AI growth is colliding with physical and institutional limits. Google has signed one of the largest corporate power arrangements yet, AMD is planning a major supply increase, and governments are testing how to govern AI incidents and high-stakes deployments. At the same time, South Korea is preparing a sovereign frontier-model program, while investors continue to fund specialized chips and autonomous security tools.
Google signs a 3.6 GW power agreement
Google has contracted for 3,590 megawatts of electricity from Constellation Energy in the PJM grid. The arrangement includes a 20-year agreement for 890 megawatts from upgraded nuclear plants and another 2,700 megawatts under a long-term supply agreement. Constellation said the deal could support more than $4.3 billion in investment, with the first upgraded nuclear output expected in 2028.
The scale shows how AI infrastructure is changing corporate energy strategy. Data-center operators can no longer assume that grid capacity will appear on demand. Google's deal responds directly to PJM's proposed approach that large new customers bring their own power or accept curtailment during peaks. Power procurement, grid rules and project delivery dates are becoming core constraints on model capacity.
AMD prepares a larger 2027 chip supply ramp
AMD CEO Lisa Su said the company plans to substantially increase chip supply in 2027. During a visit to Taiwan, Su said AMD needs more advanced wafer capacity and is planning three to five years ahead. She met Foxconn and was due to meet TSMC, while the company is also working with memory suppliers in South Korea.
The announcement highlights that accelerator supply depends on an entire manufacturing chain, not only chip design. Advanced wafers, memory, packaging and systems assembly must all expand together. AMD's planned ramp may give cloud providers and model developers more choice, but the timing and volume of secured capacity will determine whether competition translates into readily available hardware.
OpenAI and Anthropic support clearer incident reporting
At an Australian parliamentary inquiry, OpenAI said it should have informed the government sooner about an AI-related Medicare statistics breach. Chief strategy officer Jason Kwon said the company had changed internal alerts and pledged faster disclosure. Anthropic backed mandatory reporting for serious AI safety incidents and said it was finalizing an agreement for independent testing by Australia’s AI Safety Institute.
South Korea plans a $3.5 billion frontier-model program
South Korea plans to separate its domestic AI strategy into a frontier-model track and an industrial deployment track. The proposed 2027 budget includes 4.7 trillion won, about $3.49 billion, in government equity investment for a top-tier homegrown model, with private capital expected to add funding. A new competition would be open beyond the teams in the current national model program.
The plan is a major sovereign-AI bet, but officials are setting a more targeted goal than simply matching the largest US labs. The government says a model competitive with leading Chinese open systems could be realistic if the funding and compute arrive. Parliament must still approve the budget, and the program design will determine whether startups can compete alongside established companies.
Utah authorizes an AI prescription pilot
Nolla Health says Utah has approved a pilot in which its AI system can issue initial prescriptions for a limited set of topical acne treatments. The system is restricted to preapproved treatment pathways and must hand cases outside those boundaries to a physician. Patients can contact a licensed doctor, and the state or supervising clinicians can stop the program.
The pilot matters because it moves AI from advice into a regulated clinical action. Its narrow scope is the safeguard and the test. Independent outcomes, escalation rates and adverse-event reporting will be more informative than the company's claim that clinicians agree with its recommendations in more than 96% of cases. Any expansion should depend on evidence from the controlled deployment.
Capital flows to AI security and inference chips
Agentic offensive-security company Hadrian raised $40 million, taking its total funding to $65 million. The company says its platform maps external attack surfaces and uses AI agents to validate which exposures are exploitable, while humans set objectives and make consequential decisions.
TechCrunch also reported that inference-chip startup Etched is reviewing early investment offers at valuations between $40 billion and $50 billion. Etched was valued at $21 billion in its previous round and declined to comment. The talks may change or produce no deal, but the bids show how strongly investors value specialized inference hardware and alternatives to Nvidia.
What matters today
AI competition is increasingly a contest over dependable infrastructure and enforceable controls. Google and AMD are securing power and manufacturing capacity years in advance. Australia, New York, Utah and South Korea are testing different forms of public oversight, from incident reporting to clinical boundaries and direct investment. The durable advantage will go to organizations that can prove their systems have enough energy and compute, clear accountability and measurable safeguards.