There is a company in Japan that makes the seasoning in your instant ramen and the insulating film inside almost every advanced AI chip on the planet. It is the same company.
Ajinomoto is best known for inventing monosodium glutamate, a little pot of which lives next to my cooker. (My father-in-law mistook it for sugar and put two teaspoons of it in his tea once: I’m not sure he’s forgiven me yet.) Rather less well known than Ajinomoto’s fine work in seasonings is that a piece of chemistry that came out of that same materials research, Ajinomoto Build-up Film (ABF), turns out to be the dielectric layer that makes modern flip-chip substrates work. Every high-performance CPU and GPU sits on one. Ajinomoto controls something north of 95% of the film. (Compare and contrast with their better-known product: these days, Ajinomoto only actually makes 15%-20% of the world’s MSG.)
When people talk about who controls AI infrastructure, they’ll usually be thinking about Nvidia, or TSMC, or a hyperscaler with a datacentre the size of Liechtenstein. Almost none of them will have a food company in Kawasaki in mind.
They should. Because that is where the value in sovereignty is: not in the stuff that looks great in photos, or the stuff that gets all the breathless press interviews where CEOs catastrophise about AI bringing about the end of the world, but in the least glamorous, most concentrated link in the chain. It’s in the boring, infrastructure stuff that we don’t think twice about because it doesn’t occur to us to do so.
On 8 June, at London Tech Week, the government announced a £1.1 billion AI Hardware Plan. Liz Kendall called AI the “defining currency of economic and hard power” and said the countries that control the hardware will hold the keys to the future. She is right about that. But the mechanisms that are being put in place, on reading the plan closely enough to notice what it buys and what it doesn’t, don’t work.
The money breaks down roughly like this: £750 million for a national supercomputer landing in 2030, of which £400 million goes on next-generation chips. £150 million of that will be spent this summer, acting as an early customer for inference silicon (before the market gets an opportunity to decide who wins, which is another issue). There’s £120 million for an AI Hardware Innovation Programme to fund British firms to design, test and tape out novel chips. There’s up to £150 million from the British Business Bank to anchor a new fund (the BBB’s largest-ever commitment). And a skills package with a £12 million doctoral training centre in chip design and a partnership with Arm on top.
It is a serious, well-intentioned plan, and I do not want to be sniffy about public money going into hard technology for once; it’s a good thing. But look at the shape of it. This plan funds the two ends of the chain: the big compute at one end, chip design at the other; plus the people to do the design work. It is a plan built almost entirely around the layer Britain is already good at, which happens to be the layer that is most visible to someone who’s not a technologist, but who pays attention to the business pages. This is not a good thing.
Now read the whole plan again and look for the words packaging, substrate, test, or materials. They are not there. And having founded and worked at a company that designs and builds its own silicon, this is something I have spent the last fifteen years thinking very, very hard about.
A chip is not a thing you design and then own: there are a number of steps in between. A chip is a thing that gets designed, then fabricated, then packaged, then tested, and every one of those steps has its own map of organisations who can get in your way and make things grind to a halt.
Britain genuinely leads in design: look at Arm, and the long tail of talent around it. But design is the most portable layer in the entire stack. Arm’s whole business is the portability: it licenses an abstraction that anyone can implement on whatever process node is going. That is a wonderful thing to own if you are running a licensing company. It is also very nearly worthless as sovereignty, because the thing you own is the thing that travels most easily. My buddy Alasdair made a version of this argument about the Rekursiv last month: if you put the value above the silicon, it outlives any particular chip. There’s an uncomfortable corollary. Put the value above the silicon and it belongs to whoever can license it, from anywhere.
There is no leading-edge fab in Britain, and there is not going to be one. Any British-designed inference chip is fabbed by TSMC, just like everyone else’s. And TSMC’s fabs depend, in turn, on a lithography chain that is a monopolist’s fever dream: ASML for 100% of EUV machines, Carl Zeiss for 100% of the optics, TRUMPF as the sole supplier of the drive laser. There is a Japanese near-monopoly on the photoresist and a duopoly on the mask blanks. Not one of those links is British, and most are single points of failure for the entire world, not just the UK.
Nobody funds packaging, and everybody needs it. Taiwan does roughly 85% of advanced packaging; TSMC’s CoWoS capacity, the specific process that stitches an AI accelerator to its memory, is sold out and functions as the gate on how many AI chips exist right now. The dependency runs deep: the United States has committed something like $265 billion to onshoring fabs, and yet the US still flies the wafers out of its shiny new TSMC Arizona plant back to Taiwan to be packaged, because the packaging line isn’t there yet and won’t be until the end of the decade.
We don’t do substrate in the UK. Underneath the package sits the film we started with. A handful of licensees – Unimicron in Taiwan, Ibiden and Shinko in Japan, Nan Ya, and one European organisation, the Austrian firm AT&S – turn Ajinomoto’s ABF into substrates, and the top three hold about three-quarters of the market. In the 2021 shortage, lead times for some parts blew out past a year and companies as large as Broadcom ended up stuck in a queue. Ajinomoto has just put prices up 30%, effective this quarter, with the shortage forecast to run to the end of 2027, because there is nobody to buy from instead.
So this is what the geography of the British inference chips that the new government plan is paying to design and validate looks like. It’ll be designed in one of our two centres for ASIC design, Cambridge or Bristol. Fabbed in Taiwan on a machine only the Dutch can build. Packaged in Taiwan on a line only Taiwan runs at volume. Mounted on a substrate made from a film a noodle-seasoning company in Japan controls almost entirely.
Oriole Networks can build a brilliant optical interconnect to move data between those chips, but it still has to cross the planet twice before it becomes a thing you can switch on. The UK can win the layer the plan funds, comprehensively (although the funding’s probably not sufficient to do so – a matter for another article), and not move the sovereignty needle at all, because the sovereignty never captured a crucial part of the supply chain to start with.
None of this is a secret to anyone who has built silicon hardware or run a production line. The problem is always the one line on the bill of materials with a single supplier, a 30-week lead time, and a phone number in a time zone eight hours away from yours. A Raspberry Pi has hundreds of discrete components, from tiny surface-mounted electronics to functional silicon. If a single one of those is unavailable, all production stops dead. Sovereignty, at the scale of a country, is the same production line, but larger. The question is never “can we design it?” It is “who can stop us building it, and have we bothered to find out who they are?”
And this is why the current fashion for sovereignty frameworks is a bit beside the point. The EU’s Tech Sovereignty Package in June came with a four-level “assurance” scheme for cloud services; the Lords debated digital sovereignty on 23 July; Britain, characteristically, has no explicit sovereignty strategy at all and has muddled through on collaboration instead. Every one of these frameworks measures the visible stuff: where the data lives, whose cloud runs it, whether or not there’s a Union Flag fluttering bravely at the top of the flagpole on the datacentre. The dependency that would actually turn the lights out is way down the chain, in a build-up film from Japan, and none of the sovereignty frameworks so much as mention it.
At Negroni we look at manufacturing, silicon and hardware companies, and when we do, this is a big piece of the due diligence we run. Between your tape-out and a working system in a customer’s hands, how many pairs of hands touch the thing, and how many of them are a single company, in a single country, that could decide tomorrow you are no longer worth the trouble? If you cannot answer that, you do not have a supply chain. You have crossed fingers.
Kendall is right that the keys to the future are real. It’s just that a good few of them are held by a company better known for making MSG.