Amazon breaks silence on AI safety — without backing a slowdown
Two things moved in the fight over who decides how fast AI ships: the loudest quiet company in Silicon Valley finally spoke, and a Chinese lab quietly became part of a national government's AI curriculum.
Amazon ended days of silence on the AI safety debate and landed squarely in the middle: models should ship only when they are safe and fully tested, but the company declined to join the call for an industry slowdown. "We don't see it as a choice between progress and safety," an Amazon spokesperson told Reuters, which said the company gave its position in response to an inquiry rather than volunteering it. "Models should be released when they're ready and safe to use, which comes from rigorous testing and strong safeguards." The company added a caveat that reads as a warning about coordination: "There are risks if that doesn't happen collectively. We believe that as an industry, in partnership with government, we'll work through the right protections."
Amazon's answer matters more than its wording, because of where it sits in the stack. It develops its own frontier models through AWS and a dedicated AGI unit, the home of the Nova family; it is a leading infrastructure supplier to the chip-hungry labs now arguing about pacing; and its Bedrock marketplace resells competitors' models to enterprise buyers. It is also Anthropic's largest backer — $13 billion invested as of April, with up to $20 billion more contingent on commercial targets, against Anthropic's commitment to spend more than $100 billion on AWS over the next decade. The pause Anthropic CEO Dario Amodei asked for in his weekend essay is the same request Altman put on paper for OpenAI, and the company with the most to lose from a slower Anthropic declined to endorse it.
That leaves the scoreboard lopsided in the way that matters. OpenAI, xAI, Google DeepMind and Microsoft have all backed a more measured approach in principle; Nvidia and Meta said this week that assessing and controlling AI risk is each company's own job. Trump has called safety worries a "hoax" and says the only guardrail AI needs is a strong, smart president — he is expected to meet AI executives this week or next, and it is not known whether Amazon is invited. A company whose model roadmap, cloud backlog and largest equity bet all run through the pacing question is exactly the one whose neutrality should be read as a position.
MiniMax is now one of the AI tools Singapore hands its citizens under a national skills program, and the market paid for it: shares closed up 7.06% at HK$254.80, a market capitalisation of HK$88.99 billion. The company was brought in to help build Singtel AI Pass, the carrier's subscription bundle, and three MiniMax products — MiniMax Agent, Hailuo AI and MiniMax Audio — ride along with it. Singtel AI Pass is one of the premium-tool subscriptions eligible Singapore citizens can claim when they enrol in the country's new AI courses; the other options on the government portal are ChatGPT Plus, Google AI Pro and Microsoft 365 Personal. Singapore's skills agency opened more than 200 AI courses on September 16, each carrying a six-month subscription to a premium AI service.
The shape of the deal is worth naming, because it is how Chinese labs are exporting now: not through app-store distribution but by being embedded in other companies' products and other governments' programs. MiniMax's open-weight M3 is the base under Saudi Arabia's HUMAIN M3, the Arabic flagship launched this month — Saudi Arabia's HUMAIN ships its first Arabic LLM — on top of China's MiniMax — and under Genspark's Gen-1 Slides, which Fireworks AI post-trained and which now runs that product's standard mode. Overseas revenue was roughly 60% of MiniMax's first-half total. Being listed beside OpenAI and Google in a national curriculum is a different category of win: it puts a Chinese model in front of citizens who never chose it.
What to watch: whether Amazon's non-answer hardens into a formal position once Trump's meeting happens, and which other national skills programs pull Chinese models into their subscription bundles.
Should a government's AI-skills program be choosing which companies its citizens learn on? Tell us in the comments.
