When WIRED’s Will Knight left a major AI conference in Beijing last week, he came away with a single conclusion: the United States and China should set their fierce AI rivalry aside and cooperate on the risks that neither can contain alone. For UK businesses watching the two superpowers circle each other, the more useful question is not whether they will reconcile, but where Britain ends up standing if they do not.
The contest the UK did not enter but cannot avoid
The framing of frontier AI as an arms race has hardened over the past two years. Washington has treated Chinese AI advances as an economic and national security threat, imposing tight restrictions on chips and chipmaking equipment to slow Beijing’s progress. The pressure has reached individual companies: the US government recently ordered Anthropic to prevent foreign nationals from accessing its most powerful models, Mythos and Fable 5, over national security concerns. Anthropic’s response was to revoke access for everyone rather than build a screening apparatus.
Against that backdrop, the Beijing Academy of Artificial Intelligence conference reinforced an uncomfortable point. Both countries stand to lose if AI is developed too quickly and too recklessly. As models become more capable, more agentic, and more woven into everyday systems, the risk that they are used for cyberattacks or fail catastrophically grows in parallel. Because the US and China build the most advanced models, the argument runs, cooperation between them is the only thing that scales to the size of the problem.
Critical Context: The two facts sit awkwardly together. Export controls are designed to widen the capability gap; safety cooperation depends on closing the trust gap. The UK has to operate in the space between them, where neither logic fully wins.
For the UK, none of this is abstract. Britain does not build frontier models at the scale of either bloc, but it hosts the labs, the capital, the regulators, and the enterprises that depend on whatever the two superpowers ship. The country is a price-taker in the hardware contest and a rule-shaper in the safety conversation, and those two roles pull in different directions.
| Strategic dimension | US position | China position | UK exposure |
|---|---|---|---|
| Frontier model access | Restricting exports, gating powerful models | Building rapidly under constraint | Dependent on access decisions made elsewhere |
| Open-weight models | Rebooting its push (e.g. Nvidia’s Nemotron) | Market leader (Kimi, Qwen, GLM) | Heavy adopter of both, owns neither |
| Safety governance | Institute-led, security-framed | Engaged but cautious on disclosure | Early convening power via the AI safety summits |
| Compute supply | Controls the chokepoints | Subject to controls | Buys from the controlling side |
The open-weight question is the one that lands on UK desks
The geopolitics is vivid, but the decision that actually reaches a UK procurement team is narrower and more immediate: which models do we build on? Over the past few years Chinese companies have taken the lead in highly capable open-weight models, including Moonshot’s Kimi, Alibaba’s Qwen, and Z.ai’s GLM. These models are popular with US developers, and with British ones, precisely because they are open, capable, and cheap to run.
That popularity is now colliding with a security ceiling. Open-weight models have become central to research and innovation, but as they advance it becomes harder to ensure they cannot help an attacker find vulnerabilities or be turned into a cyber weapon. The latest from Z.ai, GLM 5.2, reportedly carries frontier agentic and coding capabilities. This week a leading Chinese cybersecurity firm, 360 Security Technologies, said it had built an AI model with hacking capabilities on par with Mythos. The next generation of open-weight models may match the strongest closed ones.
Reality Check: An open-weight model is not a managed service. Once the weights are downloaded, there is no vendor to revoke access, patch a jailbreak, or push a safety update. Whatever guardrails ship with it are the guardrails you keep.
Lin Yun, a professor at Shanghai Jiao Tong University who works on AI and computer security, expects attackers to gain an edge in the near term before new defences — many of them AI-driven — tip the balance back. He argues the industry will need fresh ways to guarantee that open models are current, free of backdoors and vulnerabilities, and have met safety standards. None of that infrastructure exists yet. A British firm adopting an open-weight model today is making a security bet on provenance it cannot fully verify.
There is also a signal worth reading. A source at one of China’s leading AI companies told Knight that security concerns are one reason some advanced Chinese models are no longer being released as open source at all. The open-weight abundance that UK teams have come to rely on may not last.
Britain’s real asset is the table, not the chip
The UK cannot out-compute either superpower, and it will not win the open-weight race. Its leverage sits elsewhere. The country convened the first international AI safety summit at Bletchley Park in 2023, where both the US and China signed a shared declaration on frontier risk — a rare moment of the two blocs putting their names to the same page. That convening role is the closest thing Britain has to a structural advantage in this contest.
