Compute Market Brief · 001 · Public inaugural edition

Nebius: the cost of a reported price reset.

A customer-reported October 1 increase puts GPU budgets under review. The next question is whether the new rates apply—and what useful work costs after the change.

BEP Research · September 16, 2026 (Pacific)
Public-price capture: 2026-09-17 00:32:30 UTC · Frozen edition

Download the two-page brief ↓Latest subscriber observations ↗

The research call

Budget for the possibility; verify the applicable quote. The preserved public capture still shows the existing rates. The higher figures below come from a customer repost of a purported email. They remain unverified, and they are not inserted into observed price history. A list-price move alone cannot establish realized pricing, utilization or provider margins.

Nebius · On-demand · USD per GPU-hour

Observed rates and a reported reset

Observed public rateReported Oct. 1 · unverified

H100

$3.85observed

$4.50reported

+16.9%if implemented

H200

$4.50observed

$5.40reported

+20.0%if implemented

B200

$7.15observed

$8.50reported

+18.9%if implemented

B300

$7.85observed

$9.50reported

+21.0%if implemented

All bars start at $0 and use the same $10 scale. Public capture: Sep. 17, 00:32 UTC (Sep. 16 Pacific).

Solid blue: dated public observation. Outlined: customer-reported future rate, not authenticated. These are paired price levels, not a measured time series. Download chart PNG ↓ · Data CSV ↓

01 · What changed

A report, with a meaningful budget impact.

The repost describes increases of 16.9%–21.0% across four NVIDIA GPU types, reportedly effective October 1, 2026. The original customer email has not been authenticated.

For B200, the reported move from $7.15 to $8.50 adds $1.35 per GPU-hour. At 1,000 GPU-hours that is $1,350, before other charges and discounts. The size of the budget sensitivity justifies a contract check; it does not verify the notice.

02 · The interpretation

Separate posted prices from pricing power.

A higher on-demand rate may matter to marginal buyers while reserved contracts and negotiated discounts behave differently. To evaluate pricing power, follow comparable offers, available capacity and realized contract economics together.

A provider-specific reset is also different from a market-wide move. No matched cross-provider panel or customer-billing evidence is established by this notice. This edition makes no revenue, earnings or margin forecast.

Your budget sensitivity

What would the report cost you?

Enter total GPU-hours across all GPUs. Eight GPUs running for 125 hours equals 1,000 GPU-hours.

Total GPU-hours

Illustrative cost sensitivity; not a quote or bill.

H100$3.85 → $4.50 / GPU-hour+$650additional modeled cost

H200$4.50 → $5.40 / GPU-hour+$900additional modeled cost

B200$7.15 → $8.50 / GPU-hour+$1,350additional modeled cost

B300$7.85 → $9.50 / GPU-hour+$1,650additional modeled cost

Reported October 1 rates remain unverified. The calculation assumes unchanged GPU-hours and applicable on-demand terms. It excludes taxes, storage, networking, discounts, migration and changes in workload efficiency. Confirm your region and contract before using it in a budget.

The reported regional scope

The public landing-page quotes do not attach a region to each row. These are the regions named in the customer report; confirm the provider documentation and your actual contract.

GPU	Observed public	Reported Oct. 1	Reported regions
H100	$3.85	$4.50 · unverified	eu-north1
H200	$4.50	$5.40 · unverified	eu-north1, eu-north2, eu-west1, us-central1
B200	$7.15	$8.50 · unverified	us-central1, me-west1
B300	$7.85	$9.50 · unverified	uk-south1, eu-west2, us-north1

03 · For compute buyers

Check the quote, then test the workload.

Confirm GPU variant, region, GPU count, included CPU/RAM, term, tax and effective date.
Compare the new quote with an available alternative under the same requirements.
Measure successful tasks and latency. A lower GPU-hour price can still buy less useful work.

04 · Evidence to watch next

What would change our assessment?

An authenticated original notice or official schedule would strengthen the announcement evidence.
A matching observed quote on or after October 1 would establish implementation for that specific offer.
Permissioned bills, matched competitor offers and capacity evidence would support the next commercial inference.

Sources, scope and corrections

Nebius public price page ↗ — Observed public list rates; dated capture.
Nebius compute pricing documentation ↗ — Regional scope, billing units and tax exclusions.
Customer repost of purported email ↗ — Reported future rates only; original email not authenticated.

This public sample contains only public list-price facts, a labeled report and BEP analysis. It does not expose subscriber source files, licensed market series or account-specific quotes. The edition is frozen; use the price watch for later observations. No corrections recorded at publication. Send a sourced correction to [email protected].

Dated source and calculation record ↓Monitoring methodology ↗

From a market signal to your decision

Bring the compute question your team needs answered.

A focused pilot turns a question into a sourced memo, an inspectable cost model and a briefing. Scope and fees are agreed before work begins.

Explore the compute research pilot ↗Inspect the benchmark starter ↗