Why Nebius Is Keeping Some 2027 AI Capacity Off the Shelf
Nebius mixes long-term contracts, short-term premiums and auctions. Holding back capacity pays only if the servers stay busy, and it is not yet clear the pricing will last.
AI servers can be sold like hotel rooms
Picture a hotel with 100 rooms. It could sign all of them to corporate clients for a year at a fixed nightly rate. Revenue would be stable.
But when a major international event comes to town, nightly rates can triple. A hotel that kept some rooms back can charge far more. If demand for the event falls short, those rooms sit empty.
A similar pricing strategy is now showing up in AI data centers.
AI companies rent servers loaded with Nvidia GPUs to train and run their models. When supply is tight, customers who need computing power right now are willing to pay extra. Nebius Group (NBIS), an AI cloud company, is trying to turn that gap into profit.
Three ways Nebius sells capacity
In its second-quarter results and call on August 12, 2026, Nebius laid out pricing by contract type.
| Contract type | Price |
|---|---|
| Long-term contracts | $20 million to $25 million per MW, annual contract value |
| Short-term contracts (negotiated) | $40 million to $50 million per MW, annualized contract value |
| Blackwell capacity auction | 15% above the prior highest sale price |
Short-term prices are negotiated figures, and not all capacity has been sold at those levels. Per-MW amounts are on an annualized contract value basis.
First, long-term contracts bring stability. Customers commit to using capacity for a long time. The company can forecast revenue and more easily raise capital for equipment spending against those contracts.
Second, short-term contracts bring higher prices. Customers that need large GPU clusters for a few months pay a premium. Nebius said it is negotiating contracts of up to about six months at $40 million to $50 million per MW on an annualized basis.
Third, auctions are price discovery. Nebius put part of its Blackwell GPU capacity up for auction, and it cleared at a price 15% above the previous highest sale price. That does not mean overall contract prices rose 15%.
The point is the mix. Long-term contracts secure base revenue and a financing foundation, while some capacity is left open for a chance at high short-term prices.
Are GPU prices really rising?
Ornn, which tracks compute transaction prices, showed an Nvidia B200 rental price of $9.58 per hour as of October 23. That was up 20.5% from 30 days earlier.
The older H100 rented for $3.17 per hour, up 3.9% over the same period.
30-day change in GPU rental prices (Ornn, as of October 23, 2026)
- H100: +3.9%
- B200: +20.5%
Per GPU, the B200 costs about three times as much as the H100 to rent. But the B200 differs in performance and memory configuration, so this should not be read as three times the cost of doing the same work.
The data is evidence that scarcity of the newest GPUs is showing up in short-term rental prices. Still, that price cannot be applied to every company's long-term contracts.
Does double the price mean double the profit?
Not necessarily. But it can improve profitability a lot.
Take a hypothetical AI data center. On the same equipment, annual revenue is 10 billion won and operating costs plus depreciation are 8 billion won, so operating profit is 2 billion won. If higher prices lift revenue to 14 billion won and costs stay flat, operating profit rises to 6 billion won, three times as much.
That is operating leverage.
The reverse also holds. If servers sit idle while the company waits for a better price, the picture changes. GPU depreciation, power contract costs and financing costs can keep running with no customers.
A good strategy, then, is not simply to sell high. It is to optimize price and utilization together.
Does this favor neoclouds over Big Tech?
That calls for caution.
Microsoft, Amazon and Google have enormous capital, existing customers and their own cloud platforms. Having many long-term contracts does not mean they lack pricing power. They also run usage-based pricing and short-term contracts.
Specialized AI cloud companies such as Nebius, by contrast, have a chance to sell capacity flexibly against customers' urgent demand. But they also face higher financing costs and growing dependence on particular customers.
The contest is less about size than about how well a company turns the capacity it actually holds into strong net income and cash flow.
What to watch
The key question is whether AI data centers can keep selling at high prices. Nebius has shown part of the possibility, but whether it is a durable edge will have to be tested in future results.
Competition in AI infrastructure is widening from the ability to secure GPUs to the ability to turn those GPUs into more cash than the investment cost.
Insight Times Editorial Desk
