Britain's Sovereign AI Edge Is Scientific

Dear SoTA,

Britain’s sovereign AI debate is often reduced to scale: chips, data centres and whether the country can train a general-purpose model to rival the largest American and Chinese labs. All of these matter, but they are not the whole problem.

Britain does not need to recreate every layer of the American AI stack. In AI for science, sovereign capability means maintaining a critical mass of teams doing frontier research here, supported by scientific institutions, data and compute that are difficult to reproduce elsewhere. Drug discovery is one of the clearest opportunities.

Last year, as Technical Lead in the UK Government’s Sovereign AI Unit, I worked on where Britain could build genuine sovereign AI capability. I have since joined Boltz, which develops foundation models for biology and chemistry. Moving from policy into practice has shaped my view of why AI for science is one of Britain’s strongest opportunities, and what government should do next.

Silicon Valley, with its concentration of capital and compute, was always likely to become the centre of the LLM gold rush. For Britain, the good news is that AI for science is different. Specialised scientific models can be orders of magnitude cheaper to train, partly because high-quality, AI-ready scientific data is scarce. Models trained on narrower datasets do not require the same compute as general-purpose LLMs trained on much of the public internet and creative modeling ideas can often give even greater performance in the absence of scale.

AlphaFold 3 gives a sense of the scale. It was trained using 256 A100 GPUs for around 20 days. That is beyond the reach of almost every academic lab in Britain, but still orders of magnitude cheaper than GPT-4 and within the range of infrastructure the UK already has. Isambard-AI, the country’s largest publicly owned GPU cluster, contains roughly 5,000 GH200s. Britain may not be able to compete across the entire AI stack, but in AI for science it has enough compute to play to win.

Compute, however, is only part of Britain’s advantage. The UK has a long record of scientific breakthroughs and a research base that punches well above its weight, from the Laboratory of Molecular Biology to Diamond Light Source. In AI for science, ideas and scientific judgement matter alongside scale, turning that research base into a genuine competitive advantage. AlphaFold is the most visible result, but it is no longer an isolated one.

One of the stranger things about working in technology policy is how often something on the national wish list is already being built a few Tube stops away in Kings Cross. Having recently joined Boltz, I am obviously not a neutral observer. But Boltz is exactly the kind of success story for Britain that the policy debate too often overlooks: a founding team from MIT, able to raise capital and build almost anywhere, choosing London for the company’s next phase.

The company grew out of an unusually rapid run of scientific progress. When AlphaFold 3 launched in May 2024, it demonstrated the potential of biomolecular foundation models but remained closed. Six months later, a small group of graduate researchers at MIT released Boltz-1, the first commercially usable open-source model to approach AlphaFold 3 accuracy. Boltz-2 followed, moving beyond structure prediction to estimate how strongly molecules bind, while running more than 1,000 times faster than physics-based methods. BoltzGen then extended the work from analysing molecular interactions to designing new binders.

The uptake was immediate. Boltz models have now been downloaded more than five million times and are used by more than 100,000 scientists, including teams at every top-20 pharmaceutical company worldwide. Much of this was achieved by a tiny multidisciplinary team working across machine learning, biology and software engineering.

Boltz came out of stealth in January 2026 with a $28 million seed round led by a16z. The team is now around 30 people in King’s Cross and is training its next generation of biomolecular foundation models.

Boltz now works deeply with Pfizer, GSK, Takeda and a number of yet to be announced partnerships, deploying models across their research organisations and fine-tuning them on years of proprietary data. For a company this young, earning that level of trust within three of the world’s largest pharmaceutical companies is an unusual commercial result. That adoption now extends to AI agents: Boltz models are already the de facto tools they call for biological research, with first-party Boltz API integrations in OpenAI’s Codex and Anthropic’s Claude Science and demand reaching thousands of GPUs of inference at peak.

Boltz’s ambition is to enable scientists to go from a biological hypothesis to a therapeutic-ready molecule without leaving the computer. Almost all the science and engineering behind that ambition is now being done in Britain.

Boltz is not proof that policy has finished the job. It is proof that Britain already has something worth building around. The question is how to attract the next team and keep frontier companies rooted here as they scale.

The answer is to build scientific assets that individual companies cannot reproduce alone. Government has already made a strong start: its AI for Science Strategy directs up to £137 million to the field, while ARIA’s AI Scientist programme is preparing for a future in which parts of the scientific process are automated.

Two assets matter most: datasets no company can assemble alone and laboratories capable of testing model outputs quickly. I am admittedly biased, having been involved in setting it up, but OpenBind, built around Diamond Light Source, shows that Britain’s world-class research institutions remain incredibly relevant to AI-driven discovery when joined up in the right way. Autonomous laboratories could apply the same logic to experimental infrastructure, as DeepMind’s materials laboratory in London has begun to demonstrate.

These assets will only create an advantage if Britain maintains a critical mass of people able to use them. That means making AI fluency standard across scientific training and keeping frontier teams working across machine learning and science.

Britain’s sovereign AI edge is scientific. The test is whether frontier scientific teams choose to build here, hire here and keep doing their hardest research here.

Yours,

Charlie Harris

Boltz

Author biography:

Charlie is a Member of Technical Staff at Boltz. He previously served as a technical lead in the UK Government’s Sovereign AI Unit and did his PhD in AI for Science at the University of Cambridge.

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