Britain trains AI stars, America gets the companies

Why Britain is losing its lead in the AI race

The UK trains elite AI talent and publishes strong research. Then, too often, that talent-and the companies it seeds-heads for the US. The reason is simple: capital, compute, and talent networks concentrate there.

The story is playing out in real time. After relocating his AI startup Gravitee to London five years ago, CEO Rory Blundell has been house-hunting in Denver. The company has grown to 130 staff and raised $60m this year, but pressure from investors and customers to scale in the US is mounting. London helped build the company; the next chapter may be written elsewhere.

The pattern: train here, scale there

On paper, Britain is strong. It ranks near the top for computer science education and hosts one of the largest concentrations of top AI researchers globally. Government funding for science is meaningful, and new money is flowing into AI-focused compute.

Yet the conversion rate from research to enduring companies is weak. As investor Nathan Benaich puts it: "Many AI startups start in the UK but end up scaling in the US because that's where hiring, capital, and customers converge." The data backs it up: the UK retains fewer than half of its best academic AI researchers. The US keeps roughly 80%.

Where the flywheel breaks

Three bottlenecks keep coming up: university spinout terms, growth capital, and infrastructure.

Spinouts still stuck at the starting line

The UK is excellent at generating papers but slow at turning them into companies. Founders cite aggressive university equity stakes and negotiations that drag on for months. That delay pushes teams to incorporate elsewhere or abandon the spinout entirely.

By comparison, Stanford is "miles ahead" of Oxford or Cambridge in companies emerging from AI labs, says Tom Hurd of Zeki Data. Standard, founder-friendly terms and speed win. The UK still treats each spinout like a bespoke legal project.

The capital gap is a relocation magnet

UK AI firms raised more than £2bn in 2024-more than France and Germany combined-but that's still a sliver of US private investment. Frontier-scale capital (for training large models, building evaluation frameworks, and specialized infrastructure) clusters in the US. Founders feel they must move to where the cheques, partners, and early enterprise customers live.

This is the single biggest relocation driver. Growth-stage investors want proximity. Boards push for it. And founders who delay often watch US competitors outrun them.

Salaries and retention

Compensation compounds the problem. Senior AI engineers in the US can earn up to $500,000. UK equivalents often top out near £150,000. Countries like the US, China, and parts of Europe are proactively headhunting with cash and signing bonuses. One in three founders cite access to top talent as their biggest barrier.

Result: PhDs, post-docs, and early-career engineers train here, then leave-often permanently.

Compute, electricity, and the missing clusters

Access to large-scale compute remains thin. Britain has high electricity prices and only a couple of Europe's most powerful supercomputers. That leaves startups and labs queuing for resources or tying up budgets with cloud bills.

Other countries are moving faster. France is deploying new capacity. The UK has announced an AI Research Resource, but researchers still report long wait times and limited access. For context on where Europe is building horsepower, see the EuroHPC supercomputer network.

What the state has tried (and what's missing)

Recent moves-£137m for "AI for science," a £100m "advance market commitment," and planned compute investments-are welcome. But the scale doesn't match US public and private firepower. Policy signals are improving; access on the ground still lags.

What to do now: practical actions

For policymakers

• Standardize spinout terms across universities with caps on university equity and clear IP rules. Target: term sheets in under four weeks.
• Expand the £100m advance market commitment to multi-year, multi-agency programs. Prioritize AI evaluation, safety, and high-stakes public use cases.
• Stand up neutral, frontier-scale training and evaluation clusters accessible to startups and independent labs, priced at cost and reserved via clear quotas.
• Offer time-limited incentives for clean, high-density compute: power purchase agreements, grid connections, and business rates relief tied to R&D access.
• Fast-track visas for senior AI engineers, applied researchers, and technical product leaders with same-week decisions.
• Modernize public procurement: pilot budgets, simplified security reviews, and outcome-based contracts that let startups sell into government quickly.
• Make founder equity treatment globally competitive. Improve clarity on EMI/ESO schemes and reduce friction at exit.

For universities and institutes

• Adopt a single national spinout template: low initial equity, clear IP licensing, anti-dilution protections for founders, and optional milestone-based royalties.
• Create "spinout studios" led by operators, not committees. Goal: company formed, seed round committed, and initial hires in 90 days.
• Pool compute allocations with open application windows for labs and spinouts, including public dashboards on usage and wait times.

For founders

• Consider a dual structure: UK HQ for research and IP, US entity for go-to-market and growth capital. Keep patents and core models in Britain where possible.
• Raise globally early. If your customer base or capital stack is US-heavy, plan for partner-led pilots there while maintaining UK engineering.
• Prioritize evaluation infrastructure. Build reproducible benchmarks, safety tests, and data contracts early-these shorten enterprise sales cycles.
• Use compute arbitrage: combine UK academic allocations, cloud credits, and pre-emptible instances. Share training runs across time zones.
• Compensate creatively: options with transparent value models, milestone bonuses, and research sabbaticals to retain senior talent.

For researchers and technical leaders

• Join labs with guaranteed compute, access to evaluation clusters, and active industry collaborations. Ask to see real quotas and schedules.
• Protect your optionality: negotiate clean IP arrangements before you publish code or datasets you may want to commercialize.
• Skill up where teams are thin: data engineering for training pipelines, synthetic data methods, RLHF/RLAIF ops, and model evaluation. Curated options are here: AI courses by skill.

What success looks like in 24 months

• Spinout cycle time: under 60 days from disclosure to signed terms across the top 10 UK universities.
• Talent retention: 60-70% of top academic AI researchers staying in the UK.
• Compute access: at least two frontier-scale clusters with published, researcher-friendly queues and SME allocations.
• Capital: three UK-based funds regularly leading £50m-£150m growth rounds in AI infrastructure and model companies.
• Public sector traction: 50+ live pilots using AI for science, health, energy, and safety, with a path to multi-year procurement.

The bottom line

Britain has the researchers, the institutions, and the cultural pull to build durable AI companies. What's missing is speed, scale, and founder-friendly plumbing. Fix the spinout terms, fund growth with intent, and open the compute gates. Do that, and founders like Blundell can build the next stage here-without a one-way ticket to Denver.