SpaceX's IPO filing put a number on the compute deal it struck with Anthropic earlier in May: $1.25 billion per month through May 2029, covering capacity at both Colossus and Colossus II, the data-center clusters straddling Tennessee and Mississippi with more than a gigawatt of power.

Anthropic pays a reduced fee for May and June before the full rate kicks in, and either side can terminate with 90 days' notice. In the same filing SpaceX disclosed that its AI segment lost roughly $2.5 billion from operations on $818 million of segment revenue in the March quarter, and said it is in talks to sell spare capacity to other buyers. Reuters reported alongside the filing that Anthropic is nearing its first quarterly profit, with Q2 revenue expected to top $10 billion.

OpenAI announced that an internal general-purpose reasoning model — not one trained or scaffolded for mathematics — produced a proof disproving Erdős's 1946 conjecture that square-grid constructions are essentially optimal for the planar unit distance problem.

The proof gives an infinite family of point configurations with a polynomial improvement, and was checked by external mathematicians including Noga Alon, Thomas Bloom, Tim Gowers and Will Sawin; Sawin's follow-up makes the bound explicit at more than n^1.014 unit-distance pairs for arbitrarily large n. Gowers calls it a milestone in AI mathematics, and OpenAI describes it as the first time a prominent open problem central to a subfield has been solved autonomously by AI.

Alibaba released Qwen3.7-Max, its latest proprietary Max-tier model, positioned squarely at autonomous agent work rather than chat: sustained execution across hundreds or thousands of steps, frontend and multi-file coding, MCP integrations and multi-agent orchestration, plus an office-assistant mode that reworks documents through autonomous CLI tool calls.

It is closed-weight — the Max tier continues Alibaba's pattern of keeping the flagship proprietary while promising open weights for smaller variants.

At I/O 2026 Google introduced the Gemini 3.5 family and immediately made 3.5 Flash the default model in the Gemini app and AI Mode in Search globally, claiming roughly 4x the output tokens per second of other frontier models along with 83.6% on MCP Atlas, 84.2% on CharXiv Reasoning and 1656 Elo on GDPval-AA. The same model powers Gemini Spark, a 24/7 personal agent rolling out first to trusted testers.

On the harness side, Google confirmed it is folding Gemini CLI into Antigravity — the CLI stops working on 18 June 2026, with Antigravity CLI sharing the same server-side agent harness as Antigravity 2.0 and carrying over Agent Skills, Hooks, Subagents and Extensions as plugins.

After a three-week trial in the Northern District of California, a nine-member advisory jury in Oakland unanimously found that Elon Musk's claims against OpenAI, Sam Altman, Greg Brockman and Microsoft were barred by the statute of limitations, deliberating for under two hours.

Judge Yvonne Gonzalez Rogers accepted the verdict and dismissed the case on the spot, so the merits of the breach-of-charitable-trust claim — and Musk's damages theory, put by his expert at $78.8B to $135B — were never reached. Musk said he will appeal to the Ninth Circuit; his lead counsel told TechCrunch simply: One word: Appeal.