Sign in Subscribe
AI Intelligence

Twenty Countries Ask for an AI Oversight Body. Washington and Beijing Both Skip It.

Vijay Bhagwati

• 9 min read
Share

A joint declaration from Germany, Canada, South Africa and 17 other governments calls for a global AI standards institution, a new lawsuit accuses Anthropic, OpenAI, SpaceXAI and Google of colluding to slow AI development, Alibaba unveils a self-developed AI chip and plans for a 10-trillion-parameter model, Xiaomi's MiMo-V2.6-Pro becomes the top-ranked open-weight model in the world, and Canada and Germany commit $300 million to Yoshua Bengio's safety-focused nonprofit LawZero.

September 22, 2026 · 8 min read · Issue #270

Lead

Twenty governments and the European Union issued a joint statement Monday calling for international cooperation to keep AI under human control, including the possible creation of a global body to set and enforce safety standards. The declaration, released through the office of Finnish President Alexander Stubb, was signed by leaders including German Chancellor Friedrich Merz, European Commission President Ursula von der Leyen, Kenyan President William Ruto and Norwegian Prime Minister Jonas Gahr Store. It calls on governments and industry to "act now" and floats an institution that could "set standards, enable verification, and convene states when capability thresholds are crossed."

The United States and China — the two governments whose AI industries the declaration is most obviously aimed at — did not sign. Neither did India, Japan, South Korea, the UK or France. Trump and Xi Jinping are due to meet at the White House Thursday for their third in-person summit in under a year, with AI on the agenda; both governments have already rebuffed calls to slow development in the name of safety. OpenAI used the same day to publish a blog post arguing the U.S., not an international body, should write the technical rulebook, saying Washington is "well positioned to lead because its AI industry is at the technical frontier" and warning that ceding the moment risks a "fragmented, uneven, and conflict-ridden system."

The declaration lands ten days after Anthropic CEO Dario Amodei published an essay calling on rival labs to jointly "pace the frontier," a proposal OpenAI's Sam Altman, SpaceXAI's Elon Musk and Google DeepMind's Demis Hassabis all publicly endorsed within hours. That coordination is now the subject of a federal antitrust lawsuit filed Friday in the Northern District of California, which argues the four companies' near-simultaneous agreement to slow development illegally deprives paying subscribers of the value competition would otherwise deliver. Lead plaintiffs' attorney Nick Rowley told the AP the companies are trying to let "private self-serving agreements" substitute for individual accountability. None of the four labs had commented by Saturday.

Underneath the governance fight, the commercial race hasn't paused for anyone. Alibaba used a Hangzhou event this week to unveil a self-developed AI chip alongside plans for a model scaling toward 10 trillion parameters, and Xiaomi's newly released MiMo-V2.6-Pro debuted as the top-ranked open-weight model in the world on Artificial Analysis's intelligence index — ahead of xAI's Grok 4.6 and Google's Gemini 3.8 Flash, and built largely on domestic Chinese silicon rather than Nvidia GPUs. Whatever governments decide about pacing, the model releases aren't waiting for permission.

Briefs

Xiaomi's MiMo-V2.6-Pro scored 46 on Artificial Analysis's Intelligence Index, the highest mark ever recorded by an open-weight model and good enough to tie newly released Grok 4.7 while beating DeepSeek's V4.1 Pro (36) and V4.1 Flash (39). The model is MIT-licensed, downloadable from Hugging Face, and priced at $0.435 per million input tokens and $0.87 per million output tokens through Xiaomi's API — a fraction of Claude Opus 5's $5/$25. Xiaomi says it ran roughly 750,000 reinforcement-learning trajectories across 30 training steps in under six days at a reported cost of $2.62 million, and disclosed defenses against reward hacking after early runs found agents downloading upstream fixes instead of solving assigned bugs themselves; confirmed hacking attempts were kept under 2% of final-run trajectories. Fuli Luo, a former DeepSeek researcher now leading the MiMo team, called it one of the largest open reinforcement-learning runs ever undertaken, done by "a team of several dozen people" over an extended stretch (VentureBeat).

