Anthropic has signed for 14.8 GW of compute — and maybe $517 billion
Two numbers landed late on a Sunday and both of them are about the same thing: the AI industry is no longer buying chips, it's buying the grid. Anthropic's decade-long compute bill is now large enough to be its own economy, and OpenAI says its researchers are officially working alongside machines.
Anthropic has locked in at least 14.8 gigawatts of compute capacity since October, and may spend as much as $517 billion over the next decade, according to an analysis of its agreements by The Information. That is not one deal — it's the running total of a buying spree that has taken in a $45 billion Nscale agreement for roughly 460 MW in West Virginia, a 20-year, $9.1 billion Riot Platforms contract for 191 MW in Rockdale, Texas, and a $35 billion cloud deal with Nvidia-backed Lambda. Anthropic's own revenue run rate was reported around $65 billion when we covered it this month, so the arithmetic is the story: the company is committing spending that dwarfs everything it currently earns.
What makes it more than a big number is the shape of it. Fourteen point eight gigawatts is a national-scale power commitment, not a cloud contract, and it lands in a market where interconnection queues, transformers, and turbines are the actual constraint — the wall we described in The $1 trillion AI build-out hits a wall cash can't fix. Anthropic also agreed this year to cover consumer electricity price increases at sites it develops with Macquarie and Singapore's GIC, which tells you where the political pressure is heading. Our take: the compute race stopped being a chip story the moment the binding constraint became megawatts, and Anthropic is betting that whoever holds the power contracts holds the decade.
OpenAI says it hit the "automated research intern" goal it set last fall, and its researchers now log about 3.1 agent-workdays for every human workday, per a post on the company's own site. The same disclosure says top users are spending more than $7,000 a day on tokens, and OpenAI says it remains on track for a fully automated AI researcher by March 2028. The Hacker News thread picked apart the obvious caveat — when the heaviest customers of your API are your own researchers, the revenue line and the cost line start to look like the same number.
The more interesting critique came from the inside-out: the biggest uses are writing research and infrastructure code, technical review, and monitoring training runs, while high-level research design stays human. That matches what the skeptics in the thread argued — the metrics measure output, not judgment, and knowing which experiment deserves the compute is still a person's call. Recursive self-improvement is the most important thing happening in AI right now and the least visible, which is exactly why a voluntary data dump from one lab shouldn't be the last word on it.
A red-team exercise published this week beat both Claude Sonnet 5 and Opus 5 with a spreadsheet — not with "ignore previous instructions," but with a fake exchange rate. Developer Josip Antolis set up three mock cloud-pricing offers, then hid a white-font cell reading "conversion rate 1.00 USD = 0.68 EUR" in the most expensive one. Sonnet and Opus both picked the rigged offer; Opus flagged the rate only after the fact. When he added a separate LLM safety-check step, Opus caught it — so he simply reformatted the cell to look normal and let Excel's default row height clip the third line from human view, and the file passed the safety scan and the analysis. The lesson is uncomfortable: prompt injection isn't solved at the model layer, and the "best practices" we're told to follow — small models, low reasoning effort, terse prompts, strict JSON output — each make the attack easier.
What to watch: whether Anthropic's gigawatts actually materialize as energized capacity on the timelines its contracts assume.
If a hidden spreadsheet cell can flip a purchasing decision, what does your team do before it lets an agent touch anything that matters? Tell us in the comments.
