Is it a fit?Book a call

Journal3 min

Rented GPUs got a fifth dearer in a month

In October the price of renting an H100 rose about 25% in a month, while token prices fell a quarter in four weeks. Two prices, both set by someone else.

A row of server cabinets in a long, dim server room

The cost of AI has two meters, and in October they ran in opposite directions.

Compute went up

Ornn’s index of transacted GPU rental prices settled on 26 October with an H100 at USD 3.17 an hour, up 24.8% in one month. An H200 rose 29.9% to USD 6.21, a B200 21.4% to USD 9.76. Nebius raised its own list prices from 1 October by 17% to 21%: an H100 from USD 3.85 to 4.50 an hour, an H200 from 4.50 to 5.40.

Tokens went down

Ramp’s AI Index, from its own customers’ card and bill data, saw the blended price businesses pay the two largest model vendors fall to USD 0.64 per million tokens, 25.4% lower than four weeks before. Token volume rose; token spend barely moved.

What the two have in common

Neither price was set by the company paying it. In four weeks the vendors’ shares of business spend moved six points, prices moved by a fifth or more, and a rented GPU cost a quarter more than it did a month earlier. A plan built on either meter is a plan built on someone else’s pricing.

An owned model changes which costs you carry. On your own servers, with Eigna Local, the cost is the machine, its power and a fixed running fee. A meter does not appear.

Where this touches us

Honesty first: Eigna Hybrid and Eigna Cloud train and run on dedicated rented GPUs in the region you choose, so these prices matter to us as well. It is one reason our prices are still marked proposed, and why we measure throughput per machine before we publish a cost per answer.

The lesson we take from October is not that renting is wrong. It is that the cost of a model you rely on every day should be one you can plan, on a model that does not change until you decide.

Notes

  1. Ornn Compute Price Index, settled 26 October 2026: H100 SXM, H200, B200 and A100, volume-weighted transacted prices per GPU hour.
  2. Nebius prices, from 1 October 2026; earlier prices as recorded by usagepricing.com.
  3. Ramp AI Index, 14 October 2026, token data to the week ending 11 October; Ramp notes its sample is its own customers. Commentary: Ara Kharazian.

Read next

Newsletter

Release notes.

Once a month: what we released, what we held, and why. Short, dated and true, losses included.

Double opt-in. One click to leave. We never share the list.

Each issue holds

  1. One release or testWhat passed the gate, what was held, and the reason in one line.
  2. One Journal pieceAn explainer from the people who build it.
  3. One numberWith its baseline and its date, never on its own.

The first issue comes in November 2026.

Start

Bring one task. See the proof first.

In about four weeks, on your own data in a sealed setup, we show whether an Eigna expert beats your baseline. If it does not, you keep the report.

Book a callRequest a Test Lab