Blog · Global · · 4 min read

Why GPU prices keep moving. What pushes the H100 hour up and down.

Why are GPU prices falling, then rising? Five drivers of H100 rental prices, from new chips to power, with dated index readings from 2026.

Why are GPU prices falling? In 2024 and 2025 they did. In 2026 the neocloud H100 index has mostly climbed. The honest answer is that several forces push at once, and they do not always push the same way. This post names the forces and ties each to a dated source.

Down, then up

Silicon Data, a company that sells GPU price data, published its own account on 21 December 2025. It summed up 2023 to 2025 as a move from shortage-driven pricing to oversupply-driven pricing. Its table shows the hyperscaler H100 reading at a median of $9.34 an hour from June to December 2024. For July to December 2025 the median was $6.26. These are its figures, and it has a commercial interest in them.

Then the direction changed. Silicon Data's H100 neocloud index read $2.57 on 4 May 2026. It peaked at $2.79 on 5 June, read $2.75 on 27 July and $2.78 on 4 October. On 2 November 2026 it read $2.94. The hyperscaler reading on the company's index page was $7.12 when read on 3 November.

The index is a guide to direction. It is not a price anyone paid. Casitel holds no executed prices to show you, and neither does any public index.

New chips

A new chip does not make the old chip cheap by itself. Nvidia said on 26 August 2026, with its results for the quarter ended 26 July, that its Vera Rubin platform was in full production. It named CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius as running racks. It reported data centre revenue of $89.0 billion for the quarter, up 117% on a year earlier. Those are Nvidia's figures.

Yet the H100 hour rose while Rubin ramped. Silicon Data's index page, read 3 November 2026, shows the older and newer chips on one ladder:

GPUNeocloud index reading
H100$2.94
H200$3.45
B200$6.50

The gap between chips moves too. Silicon Data's founder wrote on 4 August 2026 that the H200 premium over the H100 averaged 5.8% in May and 12.1% in July. It peaked at 16.2% on 16 July. She said the cause is not established. Memory-bound inference demand is one candidate. A young index with a changing provider panel is another.

Supply, and the power to run it

Supply pulled prices down in the earlier cycle. Silicon Data's December 2025 account credits a large rise in H100 shipments in early 2024, big colocation sites bringing clusters online, and new marketplaces that resold unused reserved capacity. It also says providers had over-estimated long-term training demand in 2023.

In 2026, new supply has been slow to arrive. TechNews reported on 4 August 2026, citing Silicon Data, that H100 rental rates were rising with strong demand, better inference efficiency and long lead times for new data centre capacity.

Power is one reason for those lead times. On 20 October 2026 Silicon UK reported that Microsoft had pulled out of Nscale's planned AI data centre at Loughton, in Essex. UK Power Networks had told Nscale that its grid connection, due at the end of 2026, would slip to the mid 2030s. The site had been planned to start work in 2027. Microsoft declined to comment on its reasons, so the link to the power delay is the press's reading, not a stated cause.

The lesson holds without it. A GPU that cannot be switched on cannot be rented, in any region. Capacity waiting for a grid connection does not ease prices, however many chips Nvidia ships.

Term and utilisation

The same GPU has several prices at one provider, because time is part of the product. Together AI's pricing page, read 3 November 2026, lists an on-demand GPU cluster H100 at $3.99. It lists $3.69 for 3 to 30 days, $3.45 for 31 to 90 days and $3.19 for 91 to 180 days. For longer terms it says to contact sales. DigitalOcean lists $3.26 for a 12-month commitment against $4.41 on demand.

A seller prices a long term lower because a long term fills the hall. Utilisation is the other side. A GPU that sits idle earns nothing, and the cost of owning it does not stop. Silicon Data says marketplaces grew by reselling reserved capacity that its owners were not using, and that this gave the market a low-end price signal.

So a falling or rising index can mean different things. It can be a change in chips, in supply, in power, or in the mix of providers in the panel.

What to take from it

  • No single driver explains a price move. Be careful of anyone who names one.
  • The index shows direction. It does not show the price for your term, your region or your size.
  • Your own term matters more than the market's mood. A 12-month price and a spot price are different products.
  • Read the contract for what the hour includes. How to buy GPU capacity lists the questions.

Exchanges want to let buyers hedge these swings. ICE and Ornn announced GPU compute futures on 19 May 2026, pending regulatory approval. CME announced H100 and B200 rental index futures on Silicon Data's indices on 11 August 2026. As at 21 October 2026 regulators had not cleared them, and the start was reported as 9 November 2026. Index methods are each company's own description.

Sources

Price moving under you? See where your quote sits.

Drop in the quote. See where it sits against the market today, and what to ask for.

Check a price

Read next. More from the compute desk.

Ask an assistant about Casitel.

ChatGPT, opens in a new tabClaude, opens in a new tabPerplexity, opens in a new tabGemini, opens in a new tabGrok, opens in a new tabCopilot, opens in a new tab

The question: What is Casitel (https://casitel.com) and what does it do for sites, capital and tenants in digital infrastructure? Use casitel.com as the source.