Stephen Casper, an MIT computer scientist who spoke at the Beijing conference, pointed to research suggesting the benefits of international collaboration on AI dangers outweigh the national security risks of working together. He drew the obvious historical parallel: the US and the Soviet Union were forced to cooperate on nuclear safety even whilst racing to out-stockpile each other. “One thing that almost everyone in AI can agree on right now is that AI doesn’t need a Chernobyl moment,” Casper said.
Strategic Insight: If cooperation between Washington and Beijing happens, it will need neutral venues, shared technical standards, and trusted intermediaries. The UK has spent two years building exactly that capacity. Its value to British business is indirect but real — a stable rule-set is worth more to most UK firms than a marginal compute advantage they were never going to have.
The catch is that a convening role only holds whilst Britain is seen as genuinely independent. Lean too far towards Washington’s export-control posture and the UK becomes a US annexe with no standing to broker. Lean towards Beijing’s open ecosystem and it loses the trust of its closest security partner. The strategic position is a narrow one, and it is not guaranteed to last.
What UK organisations should actually do
The temptation is to treat all of this as a problem for governments and frontier labs. It is not. The decisions that determine a UK firm’s exposure are being made now, in model selection, vendor contracts, and security policy.
For organisations early in AI adoption:
- Treat model provenance as a procurement criterion, not an afterthought. Record where each model comes from, how it is licensed, and what happens to your systems if access is withdrawn.
- Default to managed, closed models for anything touching sensitive data or critical systems, where a vendor can patch and revoke. Reserve open-weight models for lower-stakes, sandboxed work until provenance guarantees mature.
For organisations already running AI in production:
- Map your dependency on any single bloc’s models. A stack built entirely on Chinese open-weight models or entirely on gated US models is a concentration risk, whichever side it sits on.
- Build the switching capability now. Abstract your model layer so that an export decision in Washington or a withdrawal of open releases in Beijing is a migration, not a crisis.
Implementation Note: Model portability is the single most useful hedge against this geopolitics. A team that can swap its underlying model in days, not months, is insulated from decisions made in capitals it has no influence over.
For organisations with mature AI security functions:
- Treat open-weight models as untrusted code. Apply the same scrutiny you would to any third-party dependency — provenance checks, red-teaming, monitoring for backdoors and unsafe behaviour.
- Engage with the UK’s emerging standards work. Firms that help shape technical safety standards now will find compliance cheaper later, and gain early sight of where the rules are heading.
The challenges that do not show up on the roadmap
Four risks in this picture are easy to miss until they bite.
The first is silent dependency. Many UK teams have adopted Chinese open-weight models without recording it, because the models were open, free, and good. The exposure is real but invisible until someone asks the question.
The second is the disappearing commons. The assumption that capable open models will keep arriving is already weakening, with some Chinese labs holding back advanced models on security grounds. Strategies built on perpetual open-weight abundance may be planning for a world that is closing.
The third is standards lock-in. Whoever sets the technical standards for verifying model safety and provenance shapes everyone’s compliance costs. If the UK is absent from that process, British firms inherit rules written for other markets.
Hidden Cost: The cheapest model today can carry the highest total cost if it is built on weights you cannot verify, from a supplier you cannot influence, under rules you did not help write. Acquisition price is the smallest line in that sum.
The fourth is the cooperation whipsaw. If Washington and Beijing do find common ground on safety, the rules could change quickly — shared standards, disclosure requirements, restrictions on certain open releases. Firms positioned only for continued rivalry could be caught flat-footed by a thaw just as easily as by an escalation.
The takeaway for UK business
The headline from Beijing is that the two AI superpowers may have more reason to cooperate than to keep racing. The lesson for British organisations is to stop waiting on the outcome of a contest they cannot influence and start managing the exposure they can.
Three things matter most. Know which bloc your AI stack depends on, and do not let that dependency concentrate by accident. Build the ability to switch models faster than geopolitics can move. And treat the UK’s convening role in AI safety as a genuine asset — a more stable rule-set serves most British firms better than any compute edge they were never going to hold.
The arms-race framing makes for dramatic headlines. The quieter truth is that the UK’s position between the two blocs is a workable one, provided British businesses make deliberate choices about where they build rather than inheriting their exposure by default. For more analysis of how AI policy lands on UK organisations, explore our insights and latest news, or get in touch with a tip or a question.
Analysis based on reporting by Will Knight for WIRED, “I Met With China’s Top AI Experts. They’re Freaking Out, Too” (24 June 2026). Strategic interpretation and UK context by Resultsense.