The U.S. and China are moving AI safety into power politics - Startup Fortune

AI safety is no longer just a domestic policy debate. Washington and Beijing are now treating frontier models as part of the same risk map as trade, chips and national security.

The most important AI story out of the Beijing summit was not a new model, a new chip or a new regulation. It was Treasury Secretary Scott Bessent saying the United States and China are ready to talk about guardrails for the most powerful AI systems, including a protocol meant to keep them away from non-state actors.

That may sound like diplomatic housekeeping. It is not. Once the two largest AI powers start discussing who should access frontier models, under what conditions and with what safety practices, the conversation has moved beyond voluntary lab commitments. It becomes a question of state power, export control and crisis management.

According to Reuters, Bessent said on May 14 that U.S. and Chinese delegations at the Beijing summit would discuss AI guardrails and set up a protocol for best practices to stop non-state actors from getting hold of the most powerful models. He also framed the talks as possible because the U.S. remains in the lead, while stressing that Washington does not want to stifle innovation.

That framing matters. The U.S. is not approaching China as a neutral partner in a shared research exercise. It is approaching China as a rival that still has to be engaged because advanced AI is too dangerous to leave entirely to market competition or military suspicion.

For years, AI safety has been discussed in the language of model evaluations, red teaming, system cards and voluntary commitments from companies such as OpenAI, Anthropic and Google. Those things still matter. But the Beijing talks show that governments are increasingly focused on a narrower and harder question: who gets access to frontier capability when the models can help with cyber operations, biological research, autonomous systems or large-scale influence campaigns.

The phrase non-state actors does a lot of work here. It can mean terrorist groups, criminal networks, ransomware operators, rogue procurement channels or any organization outside formal government control. Keeping those groups away from powerful models sounds like common ground because neither Washington nor Beijing wants advanced AI tools circulating freely among hostile actors.

But common ground can quickly become policy detail. If the protocol remains a loose statement of best practices, it may produce little more than diplomatic comfort. If it becomes operational, it could shape identity checks for model access, cloud usage rules, licensing thresholds, cross-border API limits and reporting duties for companies that train or serve high-capability systems.

This is where startups should pay attention. The firms most exposed are not only the frontier labs. AI infrastructure companies, model-hosting platforms, security vendors, agent developers and enterprise AI providers could all be pulled into a world where access controls are negotiated alongside trade commitments and national security demands.

Startups need to plan for policy risk

The summit itself ended on May 15 with both Donald Trump and Xi Jinping claiming progress in stabilizing the relationship, while the Associated Press reported that disputes remained over Iran, Taiwan and trade. That context is important because AI was not discussed in isolation. It sat inside a meeting already crowded with military risk, energy disruption, semiconductor tension and commercial bargaining.

For founders, that means AI compliance can no longer be treated as a legal cleanup item handled after product-market fit. If a company is building with frontier models, serving sensitive customers or relying on advanced chips, its policy exposure is part of the business model. Investors will ask whether the company can operate if access rules tighten. Customers will ask whether deployments can survive new restrictions. Partners will ask whether the company knows where its compute, data and model weights are going.

Chip controls are part of the same story. Washington has already used export restrictions to slow China’s access to advanced semiconductors, particularly those needed to train and run large AI systems. If model guardrails become more formal, the next phase may not be only about hardware. It may be about the model layer itself, including weights, APIs, fine-tuning, inference at scale and high-risk capability unlocks.

That does not mean every AI startup should behave like a defense contractor. It does mean the better companies will build basic governance early: clear customer screening, auditable access logs, abuse monitoring, incident response, model evaluation records and a serious view of where their systems could be misused. These are not glamorous features, but they may become the difference between selling into large institutions and being seen as a risk.

The danger is that a safety protocol becomes vague enough to satisfy diplomats but sharp enough to create uncertainty for builders. Startups hate uncertainty, and rightly so. Yet the direction is becoming clear. Frontier AI is joining the list of technologies that governments will not leave alone, especially when the U.S. and China both see it as a source of economic advantage and national leverage.

The next thing to watch is whether these talks produce a real mechanism or only a political signal. If it is only a signal, companies will keep guessing. If it becomes a mechanism, the AI market will start dividing more clearly between firms that can prove responsible access and firms that simply ship fast. In the next phase of AI, safety may not slow the market. It may decide who is allowed to stay in it.

Also read: Meta’s Louisiana AI campus puts a price on public incentives • STT GDC’s India IPO plan puts AI infrastructure in the market • ADTEK is testing Hong Kong’s appetite for AI infrastructure listings