UN AI Debate Changes Little: Fragmented EU and China Rules Favour the Biggest AI Labs
The Security Council session put national positions on one stage. It did not create a global rulebook, and providers still face different market-access tests in Brussels and Beijing.
At the UN Security Council’s September 23 session on artificial intelligence, Sam Altman and Dario Amodei called for international coordination while national governments continued to set the operating rules. That split is the commercial story. The UN can bring officials and companies into one room, but labs still have to satisfy the authorities that control access to major markets. Compliance capacity, not a single global AI regulator, is becoming a competitive advantage. The UN’s account records the session.
The European Union’s AI Act is now being enforced. Since August 2, 2026, the first enforcement wave has covered general-purpose AI models, prohibited practices, transparency and AI literacy; obligations for high-risk systems phase in later. The EU implementation timeline sets the dates.
For the largest models, Article 55 is the operative clause. Providers of general-purpose models designated as systemic risk must run evaluations and adversarial tests, and they must mitigate those risks, report serious incidents and maintain cybersecurity.
China works differently. Its 2023 Interim Measures regulate public-facing generative-AI services through content rules, with filing or security assessments in defined cases. A lab selling the same model in both markets must keep evidence for two sets of regulators and controls for two sets of outputs.
That divergence turns compliance into a fixed cost, and fixed costs favour scale. A model provider needs safety engineers, legal staff, red-team capacity, incident monitoring and records that survive regulator review. The cost lands at several stages: pre-release evaluation, customer documentation, regulator reporting and post-launch monitoring. Some evidence travels, such as a model inventory or technical description. Other work does not: adversarial testing of model capability cannot substitute for Chinese filing and content controls, and a local filter does not answer the EU’s systemic-risk questions. Large cloud companies spread those costs across many models, customers and markets, and already run security operations, legal teams, identity controls and regional infrastructure. A small lab has to assemble much of the same apparatus before it can reach a comparable customer base. It can serve fewer countries and forfeit those customers, ship separate releases, or lean on a cloud partner, which reduces the infrastructure work but still leaves it answerable for how its own model behaves. The scale advantage comes from reusing evidence and staff, not from escaping compliance work.
The burden also varies by system. Article 55 applies only to general-purpose models designated as systemic risk, not to every model offered in Europe. China’s measures cover public generative-AI services and require filing or security assessment in specified circumstances. A provider that keeps a lower-capability product outside a higher-risk category may face a different compliance bill from a frontier lab, and one company can operate both kinds of service. Regulators’ designation and scope decisions therefore shape who pays the largest fixed cost.
The rules also reach product design. A provider can keep one model and vary access, safeguards and features by location, or maintain distinct releases for jurisdictions with incompatible obligations. Either strategy adds testing and release-management work. Cloud platforms can supply hosting and some audit support, but the developer remains responsible for model behaviour, incident response and documentation. Customers may therefore see different capabilities or release dates depending on where they operate. For a multinational buyer, that means one vendor can supply a different product in Frankfurt and Shanghai, and procurement has to check which version each office actually receives.
Washington’s internal split makes a common standard less likely. At the UN, President Donald Trump said the United States “totally rejects any attempt to construct a globalist scheme of control” for AI. On September 3, Senator Bernie Sanders and Representative Greg Casar announced plans to introduce legislation to pause advanced AI development and bar superintelligent systems absent safety rules. The administration resists international controls while lawmakers in the same capital propose a pause. The White House transcript and Sanders–Casar announcement set out the positions.
The strongest counter-case is convergence. Companies may adopt the EU’s documentation and testing as a global baseline if European customers matter enough, and risk inventories, technical descriptions and some evaluation records can travel. Independent evaluators could make tests more credible and portable; Microsoft President Brad Smith has argued that safety judgments should not rest only with model developers. His September 23 interview makes the case.
Convergence has limits. Regulators must agree that the same test measures the same risk and accept the result across borders. The EU’s systemic-risk testing does not resolve China’s content controls, and neither settles political oversight or data access. Common paperwork may help; a common deployment decision is harder.
Even full convergence would not rescue the smaller labs. A regime built on a small number of approved evaluators turns testing into a queue, and a queue rewards the providers that can book capacity early, fund repeated evaluation rounds and absorb delay between model versions. Those are the same providers that already carry several compliance systems. If rules stay fragmented, the largest labs win on the fixed cost of serving many regimes. If rules converge around accredited testing, they win on the fixed cost of waiting in line. Either way, each additional market is easiest to enter for the labs that can afford the most compliance, and the UN session did nothing to change that.