May was the month the capital race turned into a compute race, and the lines between competitors got blurry.
Anthropic rents Colossus 1
On May 6, Anthropic signed a deal to use all of the compute at Colossus 1, the Memphis data center owned by SpaceX, which also owns xAI. That is more than 300 megawatts and over 220,000 Nvidia GPUs. Anthropic said the capacity goes straight to higher usage limits for Claude Pro and Max subscribers. The two companies also said they would explore building compute in space.
Anthropic is paying a competitor's parent company for capacity, and that competitor is taking the money. Compute is scarce enough that neither side can afford to be principled about it. The binding constraint on Claude's growth this spring was users hitting limits, and Anthropic fixed it with whatever GPUs it could get.
$65 billion at $965 billion
Anthropic closed a $65 billion Series H on May 28 at a $965 billion post-money valuation, led by Altimeter, Dragoneer, Greenoaks, and Sequoia. Run-rate revenue had crossed $47 billion earlier in the month, up from roughly $9 billion at the end of 2025. That made Anthropic the most valuable private AI company, ahead of OpenAI's $852 billion from March.
Revenue growing five times in five months is the number to focus on. Anthropic's business has always leaned on API and enterprise usage, with coding agents as the biggest driver. It is the clearest evidence so far that paying for agents to do work scales differently than paying for seats.
Google I/O: built for the loop
At I/O on May 19, Google launched Gemini 3.5 Flash, generally available the same day. Google says it beats Gemini 3.1 Pro on coding and agentic benchmarks at lower cost and much higher speed. It also announced Gemini Spark, an always-on personal agent, and Antigravity 2.0, a desktop app for building and running agents with parallel subagents and scheduled background tasks. Gemini 3.5 Pro was promised for the following month.
Leading with Flash instead of Pro was deliberate. Agents make many calls per task, so speed and cost per call matter more than peak intelligence. Google built its biggest event of the year around the model that is cheapest to run in a loop.
A result mathematicians call a milestone
On May 20, OpenAI said an internal reasoning model had disproved the Erdős unit distance conjecture, an open problem in discrete geometry since 1946. The construction drew on algebraic number theory that researchers had not connected to the problem. Outside mathematicians verified the proof and wrote a companion paper. Tim Gowers called it "a milestone in AI mathematics."
Earlier AI math results mostly solved competition problems or cleaned up known ones. This one showed that a field had been wrong about something for 80 years.
Also in May
- Anthropic acquired Stainless (May 18), whose SDK generator was used by OpenAI and Google, and said it would wind down the hosted products. Owning the tooling around an API is now worth paying for.
- A jury dismissed Elon Musk's lawsuit against OpenAI (May 18) on statute of limitations grounds, without reaching the merits. That cleared the way for OpenAI's IPO filing.
- The EU reached a provisional agreement to simplify its AI rules (May 7).
What May tells us
Compute is the constraint, and nobody is picky about where it comes from. Revenue is following agents rather than seats. The gap between Flash-class and Pro-class models is closing fast enough that model choice is turning into a cost decision.