The big story this week is not a model, it is a bill. Anthropic agreed to pay xAI and SpaceX $1.25 billion every month for compute capacity through May 2029. That is a lab renting from a rocket company so it can keep training. Compute is now the gravity that bends every other decision in this industry.
You could see the same gravity at Meta this week. The company laid off about 8,000 people, roughly 10% of staff, with Zuckerberg telling the rest that “success isn’t a given” in the AI race. Capex is going up, headcount is going down, and the math is not subtle. When the bill from your compute partner is over a billion a month, something else has to give.
Google had a much louder week. At I/O, the company shipped Gemini 3.5 Flash to general availability and introduced Gemini Spark, a personal agent that runs 24/7 on dedicated cloud VMs and asks before doing anything serious. It is built on a new platform called Antigravity 2.0, which is essentially an IDE for steering agents. Google is finally betting on the agent layer instead of just the chat box. Honestly, I think this is the more interesting bet of the two.
On the policy side, Trump postponed signing his AI executive order, saying he “didn’t like certain aspects” of it. Whatever that means in practice, the U.S. now has a vacuum where federal AI rules were supposed to be. Nathan Lambert published a thoughtful piece on what comes next as labs run out of low-hanging scaling wins. He thinks the next phase looks less like one big release and more like a slow grind of post-training tricks and product polish. That matches what we are seeing.
Two smaller items worth flagging. Hugging Face released MinerU2.5, a 1.2 billion parameter document parser that sets a new state of the art for charts, tables, and key-value pairs. Useful for anyone touching PDFs or invoices. And Simon Willison wrote a sharp note about the curl project, which now receives more than one security report per day, four to five times the 2024 rate. The reports are well written and largely AI-assisted. Almost none of them are real vulnerabilities. The maintainers are exhausted. AI tooling does not just create new capabilities, it creates new tax on the people downstream.
What to watch: where the next billion-dollar compute deal lands, and whether anyone calls out that the cost of these labs is now starting to show up in the layoff notices.
T.
