AI Trends: 25 September 2026

CEOs take the AI safety pitch to the UN

Sam Altman and Dario Amodei told the UN Security Council this week that AI development has outpaced government oversight and needs coordinated global rules. Amodei said AI "could be a risk to humanity as a whole" if mismanaged, while Altman argued that decisions of this scale "cannot be made by labs in San Francisco alone" and need democratic, accountable governments involved. The appearance lands the same week Google, OpenAI and Anthropic are separately forming SAFA, an industry-led standards body meant to work out third-party testing, incident reporting and auditor qualifications before regulators impose their own.

Washington starts treating model access as export control

The White House asked OpenAI and Anthropic to hold off giving UK testers access to new frontier models until a US government review clears them, a notable shift toward handling model releases the way chip exports are handled. It follows a separate US-China conversation this week about setting up an AI incident-notification channel, and comes as 26 state attorneys general separately urged Congress to slow frontier deployment while preserving existing liability protections. Bernie Sanders and Rep. Greg Casar also introduced a bill proposing a pause on frontier development pending new federal rules, though it has little near-term path to passage.

Anthropic's compute and product week

Akamai signed an $11.6 billion, seven-year deal to run Anthropic's CPU workloads, sending Akamai shares up sharply after hours. Anthropic also opened a Claude marketplace connecting plugins, connectors, agents and service partners against committed spend, and took Claude Code's cloud sessions public, letting the agent keep working on a cloud machine after a user logs out. Separately, Anthropic's life-sciences group gave more detail on last week's claimed lab discovery: the array-associated reverse transcriptase enzyme system was found by 950 agents running for 21 hours, and CRISPR pioneer Feng Zhang has since endorsed the finding, though independent replication is still pending.

DeepSeek crosses $1B revenue as funding rounds keep landing

DeepSeek said it has reached $1 billion in annualized revenue while closing a $7.5 billion funding round at a valuation of roughly 500 billion yuan. It's one of several large rounds this week alongside OpenEvidence ($250M at a $15B valuation) and enterprise browser startup Island ($400M at $6.4B), underscoring that AI investment hasn't cooled even as questions about a compute "reset wall" in 2027-28 circulate among analysts.

Also worth noting

Australia's government disclosed more detail on the OpenAI agent that accessed its Medicare Statistics Reporting Portal in June: OpenAI reportedly didn't notify Canberra until roughly three months after the incident, prompting an urgent government review. Google shipped Gemini 3.8 Flash with text-to-speech across 2,000+ voices and 100+ languages, and DeepMind chief Koray Kavukcuoglu said an early post-training build of Gemini 4 could ship well before year-end. OpenAI published MentalHealthBench, a 1,215-conversation benchmark built with more than 80 psychologists across 22 countries. And OpenAI's sponsored Instagram posts promoting ChatGPT more than doubled from June to August, timed with lower-priced GPT-6 tiers aimed at converting mobile users to paid plans.

Compiled from public reporting for the 48 hours ending 25 September 2026.