The bottleneck in frontier AI is no longer chips. It is power.

The binding constraint in frontier AI is no longer silicon. It is electrons, and the next 18 months of announcements will be decided by who controls the power contract before they are about who controls the chip deal.

The mechanism is now visible. A data center announcement is structurally a power announcement: a vertically integrated, federally enabled compute-and-generation campus, sited on land a federal agency already controls, with generation authorized in one announcement and compute announced in a second. The Department of Energy's March 20, 2026 fact sheet on the SoftBank-SB Energy Ohio partnership names the load-bearing detail: 9.2 gigawatts of new natural gas generation on the Portsmouth Site, underwriting 10 gigawatts of data center demand. That is not a data center. It is a utility, built in one stroke and rebranded as a chip project.

The pattern repeats. The Nvidia-OpenAI 10-gigawatt partnership, the OpenAI-SB Energy Stargate work, and earlier hyperscaler buildouts share a shape: a public chip commitment in front, a separate power-and-site commitment behind. Wire coverage that leads with "$500 billion" or "world's largest" is reading the wrapper, not the spine.

The stakes settle cleanly. Whoever controls the power contract, the federal site, and the allied financing now decides who runs frontier training. The chip deal is downstream.