Recompute your cost per token before Q4 GPU reservations lock in stale costs

A reservation can lock in yesterday's arithmetic

GPU reservations trade committed capacity for a fixed rate, but a fixed rate can become a stale rate when on-demand prices move. The arithmetic is simple: cost per token equals the GPU-hour rate divided by tokens per GPU-hour, so the percentage change in rate becomes the percentage change in token cost when throughput holds. Nebius, an AI cloud provider, is repricing its Nvidia GPU fleet on October 1, the second on-demand reset in three months.

The H100 moves from $3.85 to $4.50 per GPU-hour, a 16.9% increase. H200 and B200 rates move to $5.40 and $8.50 per GPU-hour, increases of 20% and 18.9%. The B300 jumps from $7.85 to $9.50 on October 1, a 21% increase.

The May reset averaged 29%, and the B300 has gained roughly 56% since the pre-May rate of about $6.10 per GPU-hour. Nebius reported second-quarter revenue of $582.3 million, a 454% year-over-year rise, and AI cloud revenue of $574.9 million, a 514% rise. Revenue growth gives a provider leverage when capacity is tight, and it makes the next reservation price a negotiation variable.

Utilization changes the same arithmetic. If you rent a GPU-hour but only use part of it, the effective rate rises, and the cost per token rises with it. The inference break-even throughput is the tokens per GPU-hour at which your token price covers the GPU-hour rate plus your margin. Below that point, the workload is paying for idle capacity. Above it, the reservation can still be worth signing if the rate is competitive.

The GPU hour to token cost is the number your finance team should see, not the raw GPU-hour rate. It is also the number that makes a reservation comparable across providers, because it combines capacity price and inference efficiency into a common unit.

Run the re-pricing checklist before you sign

A reservation is a commitment to a price, not a commitment to a market. The checklist below is the compute cost re-run you need before the commitment becomes the budget. It works because it separates the rate you pay, the throughput you get, and the margin you need.

1. Recompute the new GPU-hour rate for each tier you rent, using the provider's current on-demand price sheet.
2. Derive tokens per GPU-hour from your measured inference throughput, not from a vendor demo.
3. Calculate cost per token by dividing the new GPU-hour rate by tokens per GPU-hour.
4. Compare that token cost to the cross-provider index and your margin, then renew, re-bid, or resize the workload.

The rate item catches an outdated reservation price. The throughput item prevents demo numbers from setting the cost, because throughput changes with model size, batch size, quantization, and request mix. The token-cost item converts capacity into the unit you sell, and the index item compares the provider's list price to the wider market.

Each item is testable quickly because it uses numbers you already own: the provider rate, your measured throughput, and your margin. The cross-provider index gives you an external reference, while your margin gives you the decision threshold. If the new token cost exceeds your margin, do not renew at that rate. Re-bid against the index, resize the workload to a cheaper tier, or move to on-demand until on-demand pricing settles.

Use the index as a bid, not a forecast

The Mercatus GPU Index, a cross-provider price tracker, recorded an average B200 on-demand price of $6.39 per GPU-hour across 12 providers on September 13, 2026. A cross-provider index is useful because it turns a provider's list price into a market comparison. It is not a promise that every provider will match the average, but it is the reference you use when you ask for a better reservation price.

If your current provider's new rate is far above the index, that difference is your negotiation evidence. If your workload can run on a lower tier, the index also tells you whether the cheaper tier is actually cheaper after throughput differences. Keep the comparison compact: rate per GPU-hour, tokens per GPU-hour, cost per token, index price, and margin. When the delta is larger than your margin, do not sign without a better number. Re-bid, resize, or wait for on-demand pricing to settle.