Neoclouds show how to amplify risks in AI ecosystems

Even by the fantastical standards of today’s tech sector, neoclouds such as CoreWeave run a rum type of business. They rent out servers and chips on which AI computing takes place. Yet they in turn often rent their facilities, relying on lease guarantees from chipmakers and supplier credit. Their assets, then, are smaller than one might think.

Regardless, the scarcity of AI infrastructure means these fledgling companies are in big demand. Take Nscale, which last week signed up Anthropic in a six-year rental deal worth about $45bn. The London-headquartered start-up — hailed as “one of the UK’s leading AI companies”, although one might question its Britishness — is due to go public, possibly as early as next month.

Valued at $14.6bn post-money at its last fundraising, Nscale is certainly dressed to go. Board members include Meta Platforms faithfuls Sheryl Sandberg, formerly the Facebook owner’s chief operating officer, and one-time UK deputy prime minister Nick Clegg, who ran global affairs.

Peers’ performance is encouraging. Shares in CoreWeave of the US, which has a $100bn-plus order backlog, have more than doubled since it listed last year. Nebius of the Netherlands is up threefold in the past year. Given that demand for data centres is outstripping capacity, the grouping should be able to pull in revenue. The risk, of course, is that projections of demand for AI infrastructure turn out to be overly optimistic or that high prices usher other players into the market.

When that moment comes, the unravelling will be ugly. Neoclouds take circularity to new levels: chipmakers such as Nvidia are suppliers, investors and guarantors. The risk of vendor finance, as shown in the dotcom crash of 2000, is that because the money comes from those who have their own interest in capacity being built, there ends up being too much.

Another flashpoint is that neoclouds write long-term rental contracts to big AI companies, which can extend past 10 years. But there is limited visibility on how their own costs will evolve. By then, Nvidia’s once state-of-the-art chips may look as dated as an old-school Nokia phone, and about as useful, raising the possibility it will need to invest more. Neoclouds shrug this off, reckoning that a future generation of AI start-ups may be content with older processing capability.

But termination clauses suggest customers may not share that view. While contracts vary, the deal agreed in May between Anthropic and xAI — carrying a superlative annual price tag of almost $38bn per gigawatt — can be terminated with just 90 days’ notice by either party, according to SpaceX’s listing prospectus.

On the bright side, this is a sector that has earned its Darwinian stripes. In former guises, Nscale and Ionic Digital were harnessing power and compute to mine crypto. Should they need to pivot again, heaven forbid, that may yet stand them in good stead.

[email protected]