Recompute cost per inference before Nebius's 41% memory jump

The rate change is a margin event

Nebius gave customers notice on September 17, 2026 that pay-as-you-go rates for Nvidia GPUs, AMD CPUs, and memory would increase on October 1, 2026. Nebius's October 1, 2026 adjustment is its second pay-as-you-go price rise in roughly three months, with the steepest jump being the roughly 41% increase in memory-backed capacity. The increase follows demand for computing power rising faster than supply.

For a bootstrapped AI product, that sequence matters. A GPU-hour rate increase changes the input cost of every inference. A memory-backed capacity jump changes the cost of workloads that depend on that capacity. The question is which part of the product absorbs it.

Read the bank balance before you read the dashboard. Profit is an accounting line; cash burn is the number that keeps the lights on. A rate change can leave profit looking stable while the next invoice makes the cash line thinner.

The checklist is the decision

Run this pass before the new rates hit your invoice. Each item is a quick check, and together they force a choice: pass through, protect margin, or commit.

• Confirm which rate lines changed in the new card.
• Check whether your workload mix is dominated by GPU-hour, CPU-hour, or memory-backed capacity.
• Check the selected Nvidia GPU-hour increases: H100 about 17%, H200 20%, B200 about 19%, B300 21%, a 16.9% to 21.0% range.
• H100: note the move from $3.85 to $4.50 per GPU-hour.
• HGX B300: note the move from $7.85 to $9.50 per GPU-hour.
• AMD EPYC Genoa: note the move from $0.10 to $0.13 per hour.
• Memory-backed capacity: flag the largest single increase.
• Compute cost per inference for your highest-volume workloads at the new rates.
• Compare each cost to customer price and set a pass-through threshold.
• Calculate margin impact at the current customer price.
• Use Nebius's full-discount rates of $2.93 per H100 GPU-hour and $6.18 per B300 GPU-hour to compute commitment break-even; multi-month reservations for large clusters can earn discounts up to 35% off on-demand.

Start with recent inference logs. Group requests by model, token length, concurrency, and serving stack. Divide the cost of each group by the number of completed inferences. Then multiply by the new rate for the same hardware. That gives you a GPU cost per inference you can defend.

The memory line deserves extra care. The public card did not show a memory before-and-after. Treat that uncertainty as a reason to verify your own invoice, not as a reason to assume the worst.

Set the pass-through threshold as a percentage of price. If the new cost per inference crosses that line, the product either raises price, trims the workload, or stops selling that tier. Margin impact is the difference between the old cost per inference and the new cost per inference, measured against the same customer price. If the price stays flat, the added inference cost comes straight out of margin. Put that number next to the customer price before you choose a response.

Commitments are a cash choice

A commitment discount can look like a margin rescue, but it is a cash decision. That discount only helps if your usage stays high enough to use the reserved capacity.

The May 2026 move already lifted on-demand capacity by an average of 29%, while spot capacity rose 51%. Nebius's first capacity auction cleared 15% above the prior highest price, and management said it could sell all 2027 capacity on current terms. In that market, a multi-month reservation is a forecast you are paying for in advance.

If you are choosing between neocloud and hyperscaler capacity, the comparison should be made at the workload level. AWS's October 7, 2026 capacity-block pricing by accelerator in available regions is $5.970 for P5, $6.866 for P5e, $7.895 for P5en, $14.208 for P6-B200, and $16.146 for P6-B300. Use that comparison only after you know which accelerator your inference actually needs.

If your customer price is stable and your inference volume is predictable, a commitment can protect margin. If your volume is still learning, on-demand keeps you honest. Slower growth is a deliberate choice when the next cohort of users would consume more inference than the price can cover. Check the first invoice against the threshold you wrote down.