the same instruments that capture today's premiums expose both companies, and their lenders, to the first real downturn in AI compute pricing.
For decision-makers, the practical takeaway is to treat short-term AI compute contracts as tactical procurement, not as a substitute for the budget certainty that long-term hyperscaler contracts provide. Reserve a core of multi-year capacity for baseline workloads, and use the neoclouds' spot pricing for burst demand where flexibility matters more than cost stability. The market is now offering two different products; the mistake is assuming they are interchangeable.
Why This Matters
AI compute is moving from a reserved-capacity model toward a spot market, and anyone buying GPU capacity now carries repricing risk that did not exist a year ago. The next two quarters will show whether supply catches up with demand, because the same mechanism that is printing record premiums for Nebius and CoreWeave will be the first to transmit a reversal through hyperscaler margins, enterprise AI budgets, and the GPU-backed debt financing the buildout.
Photo by Brecht Corbeel on Unsplash
Related Articles
- Nvidia Nebius Investment: 9.3% Stake in Neocloud Platform Signals Strategic Shift
- CoreWeave Revenue Backlog Hits $99.4 Billion as Infrastructure Costs Drive Losses
- Nebius Strengthens AI Infrastructure with $643 Million Acquisition of Eigen AI
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.