Friday, September 25, 2026
The Argument
A declaration signed by more than 200 figures, including ten Nobel laureates, asked governments to agree international red lines for artificial intelligence by the end of 2026. The year has fifteen weeks left, no binding instrument exists, and the United States rejected the premise outright.
GreyNoise says a threat actor pointed hundreds of agents built on OpenAI's Codex and a DeepSeek model at two PaperCut flaws, reaching remote code execution on a real victim in under four hours and compromising eleven organisations in twenty-six seconds once the campaign ran. The exploit was ordinary. The labour was not.
Positron raised $875m on Thursday at a $5bn valuation, having been worth about $1.06bn in February. Its chips are built around memory capacity and bandwidth rather than raw compute, using commodity LPDDR5X rather than the high-bandwidth memory that constrains everybody else. That choice is the entire thesis.
Four companies have committed something close to a trillion dollars to AI infrastructure. Cash left immediately; the expense does not. It arrives in slices over the useful life of the equipment, and the length of that life is an estimate the companies make themselves.
Corporate AI budgets are shifting out of innovation labs and into core operating lines, a change that is quietly redrawing how companies account for technology itself.
Nvidia has agreed to acquire Hugging Face, the hub through which a very large share of open models, datasets and tooling reaches developers, in a deal reported around $13bn. The platform's value came substantially from belonging to none of the hardware vendors. That property is what has just been bought.
Twelve disclosed AI chip funding rounds raised $5.37bn between late January and mid-August. Eight of them, and about 67 percent of the capital, went to companies building for inference rather than training. Venture money is a forecast, and this one is about where the permanent cost of AI is going to sit.
Employment for 22- to 25-year-olds in AI-exposed occupations is down about 16 percent, and companies adopting AI have cut junior hiring by around 13 percent. Aggregate unemployment among AI-exposed workers shows no detectable rise since late 2022. All three findings are consistent, because only one of them measures a door that stopped opening.
More than $45m has gone into political advertising mentioning data centres since January — $22m behind Republicans, $21m behind Democrats — and both sides are making the same case about tax breaks and electricity bills. Seventy-seven percent of Americans told Reuters/Ipsos they are worried about what the buildings do to their power bill.
The National AI Research Resource has been made permanent, with an operations centre run by the San Diego and Texas supercomputing centres. The compute behind it still arrives from Nvidia, OpenAI and Microsoft, and no appropriation figure was announced.
A new projection puts data centre investment at $31.6 trillion through 2050. The same industry's accountants cannot agree on how long a server lasts, and the useful-life estimate that decides this year's reported earnings is set by the companies themselves.
The constraint on the AI buildout has been described as electricity for two years. The nearer one is a procurement queue for hardware whose price and availability are now set by trade policy rather than by fabrication.
The pitch for putting data centres in orbit opens with free power and ends there. Power was never the part that could not be solved, and the constraint that replaces it is worse.
Organisations are sending each request to whichever model suits it, which lowers cost and quietly turns model selection into a piece of infrastructure somebody has to own.
Organisations deployed systems that produce work product and left unresolved who is answerable for it. The question arrives with the first consequential error.
Employees adopted assistants on personal accounts long before any committee finished a review. Banning it moves the usage somewhere with no logging at all.
Deployments stall on the unglamorous discovery that the organisation cannot say what its own records mean. The model is fine. The inputs were never assembled for this.
Automated support handles more contacts than ever and the phones have not gone quiet. The gap between the two numbers is where the cost went.
Generated training data started as a substitute for records nobody could license. It is now chosen deliberately, for reasons that have less to do with scarcity than with control.
The work most exposed to automation is concentrated at the bottom of the org chart, which is also where the profession's training happened. Firms are removing the rung without deciding what replaces it.
Software that acts on a company's behalf has to log in somewhere. The identity systems it logs into were designed around a person who can be fired, and the gap between those two facts is now a live governance problem.