AMD has signed the largest infrastructure commitment in its history, and it did not buy a single chip. Instead, the company locked up roughly 530 megawatts of American data-center capacity from Core Scientific under 15-year leases worth more than 14 billion dollars, with an option to expand to 2.5 gigawatts. The deal says something about where the AI race has moved. The binding constraint is no longer who designs the fastest silicon. It is who controls the power, the land and the grid connections to run it.
For most of AI's short history, the contest reduced to a single question of whose chips were faster. AMD's move signals the battlefield has widened. Sites with substations already wired, fiber already laid and power already contracted can be secured years before a rival can build equivalents, and they have become strategic assets in their own right. "Access to power, land and data center infrastructure has become critical to bringing new AI compute online," AMD's announcement said. A single large AI building can now draw 50 to 100 megawatts, enough to power a small town.
The counterparty is telling. Core Scientific was a bitcoin miner that filed for bankruptcy in 2022. Its mining sites, built for high-voltage power and heavy cooling, convert to AI use at roughly 3 to 4 million dollars per megawatt against 10 to 12 million for a greenfield build. Colocation now brings in 83 cents of every revenue dollar. The mining chips get stripped out; the electrical plant underneath transfers directly. Several former miners are making the same pivot, but none has assembled a contracted pipeline the size of Core Scientific's, now above 24 billion dollars across its AMD and CoreWeave deals.
For AMD, the infrastructure is a workaround for the thing it still lacks. Nvidia's dominance, roughly 80 to 88 percent of the accelerator market, rests less on hardware than on CUDA, a software ecosystem with 18 years and millions of developers behind it. AMD's ROCm has closed much of the gap on inference but still trails by 20 to 30 percent on training. Owning the buildings where its chips live is AMD's substitute for that missing software gravity: customers who want large AMD compute blocks in 2027 now have named sites to evaluate. To narrow the software gap directly, AMD is even paying Anthropic, whose Claude models will help optimise workloads for AMD hardware.
Nvidia, meanwhile, is answering with capital rather than concrete. The company is reportedly in talks to provide up to 250 billion dollars in financing toward a massive OpenAI data center in Ohio, and has been offering loan guarantees to operators buying its hardware. The logic is the same one driving AMD: in an era where training a frontier model can cost hundreds of billions, whether a customer can raise the money matters as much as whether it can get the chips. Nvidia is turning itself from a chip supplier into a financier of its own demand.
Underneath the headline figures sits a warning. Locking in 530 megawatts does not conjure 530 megawatts of paying customers, and analysts greeted Core Scientific's earnings with a downgrade over execution risk, noting it had gone 17 months without a new lease before this one. The infrastructure is being reserved years ahead of the demand meant to fill it. That is the wager the whole industry is now making at once: that the buildout will find its tenants, and that the money and the megawatts arrive on time.