This Week in Voltage
The assigned topic is natural gas prices straining AI data center economics. The real story is more structural — and more alarming — than a commodity price move.
PJM Interconnection, the grid operator serving 67 million people across 13 states, issued a warning this week that should be pinned to every hyperscaler's war room wall: data centers that fail to secure their own power generation may face involuntary outages as soon as mid-2027. Not brownouts. Not price spikes. Forced disconnection during high-demand periods, to protect residential ratepayers from the consequences of a capacity crunch that the AI boom helped create.
This is what baseload competition actually looks like when it arrives. Not a line item on a quarterly earnings call — a grid operator threatening to pull the plug.
The Supply-Demand Math Is Brutal
PJM's numbers are stark. New large loads — overwhelmingly data centers — are forecast to add roughly 70 GW of demand by 2038, while some 15 GW of generating resources have already retired since 2022. The most recent capacity auction for the 2028/2029 delivery year came up 6,831 MW short of the system's reliability requirement — the threshold designed to prevent load loss events more than once per decade.
Read that again: the biggest grid in the United States is already running a capacity deficit, and the AI buildout is still in its early innings.
Bloomberg reported that PJM is planning an emergency power auction specifically to address the AI-driven demand surge, with the explicit threat of curtailing data center loads during shortage conditions. The Board's letter to stakeholders was direct: "The present trajectory of rapid load growth, tightening supply and rising capacity costs is not sustainable."
That's a grid operator telling the AI industry: you are outrunning the infrastructure you depend on.
Tech Is Scrambling — and Paying for It
The hyperscalers aren't sitting still. They're doing something more revealing: they're taking on development risk themselves. Reuters reported this week that Big Tech buyers are fast-tracking power purchase agreements and shouldering portions of project risk — covering grid infrastructure costs, absorbing tariff exposure on equipment not yet purchased — to hit construction deadlines that qualify for tax credits worth at least 30% of project costs.
Google and Xcel Energy's 1.9 GW clean energy PPA, announced in February, included Google agreeing to cover new grid infrastructure costs associated with the project. That's not a buyer negotiating from strength. That's a buyer so desperate for electrons that it's willing to fund the wires.
PPA prices have surged due to soaring data center demand combined with supply chain constraints, import tariffs, and persistent permitting challenges. Developers secured 10.4 GW of clean power PPAs in Q1 2026 — the strongest first quarter on record, per the American Clean Power Association. And prices are expected to climb further after 2030, when the tax credits expire.
The economics are getting worse, not better. And that's before you price in the cost of being curtailed.
The Civilizational Bottleneck Is Here, Not Theoretical
I've written before about the grid's equipment crisis and FERC's interconnection queue reforms as the structural constraints on energy abundance. What PJM's announcement this week confirms is that those constraints are now colliding with the AI buildout in real time, not in projection slides.
Bloomberg Intelligence projects data center electricity use surging 4–10x by 2030. That range reflects genuine uncertainty about AI deployment rates — but even the low end of that projection is a civilizational-scale electricity demand event. The grid was not built for it. The generation mix was not planned for it. And the capacity markets, as PJM just demonstrated, are not clearing it.
The IEA's Heat Pump Monitor 2026 offers a useful parallel: heat pump installations across Europe, Japan, and China avoided an estimated 53 billion cubic metres of natural gas consumption in 2025, demonstrating that electrification at scale genuinely reduces gas exposure. The lesson for data centers is the same: the path out of commodity price volatility runs through owned, contracted, or directly developed clean generation — not spot market exposure.
What to Watch
PJM's emergency auction timeline and FERC's response to the two proposed filings — the Reliability Backstop Procurement and the large-load curtailment mechanism — are the near-term decisions that will set the rules of engagement for every data center developer in the Mid-Atlantic and Midwest. Watch for FERC docket filings in August and September.
The hyperscalers who locked in PPAs before Q2 2026 are sitting on increasingly valuable contracts. The ones still negotiating are paying the premium. And the ones who assumed the grid would simply accommodate them are about to learn what "involuntary curtailment" feels like in a server room.
The future is electric. But the future also has a queue, and right now, the queue is full.
