Google signs 3.6 GW power deal, a quarter of it new nuclear
Grid capacity, not chips, is becoming the binding constraint on the AI buildout — and on the same day, the labs told an Australian inquiry they can live with mandatory incident reporting.
Google has contracted 3.6 GW of power from Constellation Energy across the PJM grid — the largest electricity deal in the region's history, and the biggest single power commitment any AI company has made. The agreement covers 3,590 megawatts over 13 states, with 890 megawatts of new nuclear capacity coming from reactor uprates at 11 units across six sites in Illinois, Pennsylvania and New Jersey — roughly a quarter of the total, delivered under a 20-year contract while the remaining 2,700 megawatts run for 15 years. Constellation says it will invest more than $4.3 billion to make it happen, with the first uprate arriving in 2028 and the full 890 megawatts before the end of 2032. The deal is a direct answer to PJM's "bring your own power" proposal for new data-center load, and it arrives with a five-year technology alliance that puts Gemini Enterprise inside Constellation's operations. It matters because compute expansion has quietly become a permitting and grid-interconnection problem: after years of buying renewable credits and hedging on the wholesale market, Google is now paying to physically create generation. The pattern we flagged when utilities agreed to let Google and Nvidia dim data centers on demand — hyperscalers as grid actors, not just ratepayers — has its logical next step here: if you can't find the megawatts, buy the plants that make them.
OpenAI and Anthropic told an Australian parliamentary inquiry they would support mandatory disclosure of AI safety incidents — with OpenAI putting a concrete number on its own improved pace: a new breach disclosed in 48 hours, against three months for the last one. At the first public hearing of the federal Joint Select Committee on Artificial Intelligence in Sydney, OpenAI chief strategy officer Jason Kwon apologized over the agency's data-hacking incident, said disclosure would now begin before the facts are fully understood ("even if we don't fully understand the situation, we are just going to notify"), and backed a legally mandated framework: "The representatives of society need to make more decisions so we are not making all these decisions." Anthropic went further on the specifics, supporting the Office of AI's proposal requiring developers to report serious safety incidents because current commitments are "largely voluntary," and said it is finalizing a deal to let Australia's AI Safety Institute test its models independently. Google and Microsoft appeared before the committee the same day; the hearings run through October 9, with a report due November 30 that is expected to seed legislation. One caveat worth keeping: no source shows Kwon committing to a disclosure timeline — the 48-hour precedent and the mandatory-framework support are the on-record commitments. This is the first real post-breach test of whether frontier labs will trade voluntary self-governance for statutory duties, and on paper, right now, they say yes.
A team of researchers has stretched Yann LeCun's JEPA architecture into a single world model that works across seven domains — from fluid dynamics to clinical records — and used it to nominate a liver cancer drug combination that held up in lab samples and mice. JEPA-Anything, from PhAI Labs with collaborators at CUHK, Fudan, Bristol, Stanford, Oxford and Princeton, fixes what the authors see as standard JEPA's core weakness: one prediction head where easy patterns drown out hard ones. Instead it splits the predicted state into separate modules that each capture a different aspect, then reassembles them. The gains are strongest on physical dynamics — prediction error fell about 35 percent in intervention tests on a simplified Pong environment and by nearly half on the Burgers fluid-equations benchmark — while image tasks improved barely at all. The boldest result is the wet-lab one: the model's top-ranked candidate, pairing immune signal IL-18 with blockade of the tumor-shielding enzyme CD73, killed more cancer cells than either component alone across patient-derived organoids and mice. Keep the caveats attached: this is an unreviewed preprint, the model's Kepler's-third-law recovery came from a single training run the authors picked as best, and the cancer combination is a research candidate, not a therapy. Code and weights are already public.
What to watch: the Australian committee's November 30 report, PJM's regulatory response to hyperscalers buying their own generation, and independent reproductions of JEPA-Anything's wet-lab rankings once the weights are out.
Is a 20-year nuclear lock-up the smartest way for an AI company to solve its power problem — or the most expensive? Tell us in the comments.
