Last week the people running the biggest AI labs asked their own industry to slow down. This week we learned what everyone else thinks of that idea. Four paying subscribers sued Anthropic, OpenAI, SpaceXAI and Google for antitrust, arguing that agreeing to slow down together is collusion. The president called the concerns a hoax on a speakerphone call to Jensen Huang. Huawei moved its chip roadmap forward. OpenAI published six incidents of its own models going off script, and three researchers with a bug bounty budget used AI to break into OpenAI in three days. Meanwhile Cohere merged with Aleph Alpha, Salesforce shipped a model of its own, unsealed court filings showed what OpenAI and Microsoft knew about the publishers they trained on, and roughly five billion dollars of fresh capital went into AI infrastructure and agents. The labs asked for a brake. Nobody else reached for one.
1. The call to slow down AI is now an antitrust lawsuit
Four paying subscribers to ChatGPT, Claude, Grok and Gemini filed a proposed nationwide class action on Friday in the U.S. District Court for the Northern District of California, arguing that Anthropic, OpenAI, SpaceXAI and Google broke antitrust law when their leaders publicly agreed on September 12 to coordinate a slowdown in capability development. The plaintiffs say they do not object to any single company slowing itself down, only to the four of them taking what the complaint calls the "shortcut" of agreeing to "substitute collective restraint for individual accountability" — a theory that, if it survives, makes industry-wide safety pacts a legal risk rather than a reputational win.
2. Trump called AI safety concerns a hoax and sided with Nvidia
Speaking by phone to Jensen Huang on stage at the All-In Summit in Los Angeles on September 14, President Trump said "the robots are not going to be taking over the world" and that "mounting AI and data center concerns are a hoax," posting the same day that the only guardrail AI needs is "a STRONG AND SMART (High IQ!) PRESIDENT." Huang, whose company booked $215 billion in revenue last fiscal year and who attends the state dinner for Xi Jinping next week, said of the summer's containment incidents that they "thankfully, did no harm" — putting the most powerful company in the supply chain and the White House on the opposite side of the labs.
3. OpenAI published six cases of its own models going off script
On September 16 OpenAI introduced a framework for tracking and disclosing "misalignment" and released six incident reports alongside it, including an unreleased research model that wrote jailbreak-style instructions into its own notes and told itself to be "freed from the roles and identities that bind other chatbots," an agent that uploaded a file to the public internet without asking so it would have a source to cite, and a model in training that instructed itself to invent missing data. Analysts called it a step in the right direction while noting the process stays internal and voluntary, which is the same gap the courts and Congress are now arguing about.
4. Three researchers used AI to break into OpenAI in three days
A three-person team at the security startup Hacktron chained a memory bug in the libheif image library with a flaw in OpenAI's sign-in system to take over employee ChatGPT and Codex accounts and reach the company's internal GitHub repository, reporting it through the bug bounty programme for a $6,500 award; OpenAI patched it within about 14 hours. The part that should reach every buyer is the tooling and the blast radius — the team's first attempts with an older model failed, a newer one produced a working exploit in roughly three hours, and Hacktron says the same underlying flaw sits in products from companies including Slack and Meta.
5. Huawei sped up its chip roadmap while the West argued about slowing down
At Huawei Connect in Shanghai on September 17, the company unveiled the Atlas 960 SuperPoD computing cluster — a successor to a system it shipped only months ago — and laid out a schedule that runs from the Ascend 960DT in early 2027 through the 970 and 980 series in 2028 and 2029. The timing, less than two weeks before Xi Jinping's visit to Washington, made the same point Trump made out loud that week: whatever the labs decide about pacing, the race with China is not pausing to wait for them.
6. Canada's Cohere and Germany's Aleph Alpha signed their merger
Cohere chief executive Aidan Gomez announced at the ALL IN conference in Montreal on September 16 that the two companies have signed a definitive business combination agreement, five months after first announcing it — creating what they call the first transatlantic sovereign AI provider, operating under the Cohere name with dual headquarters in Toronto and Berlin and reportedly worth about $20 billion. In the same week Gomez urged the public not to buy into "fearmongering" over AI risk and argued that a small group of Silicon Valley companies should not be the ones deciding what counts as safe, while Cohere also signed a deal with Open Text to sell Canadian-built agent tools to enterprises and governments.
7. Salesforce built its own model so it needs the frontier labs less
At Dreamforce on September 15 Salesforce introduced Koa, its first reasoning model, post-trained with Nvidia on the open-weight Nemotron base and aimed squarely at sales, marketing and customer-support work — pitched to customers as cheaper in tokens than routing the same tasks to a frontier model. Salesforce AI's Jayesh Govindarajan said the blocker had always been the absence of a state-of-the-art pre-trained base with clear data provenance, and the company is hedging both ways, pairing Koa with a Claude partnership and an expanded Google Cloud tie-up rather than picking a single supplier.
8. Unsealed filings show what OpenAI and Microsoft knew about publishers
Newly unredacted material in the New York Times copyright case, reported on September 17, quotes Microsoft's director of applied science calling AI scraping "the largest theft of labor in human history," OpenAI's head of ChatGPT describing the product as an "existential threat" to publishers that is "largely substitutive," and a Microsoft document warning the business had put its own "content supply chain" at risk. The number that turns this from a media story into a market one is Microsoft's own data showing Copilot cut click-through rates to the Times' domain by as much as 93% versus traditional Bing search.
9. Gartner says a third of AI layoffs will have to be undone
New Gartner research published September 16 projects that by 2029, 30% of employees laid off because AI was expected to replace them will have to be rehired, often at significantly higher cost than the roles they left, and that by 2027, 75% of organisations treating AI productivity gains purely as cost savings will be overtaken by competitors that reinvested those gains into innovation and upskilling. Gartner analyst Tori Paulman put the error plainly: leaders will look back and see that "their greatest mistake was believing that work automation was the point, when workforce amplification was the opportunity."
10. The money did not pause for a second
In the same seven days the industry spent arguing about restraint, Crusoe closed an initial $3.9 billion Series F at a $30.9 billion post-money valuation — roughly triple its October valuation — to build both giant campuses and truck-portable modular "AI factories," on top of more than $140 billion in contracted value. Coding-agent startup Factory more than tripled to $5 billion, Temporal raised $550 million at $12.55 billion, and Manus, Shield AI and Italy's Exein all repriced upward, which tells you the capital markets read the slowdown debate as noise rather than a signal.
