Eleven days before President Donald Trump meets President Xi Jinping at the White House, the U.S. government has explicitly declined to impose any mandatory pause or safety licensing on frontier AI development. The White House’s June 2026 executive order and national AI legislative framework instead leave pacing and safety decisions to the labs themselves — Anthropic, OpenAI, and xAI — despite their CEOs having publicly called for a coordinated slowdown.
China is using the vacuum. At the BRICS summit in New Delhi on September 12, 2026, Xi Jinping pledged a BRICS AI open-source community, positioning Beijing as a governance leader for the Global South. The two approaches now collide on September 24.
The U.S. government has outsourced AI safety to the companies racing to build the most powerful models. No federal rule obliges Anthropic, OpenAI, or xAI to slow down — even after their chief executives asked Washington to consider it. The institutional answer arrived on September 13, when former White House AI adviser David Sacks told them to stop asking permission. “So go ahead and pace the frontier,” he wrote. “You are the ones setting it.”
On the same Sunday, Anthropic CEO Dario Amodei told CBS News on September 13 that a coordinated speed limit on AI progress would be very difficult — and that he is unsure it is possible. Eleven days before Trump sits down with Xi, the only actors able to slow frontier development are the same companies competing to build it fastest. Beijing, meanwhile, is not waiting for anyone’s permission.
The policy architecture Washington has chosen
The U.S. approach is not an absence of policy. It is a deliberate design. The June 2026 executive order explicitly prohibits any mandatory licensing or pre-clearance requirement for AI model development, even as it directs national-security agencies to build classified benchmarking processes and voluntary pre-release access for frontier labs. The national AI legislative framework from March funnels congressional action into child safety, intellectual property, free speech, and national security — while explicitly avoiding any new AI regulator. Existing agencies absorb the load.
Speaker Mike Johnson provided the congressional bookend on September 13. He told CNN that Congress has already instituted AI guardrails rooted in the White House’s national AI legislative framework, that the safety burden rests with the developers creating the models, and that an emergency moratorium would risk losing the AI race to China. No new legislation is expected before the summit.
China, by contrast, has binding rules on generative AI: algorithm filing, security assessments, content controls. It is now extending that model through BRICS, offering open-source AI infrastructure and training to member states whose procurement and research ecosystems will increasingly align with Beijing’s security assumptions rather than Washington’s or Brussels’. The EU’s AI Act imposes risk-tiered obligations and strict rulebooks, but the Atlantic gap is widening: Washington emphasises voluntary standards and innovation, while Europe regulates and Beijing drives state-led norms through multilateral blocs.
The measurable gap and what it means
Stanford’s 2026 AI Index measured the American lead over the best Chinese model at 2.7 percent in March. Anthropic’s Claude Opus 4.6 led ByteDance’s Dola‑Seed‑2.0 Preview by a margin that, two years earlier, would have been 17.5 to 31.6 percentage points. The method is a benchmark leaderboard; the implication is sharper — the measurable technical advantage that once justified export controls and licensing leverage is nearly gone.
The honest caveat: the Stanford benchmark measures Chatbot Arena scores — a proxy for model quality in conversational settings, not real-world military or cyber capability. Douglas Madory of Kentik, who was not involved in the Index, has noted in unrelated work that traffic-volume measurements capture scale, not which specific capabilities were deployed. The 2.7 percent gap is the best available metric. It is not the whole picture.
While the U.S. debates architecture, China is building infrastructure. Xi Jinping’s address at the 18th BRICS Summit on September 12 was not a rhetorical gesture. It committed China to support joint large-language-model development, host specialised AI seminars and training for member states, and construct an open AI ecosystem explicitly framed as consensus-based global governance. The 11-nation bloc — led by China, India, and Russia — now has an alternative pipeline that Western export controls cannot reach.
The distillation campaigns and what they signal
U.S. law-enforcement and intelligence agencies alleged in an early September 2026 advisory that six Chinese AI firms — DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI — engaged in industrial-scale knowledge distillation against U.S. frontier models from Anthropic, OpenAI, Google, and xAI. The agencies described the behaviour as “aggressive, malicious and targeted” copying, stating that it likely occurred with Chinese government awareness. The advisory lands as a data point, not a policy trigger; no sanctions or export-control tightening has followed.
U.S. law-enforcement and intelligence agencies have also accused Chinese labs of copying American AI models through these distillation campaigns. The exposure is not hypothetical. U.S. agencies describe Chinese labs creating tens of thousands of accounts and routing millions of prompts through American chatbots — user conversations that may indirectly train foreign models if security controls fail.
Amodei’s framework and the China problem
Amodei’s framework for slowing down is specific. Anthropic has already unilaterally committed to the first step, granting independent evaluators employee-level access to systems. The second step would require democratic-country labs to coordinate safety standards. The third step would require governments to negotiate with authoritarian states. The third step is where the plan meets the Sunday news cycle.
