The UK’s sovereign AI fund just handed taxpayer money to Jeff Bezos so that a Cambridge startup can build a “search engine” for rare metals that the chipmakers sitting on the coalition’s board already know how to find. CuspAI closed a $450 million Series B this month at a $2.6 billion valuation, with Bezos and the government’s AI fund among the investors. The coalition the company has assembled — Nvidia, Meta, Hyundai, and more than 40 other firms — is meant to use “frontier agentic AI” to discover new materials that don’t yet exist. What it will actually do is lock in a single chokepoint for materials-discovery software that every chipmaker, battery-maker, and energy-network operator will eventually have to route through, with the chipmakers who already own the supply chains writing the rules. This is enshittification by another name, four months before the launch, with the British government standing in for the user.
It is true that CuspAI’s co-founders — Chad Edwards and Max Welling — have a point in the narrow sense in which AI researchers usually do, when they say the next fifty years of industrial progress will be constrained by materials that don’t yet exist. The trouble is that the discovery of those materials is not the chokepoint. The chokepoint is the supply chains that have been deliberately throttled by the same handful of firms now sitting on CuspAI’s advisory board, and the AI software being built here will not loosen them. It will route around them, collecting rent on every transit, in the precise pattern that Cory Doctorow and Rebecca Giblin named chokepoint capitalism.
The arithmetic of materials discovery is not the constraint it appears to be. I have spent enough years reading crystallographic databases and enough evenings in my father’s workshop watching him find substitutes for parts that were no longer manufactured to know that the constraint on materials innovation is rarely the absence of candidate compositions. It is the absence of someone willing to pay the costs of characterization, qualification, and integration into an industrial process that already works. The semiconductor industry has known for at least two decades which families of high-dielectric-constant oxides, phase-change memory materials, and wide-bandgap semiconductors are worth pursuing. The cost of taking a candidate from a published crystal structure to a working process node runs into the hundreds of millions of dollars and the better part of a decade. AI software that proposes new compositions slightly faster does not lower that cost. It lowers the cost of the proposal step, which was never the binding step.
What AI materials-discovery software does lower is the cost of building a platform that sits between every chipmaker in the world and every candidate material, in a position to charge rent on both sides. The coalition CuspAI has launched — Nvidia, Meta, Hyundai, plus forty-five others — is not a research consortium. It is a customer-acquisition strategy. The “exclusive data access” the company boasts in its own press material is the asset, and that data was generated by the coalition members themselves, at their own expense, over decades. They are now contributing it to a jointly-owned discovery layer that will license its outputs back to them. The chipmakers fund the search, the chipmakers own the demand for the results, and a startup in Cambridge owns the search engine in between. When CuspAI files its next funding round, that ownership stake will be marked up on the strength of customer commitments that the customers themselves wrote into the coalition agreement. This is the standard playbook of platform extraction dressed in a lab coat.
The UK government’s role in this deserves its own paragraph, because it is the part that should not be allowed to pass unexamined. The “sovereign AI fund” investing in CuspAI — alongside the Bezos money, alongside the Kleiner Perkins money — is making equity investments of one to ten million pounds in British AI firms, on criteria that include a UK head office and British founders. The government did not disclose the size of this particular investment. What the disclosure-free structure means is that the British taxpayer is taking the same risk as the venture capitalists, on the same terms, with no special protections, in a company whose valuation already reflects the strategic value of the taxpayer’s commitment. If CuspAI succeeds, the venture investors and Bezos capture the upside through their preferred shares and information rights; if CuspAI fails, the taxpayer participates in the loss pro rata alongside them. The sovereign AI fund is, structurally, a co-investor with a public-relations advantage and a worse information position than its private-sector partners. That is not what the word “sovereign” usually implies.
