Friday, 2 October 2026
I Might. Maybe I Could Do That.
Asked Why Washington Doesn’t Own a Piece of OpenAI and Anthropic, the President Said He Might
TIME published the transcript of a White House interview on Thursday, conducted on September 28 by Eric Cortellessa and Sam Jacobs, and buried in it is the first time a sitting president has floated taking equity in the two labs by name. The question was a follow-up about the government’s Intel stake: if that was a good idea, why not do it with OpenAI and Anthropic? The answer, in full: “I might. Maybe I could do that. I have many deals like that—I do that.” On Intel he was blunter about the mechanism — “I didn’t buy it. I got it for nothing” — and put the government’s roughly 10% position, taken in an $8.9 billion deal in August 2025, at “$60 billion” by his own arithmetic. He also mentions taking a piece of Nvidia’s sales. Asked whether he would nationalise the frontier labs, he said no, twice, and described his instrument of choice instead: “The Department of Justice,” which he noted he has “used it twice, and it was very effective.” Nothing here is a deal. There are no talks confirmed — OpenAI reportedly offered Washington a 5% stake in July, and Anthropic and the administration have both denied discussing one. What there is, is a chronology: the interview happened the day after Trump had dinner with Dario Amodei, and four days after Amodei signed the White House accord committing his company to police its own models. The self-policing promise is in hand. Now the shares are being mentioned.
Source: TIME, October 1, 2026
The Four TPUs Are in Orbit, Running Fifteen Minutes at a Time
We flagged this one last week as a launch to watch, and it went: Google’s first Project Suncatcher satellite, refrigerator-sized, built with Planet, is up. On board are four of Google’s own TPUs, running a version of the open-weight Gemma model, and the headline constraint of the whole mission is thermal — the chips answer basic queries in fifteen-minute stretches and then stop to cool. That is the news, and it is better than a press release: the hardware survived ten g of launch and is computing in vacuum. The mission is scoped at about a year of orbital operation, with two more satellites next year to test whether they can talk to each other over laser links, which is the part the constellation idea actually depends on. What is new since we flagged it, beyond the fact of the thing flying, is how unromantic Google’s own people are being about the economics now that it is real. Travis Beals, who runs the project, gave the vision — “The sun puts out almost all of the power in our solar system…this project is about tapping into the best way to use solar power to run AI compute” — and then volunteered that he does not expect space-based data centres to be cheaper than terrestrial ones within five years. Deployment is not expected before 2028. Four chips, fifteen minutes, one year, and a senior director telling you in advance not to price it in.
Source: OPB / NPR, October 1, 2026
“Sovereignty Is Not Something You Rent”
Palantir named Armada its first Certified Modular Data Center Partner on Thursday, which is a dull sentence containing a specific bet. Armada builds Galleon: ruggedised, US-manufactured modular data centres you drop where power already exists instead of spending years building a substation. Palantir is putting its Sovereign AI Operating System — AIP, its Ontology product, Foundry and Apollo — inside them, and stringing multiple sites together into what the companies call a Sovereign AI Grid, so that, in their framing, critical workloads never depend on a single location. The pitch to governments and large enterprises is open-weight models on hardware you own outright, air-gapped if you want it, deployed “in months instead of years” per Armada chief executive Dan Wright. Alex Karp supplied the line the whole category has been looking for: “Sovereignty is not something you rent…owning the weights, the data, and the hardware those models run on.” Read it next to the first story in this post. Two different answers to the same question — who should own the compute — arriving on the same day, one from a president musing about equity and one from a vendor selling you the alternative.
Source: Yahoo Finance, October 1, 2026
Microsoft Shipped a Transcriber That Starts Writing Before You Finish the Word
A clean product item to close on. MAI-Transcribe-2-Streaming is Microsoft’s first real-time transcription model: first hypotheses in just over 100 milliseconds of receiving audio, a settled transcript 0.13 seconds after an utterance ends, 60 languages with continuous automatic language detection, and word error rates of 2.5% final and 2.8% on the first partial. Microsoft claims the top accuracy spot on Artificial Analysis for both final and partial transcripts and a place on that benchmark’s accuracy-versus-latency frontier, which is the claim that matters for this kind of model — anyone can be accurate slowly. It launched alongside MAI-Voice-2.1, a 23-language text-to-speech model, and a latency-optimised Flash variant, with introductory pricing at $0.54 per hour of audio to the end of the year, available through Microsoft Foundry, the MAI Playground, Vercel and OpenRouter. Nothing about this is a landmark. It is the eighth or ninth month running in which Microsoft has shipped its own audio model rather than buying one, and that cadence, not the millisecond figure, is the story.
Source: Microsoft AI, October 1, 2026