In his CBS News interview and companion essay on September 13, Amodei identified China’s non-participation as the “toughest dilemma” and warned that military and competitive incentives to pull ahead are so strong that a speed limit may be impossible. He paused. “I think that’s going to be very difficult,” he said, “and honestly, I don’t know if it’s possible.”
| Entity | Current rule | New rule | Effective date |
|---|---|---|---|
| White House (National AI Legislative Framework) | No mandatory licensing | Seven pillars: child safety, IP, free speech, national security; no new AI regulator | March 20, 2026 |
| Executive Order 14409 | Voluntary classified benchmarking | No mandatory licensing or pre-clearance; voluntary pre-release access for labs to federal government | June 2, 2026 |
| U.S. Congress (per Speaker Johnson) | Existing guardrails on developer responsibility | No new emergency moratorium legislation before summit | September 13, 2026 |
| Sources: White House National Policy Framework for Artificial Intelligence (March 2026); Executive Order 14409 (June 2026); Speaker Johnson’s CNN appearance summary (September 13, 2026) | |||
The Nobel Prize signal
OpenAI CEO Sam Altman told Fortune on September 12 that Presidents Trump and Xi “would get the Nobel Peace Prize together” if they could agree on shared standards and testing for AI development, calling such a pact “a wonderful accomplishment” that could be outlined in a single-page document. The comment is not policy. It is a window into how thoroughly the industry’s safety conversation has shifted from what Washington should require to what it might bless if the two executives’ offices happen to agree.
The structure is now visible. Washington has built a voluntary framework and called it governance. Beijing has built a bloc and called it leadership. The September 24 meeting is the first moment those two architectures collide in the same room.
The voluntary architecture is the real architecture
The June 2026 executive order explicitly prohibits any mandatory licensing or pre-clearance requirement for AI model development, even as it directs national-security agencies to build classified benchmarking processes and voluntary pre-release access for frontier labs. The national AI legislative framework from March funnels congressional action into child safety, intellectual property, free speech, and national security — while explicitly avoiding any new AI regulator. Existing agencies absorb the load.
The BRICS open-source initiative Xi announced is not theoretical. It promises low-cost large language models and training to Global South governments and firms, pulling critical infrastructure and procurement into ecosystems whose governance standards are set in Beijing. For a Western energy grid, financial network, or defence supplier plugged into BRICS economies, that translates into a dependency on AI stacks it neither controls nor fully trusts. The voluntary U.S. framework offers no comparable multilateral counteroffer.
The summit’s joint statement, expected on September 24 evening Washington time, will signal whether AI governance lands in the cooperation or competition column. If AI earns explicit mention with safety or cyberattack monitoring language, the door opens for lab-level coordination. If it is confined to general innovation language — or omitted — expect export controls and BRICS open-source efforts to continue on divergent tracks. The Trump-Xi table is the arena. What it cannot produce, the companies racing each other will decide.
Beyond the headline
The power behind it
The real power shaping this debate is the institutional architecture Washington has chosen: a national AI framework and executive order that deliberately avoid mandatory licensing and shift responsibility to existing agencies and private standards. That design keeps frontier decision-making concentrated in a small cluster of firms whose commercial incentives and liability fears now effectively stand in for democratically negotiated safety rules, even as Beijing uses BRICS to embed state-centric governance and open-source ecosystems that could lock in its influence over Global South AI norms.
The timing
This confrontation over AI speed limits is unfolding just as the measurable technical gap between U.S. and Chinese frontier models has nearly evaporated and before any binding safety regime is in place. The September 24 summit therefore lands in a narrow window: Washington still has a small performance edge and chip-control leverage, while China is consolidating a BRICS governance bloc. That makes this month unusually consequential for whether AI remains an arena of loosely coordinated rivalry or begins to acquire even minimal bilateral guardrails around cyber and military use.
The reach
One underappreciated actor in this story is the BRICS AI open-source community Xi is pushing, which could quietly reshape Western cybersecurity and industrial planning. By offering Global South governments and firms low-cost large language models and training, the initiative can pull more critical infrastructure, procurement, and research into ecosystems whose security assumptions and governance standards are set in Beijing rather than Brussels or Washington. For Western energy grids, financial networks, or defence suppliers plugged into BRICS economies, that shift could translate into new dependencies on AI stacks they neither control nor fully trust.
Three decisions the September 24 summit forces
With the Trump-Xi meeting eleven days away and no mandatory U.S. safety rules in place, Western practitioners face choices the policy architecture leaves to them.
