The labs go looking for small sites
Anthropic and OpenAI have spent a year announcing gigawatts. Both are now hunting 20 to 30 megawatt sites, because the scarce thing is not capital but power that already exists.
Aleksandr Popov via UnsplashEvery announcement in AI infrastructure for the past year has been measured in gigawatts. The interesting move this week is three orders of magnitude smaller.
Anthropic and OpenAI are both looking for data centre deals in the range of 20 to 30 megawatts, CNBC reported on Friday, citing people familiar with the discussions. The sites under consideration are ones that already have power, in the United States, the United Kingdom and the Nordics, and the labs are pursuing them alongside the enormous campuses they have already committed to.
The reason is arithmetic about time rather than money. A gigawatt campus needs a grid connection, and a grid connection needs a queue position, a substation and years. A 25-megawatt building that is already energised needs a lease. Jabez Tan of Structure Research put the advantage in CNBC's report as "speed to usable capacity", which is the whole thesis in three words.
What makes it work technically is the shift in what the compute is for. Training one frontier model wants a single tightly coupled cluster, because the machines have to talk to each other constantly. Inference — running the model for users — is embarrassingly divisible, and can be spread across many modest sites without penalty.
The mix is moving that way fast. JLL, cited in the same report, expects inference to rise from about 9% of global AI workloads last year to 37% by 2030, and to overtake training as the dominant workload in 2027. Infrastructure bought for the training era is not the infrastructure the inference era needs.
Small is a relative term here. Rod Evans of Nscale told CNBC that a 20-megawatt cluster of some 10,000 GPUs represents roughly $2bn of capital. These are not side projects; they are the size of a serious enterprise data centre, bought as an increment rather than as a campus.
The gigawatt commitments have not gone anywhere. Anthropic has 5GW of capacity coming with Amazon, a further 5GW of Google TPU capacity through Broadcom, and a 460MW, $45bn arrangement with Nscale in West Virginia; OpenAI has 3GW in Georgia, 8GW in Ohio and the wider Stargate programme. The small sites are what fills the gap until those arrive.
This is the trade we described in August, when the binding constraint in the build-out turned out to be a place in the interconnection queue rather than money or chips. When capital is abundant and electricity is not, the premium moves to whoever already has the electricity.
It also explains a set of otherwise odd buyers. Owners of small, powered, unglamorous industrial sites — including the ones built for cryptocurrency mining — are holding an asset the two best-funded companies in the industry now want, and they can charge for it accordingly.
OpenAI is raising at more than $1.2tn partly to keep paying for this. The strategy the money buys is no longer only the biggest possible machine. It is also a hundred ordinary buildings that can be switched on this year.