Energy Joins the Stack

NextEra Energy announced a $66.8 billion all-stock acquisition of Dominion Energy on May 18, framed by both companies as a bet on AI data center demand. The combined entity carries a 130-gigawatt construction backlog, a $59 billion annual capital budget, and around 10 million customer accounts across Florida, Virginia, and the Carolinas. Dominion owns the grid that powers northern Virginia's data center alley, and NextEra is already the country's largest renewable developer. The merger sits at $420 billion in combined enterprise value and lands the new entity third in US energy by size.

What looks like a utility merger is a layer of the AI stack quietly consolidating. Every frontier-AI buyer already carries a compute-provider dependency one level beneath the model itself: Claude Enterprise sits on a SpaceX-disclosed compute base, ChatGPT Enterprise runs on Azure, Gemini runs on Google's TPU stack. Compute providers buy power, and the power providers just consolidated to capture that demand. A third dependency now sits behind every frontier-AI contract: the energy provider feeding the compute provider that runs the model the buyer signed for.

The visible projects make the bet specific. NextEra is contracted to bring Google's Duane Arnold nuclear plant in Iowa back online for data center load. Meta has committed to a 2.5-gigawatt solar-and-battery deployment with NextEra across Texas and New Mexico. The deal materials reference 30 potential data center campuses across the combined service area, targeting 40 by year-end. Two of those campuses, in Texas and Pennsylvania, are 10-gigawatt hubs with US-Japan government co-ownership, structured under the Trump administration's reported $550 billion Japan trade package. That last detail is the one most AI coverage will miss: energy infrastructure for AI is now partly a geopolitical asset, not only a utility-tier supply decision.

The buyer-side implication has a clear shape. Through 2025, energy supply was an operations cost that procurement treated as a line item. Through 2026, energy supply becomes a strategic variable that AI buyers have to map. If a hyperscaler's frontier-AI workloads run on a grid serviced by a regional utility that just got absorbed into a 130-gigawatt construction queue, the buyer's effective compute supply chain now includes a 2027 merger close, a regulator's review of that close, and the construction timelines of whatever data center campuses come online behind it. None of that was on the AI procurement matrix a year ago.

The steelman is real. The deal does not close until 2027. Antitrust review for a $67 billion utility merger isn't trivial. Hyperscalers can and do build their own power infrastructure, and Meta's 2.5-gigawatt commitment is partly a hedge against exactly that dependency. There's a version of the next eighteen months where the utility tier consolidates and the AI buyer never feels it, because the hyperscalers shoulder the energy-sourcing problem themselves and pass the cost through as a flat per-token rate. That's possible.

But the bet itself is the signal. NextEra committed $67 billion in equity to capture AI-driven energy demand permanently. That pricing decision sets the cost floor for what AI energy will cost over the next decade, because no other US utility can underwrite the buildout at that scale and the merger removes one of the two parties most likely to compete on price. CEO John Ketchum told reporters the demand for electricity is rising "unlike anything we've seen in generations." That framing, from the company writing the check, is more load-bearing than any analyst note this week.

What to Do With This

If your firm runs anything material on a frontier-AI vendor, whether that's Claude Enterprise, ChatGPT Enterprise, Gemini, or a model API at production scale, add three lines to the vendor due-diligence template.

First, ask the AI vendor to identify the primary energy provider behind its primary compute provider, and to disclose any committed-capacity or long-term-supply agreements. Second, add a clause requiring 90 days' notice of any change to that energy provider's regulatory status, including ongoing merger reviews. Third, when comparing frontier vendors, include grid-region exposure as a scored dimension in the selection matrix. The vendors will say it's not relevant. It is.

Also on the Radar

SpaceX Scheduled Its IPO for June 12 With a Reported $75B Raise

SpaceX set June 12 as the IPO date, with a reported $75 billion raise, roughly three times the size of Saudi Aramco's 2019 listing. The Anthropic compute dependency disclosed in the SpaceX S-1 now has a public-market disclosure clock attached to it, and every buyer holding a Claude Enterprise contract gets the first earnings-call read on Anthropic's compute economics by way of SpaceX.

Anthropic's $30B Funding Round Is on Track to Close This Week

Bloomberg reported May 22 that Anthropic is on track to close a $30 billion-plus funding round at a $900 billion valuation as soon as the week of May 26. The pre-IPO cap-table close lands well ahead of the October IPO timeline, tightening the runway for any disclosure that would change the buyer-side compute calculus.

Anthropic Acquired Stainless, an SDK-Generation Startup

Anthropic acquired Stainless on May 18, an SDK-generation startup whose customers include OpenAI, Google, and Cloudflare. Buyer signal: Anthropic now owns developer-tooling infrastructure its frontier-AI competitors depend on, which complicates any "pure model vendor" framing for procurement teams comparing labs.