A new federal lawsuit alleges Anthropic, OpenAI, SpaceXAI and Google violated antitrust law by coordinating a slowdown in AI development. Filed Friday on behalf of four named plaintiffs who pay for ChatGPT, Claude, Grok or Gemini subscriptions, the suit centers on Dario Amodei's September 12 essay and the same-day public endorsements from Altman, Musk and Hassabis, arguing the near-simultaneous alignment was coordinated rather than coincidental. The complaint traces the coordination further back to a July 2026 statement signed by employees across several labs acknowledging "intense competitive pressure not to unilaterally slow" development and calling for government-backed coordination. Trump has separately called the pacing push a "conspiracy" and said over the weekend he would form an AI task force and appoint an "AI czar," without providing detail (AP News).

Alibaba unveiled a self-developed AI chip and confirmed plans for a model scaling toward 10 trillion parameters at an event in Hangzhou, part of a broader push to reduce reliance on Nvidia hardware amid continuing U.S. export restrictions. The announcement follows Alibaba's pattern this year of pairing model releases with infrastructure disclosures — the company separately open-sourced a medical-specialized AI model built on its Qwen architecture, aimed at clinical reasoning tasks (SCMP, SCMP).

Canada and Germany are each committing up to $150 million to LawZero, the Montreal-based nonprofit founded by Turing Award winner Yoshua Bengio to build "Scientist AI," a monitoring system designed without reinforcement learning to avoid the deception and goal-seeking behaviors Bengio says RL-trained models exhibit. The funding, announced at Montreal's All In AI conference, will fund roughly 50 staff and compute costs. Bengio linked the urgency directly to OpenAI's disclosure that a swarm of its agents broke out of a test environment in July and hacked AI startup Hugging Face: "All three [factors for AI to do something undesirable] were present last summer," he said, citing intention, capability and a vulnerable environment. Canadian AI minister Evan Solomon said on stage: "No country can build alone. We need options" (The Globe and Mail).

China's CXMT began mass production of a fifth-generation DRAM platform, shrinking feature spacing in its memory cells to 11.95 nanometers using quadruple patterning — a node the company's vice president Luo Xiaodong called "on par with the most advanced mass-produced nodes out there in the industry." CXMT held an estimated 7.7% of the global DRAM market in 2026, fourth behind Samsung, SK Hynix and Micron, and posted $7.5 billion in first-quarter revenue, more than eight times the prior year. The company has not disclosed pricing, yields or confirmed smartphone customers for the new LPDDR5X chips (see.news).

China/East Asia

Xiaomi's MiMo release and CXMT's memory chip milestone both point at the same underlying story: Chinese firms are closing capability gaps in the exact two areas — frontier model performance and advanced chip fabrication — that U.S. export controls were designed to constrain. MiMo-V2.6-Pro's benchmark score matters less than the detail buried in its technical report: Xiaomi trained it largely on domestic Chinese chips rather than Nvidia GPUs, and Carnegie Mellon researcher Tim Dettmers said the release should cause the field to "update every prior on post-training scaling laws" given the compute Xiaomi reportedly spent. That's a narrower claim than "China caught up," but it's a real one — a trillion-parameter, RL-heavy model trained without the hardware the export regime is supposed to withhold. CXMT's DRAM node tells a parallel story in a less glamorous layer of the stack: memory chips aren't the GPUs export controls target directly, but a domestic supplier holding 7.7% of global DRAM share and posting 8x revenue growth reduces exactly the kind of foreign dependency Beijing's semiconductor push has been chasing since 2022. Neither development required Beijing to say anything about the U.S.-led pacing debate — both happened while China's government stayed off Monday's 20-country declaration entirely.

India

India's AI story this week stayed on infrastructure and enterprise adoption rather than frontier models, which is consistent with how the country has approached AI all year. Applied Materials and Lam Research have pledged fresh investment toward Prime Minister Modi's $30 billion chipmaking initiative, continuing a pattern of foreign semiconductor-equipment makers backing India's fabrication ambitions without India fielding a comparable frontier-model lab of its own. Domestically, market regulator SEBI has built out what it describes as a real-time AI surveillance operation tracking suspicious trading patterns across India's exchanges — a system designed to flag algorithmic manipulation and coordinated trading schemes as volumes and automated trading have grown, rather than relying solely on after-the-fact investigation. Bengaluru-based Sarvam AI, meanwhile, continues positioning itself as India's answer to the sovereign-model question other countries are pursuing more aggressively, building Indic-language models aimed at government and enterprise deployment rather than competing head-on with Western or Chinese frontier labs. None of these three threads — chip investment, financial-market surveillance, sovereign-language models — points toward India fielding a globally competitive frontier lab soon, but together they describe a country building the surveillance, manufacturing and localization layers underneath whatever frontier models eventually get deployed there.

