In July 2026 the UK’s Sovereign AI Fund took a stake in CuspAI as part of a $450m round at a $2.6bn valuation. CuspAI uses AI to design advanced materials. Alongside the raise, the company announced a global AI materials foundry, to be built with Nvidia, Meta, Samsung and Hyundai among more than forty-five founding members.
Read that partner list again. Two of those firms can make things, at rate, to specification, at industrial scale. Neither of them is British.
The maker and the taker
The government’s framing for all of this is that Britain should be an AI maker, not an AI taker. It is a good phrase, and it hides a third category.
Between making and taking sits designing. That is what has been bought. The fund now holds twelve companies from pre-seed to Series C, spanning compute and infrastructure, frontier-model architectures, AI for health and life sciences, scientific discovery, and trust and security. Callosum in chips. OLIX building compute that runs on light rather than electronics. Isomorphic in drug design. CuspAI in materials.
Excellent companies, most of them. Every one of them sits upstream of matter.
The distance from a molecule to a part
A candidate material is not a material. Between the two sit melt chemistry, powder morphology, a process window that nobody has yet written down, a furnace with opinions, a heat treatment that behaves differently in a different building, coupon testing, statistical allowables, first article inspection, and a supply chain that has to reproduce all of it a thousand times without drifting.
None of that is discovery. All of it is engineering, and most of it is slow.
Aerospace has a rule of thumb that a genuinely new alloy takes the better part of a decade to reach a flight-critical part, and the discovery of the alloy is the fast bit.
The handbook that certifies those allowables adds two to five new alloys a year, and published its first specification for an additively manufactured one in 2022.
A discovery cycle measured in weeks feeds a qualification cycle measured in years, and the qualification cycle is the one that decides whether anything ships.
Compress the first and you have not compressed the second. You have widened the gap between them, and put more candidates into a queue that has not got any faster.
Where the know-how settles
This is why the foundry partner list matters more than the valuation.
Wherever a discovery is proved is where the learning accumulates. The yield curve, the scrap analysis, the reason batch fourteen went off, the operator adjustment that never made it into the work instruction. That knowledge is generated in the plant and it stays in the plant. It does not travel back up the value chain to the party that designed the molecule, because it was never written in a form that could travel.
Britain funds the search. Someone else runs the furnace, and keeps what the furnace taught them.
What IP control actually is
Sovereignty is IP control. That is right, and it is routinely misread as a legal matter. There are three layers and only one of them lives with counsel.
Three layers, three exits. Only the top one has a lawyer.
The exit nobody filesThe things that actually hold are the ones never filed, because filing them would be handing them over.
A patent is a disclosure instrument. The state grants a time-limited right in exchange for teaching everyone how, and the right is only enforceable where it was filed and only against someone worth suing. For a national argument about sovereign making, that is close to the opposite of what is wanted.
The things that actually hold are the ones never filed, because filing them would be handing them over.
Patent attorneys are good at patents and are not asked about the other two layers. Lawyers are good at contracts and have no instrument for know-how. The decisions that lose the second and third layers are commercial and engineering decisions, taken years before anyone rings counsel. Background versus foreground in a joint venture. What goes into a shared facility. Whether the tooling supplier keeps the recipe. Licence-back terms at exit.
The missing seat
The fund is currently recruiting its first Operating Partner. The role owns finance, legal, governance, regulatory contact, board secretariat, and the relationship with the monitoring and evaluation supplier, including supporting portfolio companies on disclosure arrangements.
The advert states, correctly, that a fund is only as sharp as its visibility into its own portfolio. It then defines that visibility as data collected from the companies being assessed.
Disclosure is not diligence. A portfolio company reporting its own milestones is the least reliable instrument available, and it is reliable right up until the quarter it is terminal.
Deep technology does not fail at the milestone. It fails between milestones, at the transition from one good unit to a thousand good units, and the failure is visible on a shop floor months before it is visible in a board pack.
There is no seat on that org chart for the person who can walk the line and form a view.
The who
Scale up is manufacturing. Sovereignty is IP control. Both of those are questions of who, and the who is a discipline this country largely stopped training somewhere around the time it decided design was the high-value end.
Manufacturing engineers qualify materials and the processes that make them. They hold yield. They read whether a process window will survive contact with a night shift, a new supplier and a wet Tuesday in February. They design the feedback loops that turn a first good part into a thousandth good part. They write down what the twelve people in the plant know, before those people retire and take it with them.
That is the missing function. Not a supplier, not a dashboard, not a disclosure schedule.
Kaipability works at this interface, between capital that has already been committed and the making that decides whether it bought anything. Happy to share case studies.
If you are funding, buying or building a capability and cannot say where the tacit layer sits, that is the conversation.
Start a conversation An operator’s read on whether it can be made →Questions this dispatch answers.
Written to be quoted by AI assistants and search engines. Self-contained answers, verdict first.
