Slow, Not Scarce

Dear SoTA,

Gallium is a soft silvery metal that melts in a warm hand. It goes into the power components that step the high voltage electricity arriving from the grid down to the low voltage actually used by a server, and into the optical components that move data between machines as light rather than as electricity. Roughly ninety-nine percent of the world’s newly refined supply comes from China, which put exports under licence in 2023 and stopped them to the United States entirely at the end of 2024. Britain refines none of it and imports every gram it uses.

Gallium is found in traces inside the bauxite that aluminium is made from and the ores that yield zinc, so it can be recovered only where those ores are refined. That makes the ninety-nine percent a fact about China’s refineries, not about where the bauxite is (if you are curious: Guinea, Australia and Brazil, mostly).

Overall, concentration tells you that you might be exposed, but not whether that should worry you. That answer hinges on how quickly you could get the same thing from somebody else, which is a question about how long new capacity takes to build rather than how much of the stuff exists.

Gallium is not rare: it is slow.

The UK AI Hardware Plan, published in June, puts £1.1 billion over four years into a national supercomputer, chip design, photonics, power electronics, and the training and salaries of the people who work on them. It states plainly that we will not try to reproduce the whole AI supply chain, which is the right call, and it names power electronics built on silicon carbide and gallium nitride as fundamental to efficient AI computing. It does not mention gallium once. Britain shut its last zinc smelter in 2003, and the aluminium plant at Fort William buys its alumina already refined, which is the stage where gallium would have come out. It comes out somewhere else now. The plan’s question, what Britain is good at, returns the industries that survived, not the ones whose absence now costs us.

Everything downstream of chip design happens abroad. This June I spent a Saturday tracing AI’s physical supply chain, from raw materials through to rented computing, and broke it into eight stages, though one could split it into six or twelve without changing the broad picture. The machines that print the finest circuitry are made by a single Dutch company. Taiwan produces almost all advanced AI chips and most of the advanced packaging. South Korea makes almost all the high bandwidth memory, the stacked chips that feed data to a processor fast enough to keep it working. Britain appears once, in chip design.

Dutch machines, Taiwanese fabrication, Korean memory, American compute: among these four near-monopolies, only some are critical to Britain, and the rule of thumb for which is which is how quickly you could buy the thing from somebody else. For example, renting computing power in somebody else’s data centre is a business around eighty percent American, but companies move their work between providers within months and do so routinely. Heavily concentrated, and quick to leave, which calls for a procurement rule rather than an industrial programme. Government and large buyers write portability into their contracts, insist the data comes out in a format somebody else can read, and know the exit terms before they sign.

Gallium is the opposite case: just as concentrated, but with nowhere else to go. No local aluminium industry means no gallium, and rebuilding one is a decade of work we have not started. Where a supply is concentrated in one country and slow to reproduce anywhere else, building our own fails on timing. A smelter commissioned in answer to a shortage will be complete years after that shortage has done its damage. The only levers left are holding more of it and wasting less of it, which for gallium means recovery. From the offcuts of semiconductor manufacturing, where specialist plants already recover between a quarter and a half of the waste, and in time from end-of-life equipment, where recovery is currently zero because taking a device apart by hand costs more than the gallium inside it. The first is a collection problem at a size this country can afford; the second is a research problem, and worth funding as one. It would not make us independent, but it would put a floor under how long we last after a stoppage upstream.

Building a domestic industry buys resilience only when something cannot be bought quickly and its inputs are available everywhere. Chip design is the clearest example, and where most of the plan’s money goes, alongside photonics and chip stacking. Those are the right choices, and this is the one place in the chain where asking what Britain is good at and asking what would hurt most if it stopped give the same answer.

The same logic applies to electricity, except that nothing here is concentrated in one country: the bottleneck is one we built. Britain was right to build generation, because wind is available everywhere and slow to install, which is the combination that rewards owning a thing outright. What we did not build were the wires. Carrying Scottish power to English demand needs lines that take a decade to consent and construct, so electricity the country already owns sits behind wires that cannot carry it. In 2025 we paid Scottish wind farms about £380 million to switch off and English gas plants about £1.08 billion to switch on, on Octopus Energy’s Wasted Wind figures, because the grid could not carry enough power south. That is roughly £1.5 billion spent replacing electricity that already existed, and the National Energy System Operator expects those constraint costs to reach £4 to £8 billion a year by 2030 if grid upgrades continue to lag.

The same bottleneck shows on the demand side. Britain’s electricity demand peaks at around 45 GW (imagine the whole country running everything at once on a winter evening). Requests to connect new demand to the grid went from 41 GW in November 2024 to 125 GW by June 2025, nearly three times that peak, with data centres accounting for 73 GW of it, across about 315 projects. Most of it will never be built.

If the wires cannot be built quickly, the other lever is where the demand goes, and data centres are unusually easy to site. A steelworks needs a port and a chemical plant needs a pipeline, but a data centre needs electricity and a fibre cable. Some workloads want to be near their users, but training is not one of them: a run does not care whether it happens three hundred miles further north. London and the Thames Valley hold something like three-quarters of British capacity, while ninety-eight percent of the power we throw away is wasted in Scotland. The pipeline is finally moving north, which is the first sign that anyone has noticed.

The AI Growth Zones are aimed at exactly this mismatch, steering new compute towards places with spare power. Ofgem’s July consultation is not, and it says so. Its proposed Data Centre Commitment Fee would require projects above 40 MW to post security worth £237,500 to £712,500 per megawatt from the moment they accept a connection offer until the day they energise, returned if they build and forfeited if they walk away, which makes it collateral against speculation rather than a price on the connection. The fee is sized as a share of what a data centre costs to build, not as an estimate of what the capacity is worth, and Ofgem’s stated reason for choosing capital cost is that it scales with the size of the project. Where anything should go is held over to a separate pillar, led by government, along with the auctions that would price capacity properly. So the security is identical whether a project sits beside curtailed Scottish wind or in west London. The consultation closes on 16 September.

Which of these Britain should do first - gallium recovery, the chip design and photonics the plan already funds, or the wires that would carry Scottish wind to English demand - I cannot tell you.

What I can tell you is how I would think about it. Ask the two questions separately: how much of it sits in one country, and how long it would take to get elsewhere. Concentrated and slow to replace, as gallium is, and you hold more and waste less. Concentrated but easy to leave, as rented computing is, and you need a contract rather than a factory. Slow to acquire but available everywhere, as chip design is, and you build. Building is the default answer, but it is the right answer only once in three, and in the other two we spend money for no gain in resilience.

The dataset behind these figures is open for anyone to use, check or correct, and several of its values are still marked unverified because I could not find a source I trusted. Some of you work inside these industries. If one of my numbers is wrong, I would far rather hear it than be right in private.

Stay wonderful ☀️

Yours,

Marco

Founder, Prismatic Labs

Share

Author biography:

Marco Zaccaria Di Fraia, PhD, is the founder of Prismatic Labs, which builds instruments for measuring, monitoring and managing the physical footprint of AI. The concentration map behind this letter, Culm, is open under Apache-2.0 and won the Breaking Barriers to AI Safety hackathon (LISA x BlueDot Impact, June 2026).

Write to the Society for Technological Advancement on letters@ilikethefuture.com.