All news

Locked-in chip supply shows AI compute will stay scarce

A half-open graphite drawer holds only a few silicon chips on white felt, one of them with a cobalt sheen

Nvidia's data center sales rose 71% and its supply commitments more than tripled. Access to AI compute is now a strategic asset, not a commodity.

Nvidia has locked in $9.00 billion of inventory purchases and long-term supply obligations, up from $2.54 billion a year earlier. The figure came with results for the quarter ended 30 January, reported on 16 February, in which data center revenue rose 71% from a year earlier to a record $3.26 billion.

Commitments that more than tripled in a year are a supplier's forecast written into contracts. Nvidia attributes them to longer lead times throughout the supply chain, and a company that books manufacturing capacity this far ahead expects demand to last. Buyers of AI compute now compete for supply, not only for price, which turns access to compute into a strategic asset rather than a commodity.

The demand is broad. Ampere-generation GPUs drove the growth, used for both training and inference in cloud computing and in AI workloads such as natural language processing and recommender models. The cloud providers that reserve that capacity first will set prices for everyone else. Nvidia earns a gross margin of about 65% across its business, and that margin is built into the price of every GPU hour an AI company rents.

Regulators have drawn a line around that position. Eight days before the results, Nvidia and the owner of Arm terminated their deal, worth about $40 billion, because of significant regulatory challenges; the US Federal Trade Commission had sued to block it in December. Authorities will not let the leading AI chip supplier control a processor architecture its rivals depend on, and that keeps room open for competing chips and designs.

Compute belongs in a startup's plan as a strategic input with lead times. A team whose roadmap depends on training needs capacity commitments early, through cloud credits or reserved instances. Models designed for efficiency from the start protect the business once a product scales, because inference cost decides gross margin. Software that lets models run well on several kinds of hardware gains value as large buyers look for a second source, and the workloads named as growth drivers, language models and recommendation systems, show where the next application companies will need the most compute.

Scarcity could still ease faster than the contracts imply. Whether supply obligations keep rising or lead times shorten during the year will settle it, along with the new data center products Nvidia has promised for its developer conference later this month and how they are priced. Alternative accelerators and cloud providers' in-house chips are the other test, since they decide whether buyers get a credible second source.

Fanaura is a family office that treats the long horizon as an edge, and over that horizon AI compute behaves like a scarce industrial input with one dominant supplier. Our venture thesis therefore favours AI-first teams that make compute access and efficiency part of strategy, and tools that give buyers a real choice of hardware.

For founders

Building software that runs AI models efficiently across different chips? 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.