The deeper problem is the political economy this investment entrenches. The UK has spent the last decade building an AI hardware strategy on the premise that sovereign capability in compute, chips, and foundational models requires public money to anchor private capital. That premise is defensible. The version of it now being executed — public money anchoring a materials-discovery platform whose outputs feed the supply chains of US-headquartered chipmakers, with Meta and Nvidia writing the coalition’s roadmap — is a different proposition. It is industrial policy as tributary. The British taxpayer is funding the upstream research that lowers the input costs of firms whose headquarters, shareholders, and ultimate loyalty sit elsewhere. The Cambridge lab gets the prestige; the returns flow to Menlo Park and Santa Clara.
The four-forces framework Doctorow uses to diagnose platform decay applies here before the platform has even shipped. Competition: CuspAI’s own coalition includes the firms that would otherwise compete to build materials-discovery tooling internally, so the competitive discipline that would normally force a discovery platform to share value with its customers has been pre-emptively negotiated away. Regulation: there is no UK or EU regulatory framework that treats materials-discovery software as the kind of chokepoint infrastructure that requires interoperability mandates or data-portability obligations, because the relevant regulators are still thinking about AI in terms of model capabilities rather than market structure. Self-help: any academic lab or downstream chipmaker that wanted to build a competing discovery layer would need access to the same coalition-curated datasets that CuspAI now controls, and those datasets sit behind contribution terms that effectively prevent exactly that use. Labor: the coalition’s advisory board includes John Giannandrea, formerly of Apple and Google, and Abhi Talwalkar, board member at AMD and chair of Lam Research — the labour that built the underlying materials science is now represented in the venture by people whose next career move is the IPO, not the foundry floor. The four forces are the conditions for extraction, not the outcome; if the AI software does not produce useful candidates, there is nothing to extract. But the coalition’s structure ensures that any success, however small, is captured.
The remedy is not complicated, and it is the remedy Doctorow has been prescribing for a decade — though the sovereign AI fund’s mandate as a commercial equity co-investor likely prohibits it from imposing public-interest conditions on its portfolio companies without the kind of statutory reform Parliament has not attempted. The UK government’s investment should have come with structural conditions: a commitment that any materials-discovered output funded by public money enters the public domain on reasonable, non-discriminatory terms; an interoperability requirement on the discovery software so that academic labs and downstream users can plug in alternative models; a portability obligation on the coalition’s training data so that the dataset built partly with public money cannot be locked behind a proprietary API. None of these are technically difficult. They are politically difficult, because they require the government to treat its own equity investment as a policy instrument rather than a marketing one. The government would rather be photographed next to Jeff Bezos at a Cambridge press conference than write the contract that would make the press conference matter.
It is worth naming what the coalition’s own framing concedes. Edwards and Welling say the world needs materials that don’t yet exist. That is true, and it has been true since at least the 1970s, when the first high-temperature superconductor patents were filed. What has changed is not the existence of the candidates but the willingness of the firms with the capital to qualify them to do so. The CuspAI coalition will not solve that. It will, at best, accelerate the proposal step for a small number of compositions the coalition members already knew to ask about, and license those compositions back to the same coalition members under terms the coalition itself negotiated. The UK government has paid to enter a club whose membership card is the only thing the club actually sells.
The Nscale data-centre financing MSI covered earlier this month was a different deal in form but the same structure underneath: public-adjacent capital underwriting infrastructure whose value accrues to the chipmakers and hyperscalers whose capex plans the infrastructure serves. The pattern is consistent enough that it deserves a name. It is the British state as junior partner in a US-headquartered supply chain, dressing up the partnership as sovereign capability. CuspAI’s “search engine for rare materials” will not change the geopolitical position of the UK’s chip industry. It will, however, give Jeff Bezos a tidy return and a press release. The Department for Science, Innovation and Technology got its photograph. The taxpayer got a minority equity stake in a company whose board includes Nvidia, Meta, and Hyundai, and whose most valuable asset is the exclusive access to data those firms contributed and could just as easily have contributed to a publicly-governed alternative. A parliamentary question on the sovereign AI fund’s disclosure rules and the terms of the CuspAI investment would be a start. The work is to be done. None of it was done here.