Equinix’s Nuclear Power Deal Ditches the Grid Queue

Equinix just did something no colocation provider had done before: it signed a non-binding agreement for up to 6 gigawatts of nuclear power that won’t exist for over a decade. Equinix’s nuclear power deal, unveiled around August 25, 2026, envisions that capacity coming online by 2040 — a date so far out that today’s toddlers will be filing taxes before the reactors switch on. That’s a strange detail for an “announcement,” and it’s also the whole story. This isn’t really an energy deal. It’s a market-structure deal, and once you see the mechanism underneath it, you can’t unsee it.

Here’s the claim nobody else is making about this: FERC’s June 2026 order, written explicitly to speed up how fast data centers connect to the grid, is the reason deals like Equinix’s now exist. Operators looked at the fix, decided it still wasn’t fast enough, and are responding by routing around the public grid entirely — pre-buying power from reactors that haven’t broken ground. They’re trading one multi-year bottleneck for a different one, on the theory that firm power you control beats grid power you’re queued for. I’ve spent enough time around both particle accelerators and hyperscale power rooms to recognize the pattern: when the wait becomes the risk, people stop waiting and start building their own supply chain.

Equinix’s Nuclear Power Deal: What Just Got Announced

The Equinix agreement is non-binding, which matters — it’s a statement of intent, not a signed offtake contract with penalty clauses. But the scale is the point: up to 6GW by 2040 would make Equinix, a company known for renting server rack space rather than building power plants, one of the largest private nuclear buyers in the country. That puts colocation providers — the neutral, multi-tenant data center operators that host everyone else’s servers — in the same conversation as the hyperscalers who started this trend.

It’s not an isolated move. In January 2026, Meta partnered with TerraPower on up to eight Natrium reactor plants, with the first two units — 690MW combined — targeted for delivery in 2032. Amazon, Microsoft, and Google have all made comparable bets over the past two years. What changes with Equinix is who’s writing the check. This is no longer just the companies that own the AI models buying their own power. It’s the landlords doing it too, which tells you the entire industry now assumes the grid alone won’t get there in time.

The AI Power Crunch, By the Numbers

The demand curve behind all of this is not subtle. The IEA projects global data center electricity consumption roughly doubles, from 485 TWh in 2025 to approximately 950 TWh by 2030. That’s not a rounding error in a national grid — it’s an entire new industrial sector materializing on top of infrastructure that was sized for a world without trillion-parameter models chewing through electricity 24/7. We covered how this demand shows up on ordinary bills in Phantom Data Center Power Demand Is Raising Your Bill, but the chart below is the blunter version: this is the number every utility planner, every grid operator, and now every colocation CFO is staring at.

Why FERC’s Interconnection Order Backfired

Now for the part that turns this into a genuinely counterintuitive story instead of another “AI needs power” retread. On June 18, 2026, FERC issued Section 206 orders to six major U.S. grid operators — PJM, MISO, SPP, CAISO, ISO-NE, and NYISO — giving them roughly 60 days, until mid-August, to justify or reform their interconnection and tariff rules for data centers and other large loads. FERC also imposed a separate 30-day mandatory reliability-report requirement, forcing operators to show how they’ll secure enough generation capacity to keep the lights on. Read at face value, this was regulators doing exactly what everyone had been asking for: shrinking the queue that had left large loads waiting years just to get a grid connection studied.

Except the practical effect looks close to the opposite of what was intended. Interconnection queues in PJM and other regions have stretched for years — long enough that even a faster, reformed process still means multi-year waits measured against a construction timeline that used to move much quicker. New reactors, even the “advanced” small modular designs getting all the headlines, still take on the order of six to ten-plus years to license and build. So operators are now comparing two slow processes and picking the one they can control. A faster public queue is still a queue you don’t own; a reactor you’ve pre-purchased is a supply chain you do. That’s not a small distinction to a company planning a $10 billion campus five years out.

This is the same logic that pushed some of the largest particle physics facilities I’ve worked around to negotiate their own dedicated power arrangements rather than sit inside a national grid’s general allocation — when your load is big enough and your timeline is fixed enough, you stop being a customer in line and start being an infrastructure developer yourself. Data centers have now crossed that threshold. We laid out the underlying grid-versus-compute tension in AI Data Center Power Bottleneck: Grid Beats GPUs, and this is that argument’s logical next step: if the grid bottleneck won’t yield, bypass the grid.

Behind-the-Meter Power: Equinix’s Nuclear Deal Isn’t Alone

The technical term for what Equinix, Meta, and others are doing is “behind-the-meter” generation — a private, dedicated power arrangement that sits on the customer’s side of the utility meter, rather than depending on power delivered through the public grid and its interconnection queue. It’s a deliberate end run around the exact process FERC’s June order was meant to fix. Instead of waiting for a grid connection study, the data center operator secures its own generator and skips the queue’s timeline almost entirely — trading a public bottleneck for a private, self-built one that at least moves at a pace they can negotiate.

Line these commitments up chronologically and the pattern gets hard to miss. FERC orders faster grid connections in mid-2026. Meta had already locked in TerraPower’s first Natrium units for 2032, months earlier. Equinix pushes the horizon out further still, to 2040, betting that six gigawatts of nuclear capacity is worth committing to now even without a binding contract. Each deal reaches further into the future than the last, which is the opposite of what you’d expect if the underlying power problem were actually getting easier to solve. We’ve tracked how this capital pileup is straining budgets and grids alike in The $1 Trillion Data Center Boom Is Breaking the Grid.

What Equinix’s Nuclear Power Deal Means for the Grid

None of this makes the public grid irrelevant — most data centers still need it for redundancy, and most of these nuclear agreements are non-binding statements of intent rather than steel-in-the-ground commitments. But the direction of travel is unmistakable: the more predictable and controllable path is now the one that runs through reactor licensing and construction, not utility interconnection studies. That’s a genuinely strange outcome for a regulatory order whose stated goal was to make the grid path faster.

Watch what happens as FERC’s reformed rules actually take effect over the next year. If interconnection timelines meaningfully shrink, some operators may quietly return to the public queue and let their nuclear options lapse unexercised. If they don’t, expect Equinix’s nuclear power deal to be the template rather than the outlier — more colocation providers, not just hyperscalers, pre-buying gigawatts from plants that exist only on paper. Either way, the real signal isn’t the reactor. It’s that the companies building AI infrastructure no longer trust the grid’s clock, and they’re willing to bet on a slower one they own instead.