Measure The Runway Hit Before Switching GPU Providers

Nebius’s October 1, 2026 on-demand GPU price increases are now effective, with H100, H200, B200 and B300 rates up roughly 17%–21%. Recalculate startup runway using the additional cost of exposed usage—not the headline percentage. A 20% GPU price increase does not mean a 20% increase in company burn, and the price change alone is not a reason to migrate.

Select your GPU and monthly billed GPU-hours; add cash and burn to calculate the runway change.

The Increase Applies To Specific Resources And Rates

As of October 6, 2026, Nebius’s public pricing page lists these USD on-demand rates per GPU-hour. Percentage increases are calculated from those prices.

GPU	Old Rate	New Rate	Increase
H100	$3.85	$4.50	16.9%
H200	$4.50	$5.40	20.0%
B200	$7.15	$8.50	18.9%
B300	$7.85	$9.50	21.0%

These are per GPU, not per multi-GPU VM, and exclude applicable taxes. The same page shows unchanged on-demand rates for RTX PRO 6000 and unchanged starting prices for L40S configurations.

The change also reaches some CPU-only workloads. Nebius’s compute documentation identifies increases for non-GPU AMD Epyc Genoa CPU and RAM, and non-GPU Intel Ice Lake RAM. Audit those resources separately; do not apply a blanket GPU percentage to the whole invoice. The figures above do not quantify the CPU and RAM increases.

For commitments or negotiated pricing, check your agreement and account-specific quote. A public list-price change does not establish what happens to an existing contract. The documentation also specifies separate ILS pricing for companies from Israel; the USD table is not a substitute for those rates.

Billed GPU-Hours Determine The Cash Impact

Start with billed running time, not GPU utilization. Nebius bills running VM compute by the second; stopped VM compute is not charged, but storage remains chargeable. A running GPU waiting for work still belongs in your cost model.

For each affected GPU type, additional monthly cost equals forecast billed GPU-hours multiplied by the difference between the new and old effective rates.

Use your payable rates after applicable discounts—not automatically the list prices—and model taxes and credits separately. For a mixed fleet, calculate each GPU type’s change and add the results. The calculator above models one GPU type at public USD list prices, with unchanged usage and no discounts or taxes.

Eight Running H100s Add $3,796 Per Planning Month

Assume a standardized 730-hour planning month. Eight continuously running GPUs produce 5,840 GPU-hours:

Monthly Measure	Amount
Old Cost: 5,840 × $3.85	$22,484
New Cost: 5,840 × $4.50	$26,280
Additional Cost	$3,796

Suppose the startup has $1.2 million in available cash and previously burned $100,000 net per month, including the old GPU cost. With all other cash flows unchanged, monthly net burn becomes $103,796.

Previous runway is $1,200,000 divided by $100,000, or 12 months. Revised runway is $1,200,000 divided by $103,796, or about 11.6 months.

The H100 rate rose 16.9%, but company net burn rose only 3.8%. That distinction keeps the response proportional: the company has lost some runway, not one-sixth of it.

This is a static illustration, not a cash forecast. Build a monthly schedule if usage grows, annual payments come due or credits expire. Credit expiry is a separate cash shock; do not hide it inside the price increase. Use the burn-rate and runway guide to distinguish cash burn from accounting expense.

Remove Paid Idle Time Before Changing Infrastructure

The first savings candidate is running compute that produces no useful work. Assign ownership of development VMs, unused environments and oversized deployments. Stop compute when it is not needed while preserving necessary data. Storage charges persist, so stopping a VM does not make its entire footprint free.

Then benchmark cost per useful result. For training, compare cost to reach the required quality target. For inference, compare cost per successful customer task at acceptable latency and quality.

Batching, model changes and different hardware are candidates to test—not assumed savings. A cheaper GPU-hour can require more hours. Any comparison that holds hourly price constant but ignores throughput, retries or quality can point toward the wrong provider or model.

Spot Capacity Requires An Interruption Budget

Nebius says preemptible VMs may stop at any time and do not guarantee availability. The same documentation warns that local SSD data is lost when a preemptible VM stops, so checkpoint to durable storage.

The public pricing page also says spot prices are dynamic. Budget checkpointing, retries and fallback capacity before treating advertised “from” prices as durable savings. Spot capacity belongs in the comparison only where the workload can tolerate interruption; it is not a like-for-like replacement for every on-demand deployment.

Compare Commitments And Migration On Cash Payback

Nebius advertises commitment discounts of up to 35% for large-scale, multi-month reservations. Request the actual price, payment schedule, minimum obligation and flexibility terms. The maximum advertised discount is not an account-specific quote. A lower rate can still reduce financial flexibility if demand falls.

For migration, cash payback in months equals one-time incremental switching cash cost divided by expected net monthly cash savings.

This calculation requires positive monthly savings and assumes they remain stable. Include additional engineering spending, overlapping infrastructure, data movement and validation costs. Existing salaries are not automatically incremental cash costs, but diverted engineering time still matters.

If switching postpones a customer launch, model the effect on receipts and milestones separately rather than treating migration as free. The draft evidence provides no migration-cost figure or alternative-provider quote, so it does not establish a migration payback period.

Update the operating forecast with three explicit figures: additional monthly cash cost, revised runway and cost per customer task. Assign owners and dates to the savings you expect. Optimize, negotiate or move according to those numbers—and whether the change preserves the milestone the cash is meant to fund.