- US-based AI developer or executive
You must decide whether to voluntarily implement stricter safety protocols and potentially slow development, or accelerate to maintain a competitive edge against Chinese advancements and alleged IP theft. The government will not mandate a pause. Review Anthropic’s proposed slowdown framework and OpenAI’s commitment to grant independent evaluators employee-level access. Consult your general counsel on product-liability exposure — Sacks cited it as a commercial reason for caution on frontier releases.
- Western supply chain manager for advanced semiconductors
The U.S. advisory accusing six Chinese AI firms of industrial-scale distillation creates a clear trigger for tightened export controls. Monitor Commerce Department and Treasury announcements in the 1-2 months following the summit. Audit your current sales contracts for compliance clauses covering remote access to U.S. cloud compute and chip re-export restrictions. The Stanford Index shows the gap narrowing; enforcement escalation is probable.
- Western cybersecurity professional with APAC exposure
Evaluate and strengthen your organization’s API access controls, anomaly detection for model usage from foreign networks, and contractual guarantees with third-party AI vendors — especially those operating in or with ties to BRICS nations. The U.S. advisory’s “tens of thousands of accounts” detail means commodity account-creation defences are insufficient. The BRICS open-source AI zone creates a second exposure vector: models governed by non-Western security standards on logging and retention.
- Western investor with APAC emerging market exposure
Re-evaluate your portfolio’s exposure to technology companies in BRICS nations. The 2.7 percent performance gap undercuts Western export-control leverage and makes Chinese open-weight models viable alternatives for European and U.S. firms. Adjust investment strategies to account for the potential erosion of Western tech advantages in markets where BRICS open-source AI gains traction — India, Brazil, and South Africa are the most immediate arenas.
FAQ
Practical impact on U.S. AI export controls to China?
U.S. AI and semiconductor export controls already restrict sales of advanced Nvidia GPUs and certain manufacturing equipment to Chinese entities, with repeated tightening since 2023. Policy analysts expect further measures if industrial-scale distillation and model copying continue, potentially including sanctions on named labs or broader restrictions on remote access to U.S. cloud compute. Companies providing AI services to Chinese customers should monitor Commerce Department and Treasury announcements closely and review contracts for compliance clauses as rules evolve.
How a voluntary AI slowdown could operate in practice?
Amodei’s proposed slowdown framework envisions labs granting independent evaluators employee-level access to systems, coordinating safety standards among democratic-country companies, and seeking eventual commitments from authoritarian states. In practice, this would likely require an antitrust waiver or safe harbour to let competing firms share information and coordinate release pacing without violating competition law, plus clear criteria for when to slow or halt scaling. Without matching Chinese participation or government enforcement, any slowdown would remain a self-imposed constraint and could be limited to specific dangerous capabilities rather than blanket pauses.
What Western businesses should know about using Chinese or BRICS-aligned AI models?
Western firms increasingly encounter Chinese or BRICS-aligned models, especially open-weight systems that can be deployed on-premise. Key practical issues include data-localisation laws, intellectual-property protection, incident-reporting obligations, and alignment with domestic privacy rules like GDPR. Businesses should conduct vendor due diligence on where inference and training occurs, how logs are stored, and whether models are subject to Chinese security reviews or content controls, and ensure contracts address jurisdiction, data handling, and audit rights before integrating these systems into customer-facing products or critical workflows.
Explainer
- Frontier model
- An AI model at the cutting edge of capability, typically defined by performance on benchmarks measuring reasoning, coding, and general knowledge. The U.S. executive order defines a “covered frontier model” based on classified cybersecurity benchmarks. The term is contested: no universal threshold exists, and the designation itself is a policy choice.
- Knowledge distillation
- A technique in which a smaller “student” model is trained to replicate the outputs of a larger “teacher” model. It is widely used for legitimate model compression. The U.S. advisory accuses Chinese firms of using it at industrial scale against American models without authorisation, calling the behaviour “aggressive, malicious and targeted” copying.
- BRICS
- An 11-nation bloc originally formed by Brazil, Russia, India, China, and South Africa, expanded in 2024 to include Egypt, Ethiopia, Iran, Saudi Arabia, and the UAE. The 2026 summit in New Delhi became the venue for Xi Jinping’s AI open-source community proposal, positioning BRICS as a governance platform for Global South AI norms.
- Chatbot Arena
- A crowdsourced benchmarking platform where users rate responses from anonymous AI models against each other, producing Elo scores that rank performance. Stanford’s 2026 AI Index used it to measure the U.S.-China gap. The method reflects user preference, not lab-controlled evaluation — a distinction that matters when interpreting the 2.7 percent figure.
- Elo score
- A rating system originally developed for chess, now adapted to rank AI models by pairwise preference comparisons. A 39-point Elo difference on the Chatbot Arena leaderboard translates to the 2.7 percent performance gap Stanford reported. Small Elo deltas can reflect statistically significant but practically narrow capability differences.
RELATED STORIES