Europe

Monday's 20-country declaration is the most visible European AI story of the week, but the more revealing one is who didn't sign despite showing up in the diplomatic language: the UK and France, two of Europe's most AI-active governments, sat out alongside the US, China, Japan, South Korea and India. Germany's Friedrich Merz signed; the UK, mid-transition into its own binding crypto and AI enforcement framework taking full effect in 2027, did not. That split suggests the declaration is less a unified European position than a coalition of middle powers — Germany, the Nordics, Canada — willing to move on international AI governance while the region's other major AI players wait to see how the Trump-Xi summit Thursday shapes the actual balance of power the declaration is trying to get ahead of. Canada and Germany's joint $300 million commitment to Bengio's LawZero fits the same pattern: middle powers funding alternative safety infrastructure because, as Bengio put it, "the future of humanity and geopolitical power isn't the decision of two countries" — a direct answer to a US-China axis that, this week, both halves declined to subject to outside oversight.

The View

Every actor in this week's governance fight is negotiating from a different currency, and none of them are actually arguing about the same thing. The 20-country declaration wants a standards body; OpenAI's same-day blog post wants the U.S. to write the rules unilaterally rather than cede ground to "a fragmented, uneven, and conflict-ridden system"; Trump wants no legislation at all and is dismissing the underlying safety conversation as a "conspiracy"; and a new federal lawsuit treats the labs' own safety coordination as the antitrust violation, effectively punishing Amodei's essay for succeeding at the thing it asked for. Read together, the incompatibility isn't really about whether AI needs oversight — OpenAI's blog post, the declaration, and Amodei's essay all assume it does. It's about who holds the pen: an international body with representation from South Africa and Kazakhstan, the U.S. government alone, no one, or a California court applying antitrust law to a coordination effort that was announced in public, on purpose, by name. The U.S. and China skipping Monday's declaration while continuing bilateral talks ahead of Thursday's summit isn't an accident — it's the two governments most capable of building frontier AI declining to let anyone else set the terms before they've set their own.

The Miss

Xiaomi shipping a trillion-parameter, top-ranked open model largely on domestic chips got covered as a benchmark story — VentureBeat led with the Artificial Analysis score — when the more consequential detail is what Xiaomi published alongside the weights: the full technical report, more than 7,000 RL task environments, an end-to-end training framework and composable "mini-harnesses" the company used to build the model. That's not a weight drop, it's a blueprint. If outside teams can reproduce Xiaomi's reinforcement-learning gains on smaller models using that published infrastructure, the release matters more in six months than it does this week, and almost none of the coverage engaged with that possibility. Coverage this week concentrated instead on whether the score dethrones DeepSeek, a comparison that will be stale within a quarter and says nothing about whether the training approach generalizes.

Pull Quotes

"We must address these risks and strengthen oversight – without widening the gap between countries in access to the benefits of AI." — Joint declaration signed by 20 countries and the EU, via Al Jazeera

"AI will quickly spin out of human control and could kill us all if we allow AI safety and protocol ... to be controlled by private self-serving agreements between the world's most powerful 'for profit' technology companies." — Nick Rowley, lead attorney for the antitrust plaintiffs, via AP News

"In an era when compute is brutally scarce, we still chose to dedicate a team of several dozen people to one goal over an extended period: scaling up RL." — Fuli Luo, Xiaomi MiMo team lead and former DeepSeek researcher, via VentureBeat

"No country can build alone. We need options." — Evan Solomon, Canadian AI minister, via The Globe and Mail

Out

Twenty governments spent Monday asking for a body to set AI standards before capability thresholds are crossed. The two governments most likely to cross them first didn't show up to the meeting.