All news

Early access to Hopper chips becomes a cost advantage

A row of white blocks on a pale stone table with one cobalt block placed ahead of the line

Nvidia's H100 is in full production, with cloud instances due next year. Teams that train on it first will buy the same AI work for less.

Nvidia's H100 promises the same AI work for a third of the cost, and it is now in full production. On 20 September the company said the data centre GPU, built on the Hopper architecture, cuts total cost of ownership threefold against the previous generation for the same AI performance; partners including Dell, Hewlett Packard Enterprise, Lenovo and Supermicro ship the first systems in October.

Compute gets cheaper one chip generation at a time, and the first buyers capture each step. If the threefold claim holds in practice, a lab training on H100 next year gets the same result for about a third of what a rival pays on older chips, or a much larger model for the same budget. Access, not price, decides who collects that gain.

The clouds set the timetable. AWS, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure are among the first to offer H100 instances, starting next year, and those that install the chip first will set the price of large-scale compute over the coming year and win the largest customers, the pattern locked-in chip supply already showed. Its Transformer Engine targets the large language models now driving demand, the workloads that carry the biggest bills.

Nvidia's own numbers show where the money has moved. In its last quarter data centre revenue grew 61% to $3.81 billion while gaming revenue fell 33%, and five days before the keynote Ethereum completed its switch to proof of stake, ending mining on the network that had driven crypto demand for graphics cards. Consumer graphics weakens as AI in the data centre grows. Games still matter: the same keynote launched GeForce RTX 40 cards that use AI to generate whole frames, and with neural rendering in consumer hardware there is room for tools built around AI-generated graphics.

Compute belongs in the strategic plan, not the monthly bill. A team whose roadmap depends on training should reserve cloud capacity on the new chips early, or negotiate hard on previous-generation hardware, which will get cheaper as H100 arrives. Software that makes models use new hardware features, such as lower-precision arithmetic and transformer-specific optimisation, has a clear market.

A threefold saving remains a vendor's figure until customers can rent it. When H100 instances go live on the major clouds, their hourly price against the previous generation will show how much of the gain reaches buyers. The US licence requirement for A100 and H100 sales to China, which Nvidia disclosed in August, puts part of its data centre growth in question. For smaller teams the decisive number sits elsewhere: prices of used and previous-generation GPUs after the end of Ethereum mining decide how cheaply they can train.

Like a staircase, the price of AI compute falls one chip generation at a time, and the gains go first to those with access. Fanaura's approach to AI-first founders favours teams that plan compute as carefully as product and turn each step down in cost into a better product rather than a thinner margin.

For founders

Building software that gets more out of new AI hardware, or tools for AI-generated game graphics? Send us the project.

Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.

Apply in five minutes
Subscribe

Get the news by email

One short note each weekday. No marketing, and you can unsubscribe at any time.