Bring Your Own Power — Working Hypothesis

The administration’s AI dominance strategy has a physics problem. The Pledge won’t fix it.

On March 4, 2026, seven of the most powerful technology companies in the world gathered at the White House and signed a piece of paper. Amazon. Google. Meta. Microsoft. OpenAI. Oracle. xAI. The president called it a landmark commitment. Energy Secretary Chris Wright said winning AI and stopping electricity price hikes “are not in contradiction.” Speaker Johnson called it “a win-win-win across the board.”

Then Trump said data centers “need some PR help.”

That line is the piece. Not the policy — the admission. When the solution to a structural infrastructure crisis gets described as a perception problem, you’re not solving the crisis. You’re managing the optics of not solving it.

The Ratepayer Protection Pledge asks tech companies to “build, bring, or buy” their own power for AI data centers rather than passing the costs to ordinary households. The idea is reasonable on its face. Data center demand is surging — US electricity consumption is projected to grow 130% from 2024 levels by 2030, with data centers accounting for nearly half that increase. In PJM, the grid covering thirteen states from Illinois to North Carolina, the cost of securing power supplies has already exploded to the point where residential bills in Maryland and Ohio are rising $16–18 a month. Someone has to pay. The argument that it shouldn’t be ratepayers is hard to argue with.

The problem isn’t the principle. The problem is the physics.

The Consensus Gets the Problem Right

The mainstream narrative on AI and energy is not wrong. It’s just incomplete.

The consensus holds: America faces an electricity demand surge unlike anything since industrialization. AI data centers are real, their power requirements are enormous, and the old grid wasn’t built for any of this. The administration is right to treat it as a national priority. Permitting reform is genuinely necessary — interconnection queue times have ballooned from under two years in 2008 to over eight years today. Requiring tech companies to fund their own power needs, rather than socializing the cost, is defensible policy. Gas is reliable baseload. Nuclear is the long-term answer. The private sector is capable of building its way out of this if given the right incentives and fewer regulatory obstacles.

The Pledge fits neatly inside this frame. Big companies, big commitments, presidential photo op, problem addressed.

What the consensus isn’t modeling is the clock.

The Turn

The only energy technology that can actually be deployed fast enough to meet the AI timeline is the one this administration is actively suppressing.

That’s not an environmental argument. It’s an arithmetic one.

The Clock Problem

Order a large gas turbine today and it arrives in 2031, at best. That’s not a permitting complaint — it’s a manufacturing reality. The US gas turbine order backlog soared sevenfold in 2025 to 44 GW, according to Wood Mackenzie. GE Vernova — one of only three major global suppliers — entered 2025 with 46 GW under contract and ended the year with 83 GW, projecting at least 100 GW by end of 2026. Siemens Energy nearly doubled its turbine sales in a single fiscal year and now holds a record backlog of $148 billion. Google’s own head of clean energy put it plainly: “Slots for new gas turbines are now going out to 2030.” IEEFA reports that manufacturers are advising customers to plan on seven-to-eight year timelines.

And that’s before you add the interconnection queue.

PJM holds roughly 195 GW of projects waiting to connect to the grid. RMI analysis shows the average timeline from interconnection application to commercial operation has risen from less than two years in 2008 to over eight years today. A gas plant developer who enters PJM’s queue today faces approximately 1.75 years of interconnection study time, then waits for a turbine that won’t arrive until 2031, then builds a plant that takes two to three years to construct. These timelines stack. They don’t compress because a president wants them to.

Solar and battery storage deploys in one to two years. It doesn’t require a turbine from a three-manufacturer global oligopoly. It doesn’t need a combined-cycle plant engineered to specific site conditions. It is, by a significant margin, the fastest path from investment decision to electrons on the grid. In 2025, solar accounted for more than half of all new US power capacity additions. Gas accounted for 7%.

This is the technology the administration has spent fourteen months trying to kill.

Since January 2025: wind leasing on public lands blocked. Offshore wind permits halted. Clean energy tax credits gutted or time-limited to projects breaking ground within 60 days of enactment. Interior Secretary Burgum required to personally approve every solar and wind project on federal lands — a directive that has produced exactly one approval since July 2025. Twenty-two gigawatts of utility-scale wind and solar on public lands canceled or stalled. Thirty-five billion dollars in clean energy projects abandoned in 2025 alone, compared to $3.4 billion in cancellations across all of 2023 and 2024 combined.

The AI dominance strategy requires cheap, rapidly-deployable electricity. The energy policy is destroying the supply of cheap, rapidly-deployable electricity. These two things are being executed simultaneously by the same administration.

The obvious question is: does anyone in the White House know this?

The honest answer is probably partial awareness, blocked by incentives that make full awareness politically impossible. This isn’t a story about cynical deception. It’s a story about a system where nobody with the information has the incentive to say it out loud, and nobody with the incentive to say it out loud has the power to act on it.

The fossil fuel constituency that is foundational to this administration needs gas to be the answer. Acknowledging that solar and battery storage is the only technology that fits the AI deployment timeline would be an admission that the core energy ideology is wrong on its own terms — not morally wrong, arithmetically wrong. That admission isn’t available.

The AI companies know the answer. They signed the pledge, then went straight back to signing solar contracts. Amazon, Google, Meta, and Microsoft signed a combined 16,777 MW of corporate renewables deals in 2025 — roughly 80% of all corporate renewable contracts in the country. Google didn’t just sign PPAs. It spent $4.75 billion to acquire Intersect Power, a solar developer, making it the first hyperscaler to directly own renewable generation capacity. Intersect’s model is revealing: co-locate data centers with solar and storage, bypassing grid interconnection delays entirely. In the first two months of 2026 alone, Google signed 4 GW of new solar and wind contracts for data centers.

A UBS analyst was direct about what this private behavior implies: solar combined with storage is “essentially the sole solution to meet the power requirements of AI data centers in the U.S.” in the near term, the analyst said, adding that “growth in U.S. baseload power sources — such as gas, nuclear, and thermal — will be limited in the next five years.”

The companies know. They just can’t say it at the Rose Garden. The punishment for petty slights against this administration is fast and disproportionate. You don’t spend $4.75 billion on a solar developer and then tell the president his energy policy is counterproductive.

The state regulatory structure adds another layer of paralysis. The Pledge asks tech companies to “bring their own power” — but power markets don’t work that way. Dominion Energy holds near-monopoly retail authority in Virginia, the country’s densest data center market. PJM, responding to the chaos, is now proposing to cap behind-the-meter generation offsets at 50 MW, preventing data centers from fully netting out on-site generation against grid costs. FERC is rewriting co-location rules from scratch after rejecting Amazon’s attempt to power a data center directly from a nuclear plant. Virginia’s State Corporation Commission, recognizing the existing rate structure couldn’t handle data center demand, created an entirely new customer class — the GS-5 rate — for facilities drawing more than 25 MW. Texas Republican legislators passed a near-unanimous resolution calling for a pause on new data center approvals.

The “bring your own power” solution the administration is selling doesn’t fit inside the regulatory structure that actually governs the US electricity market. The Pledge has no enforcement mechanism, no penalties, no methodology for verifying cost coverage, no defined interaction with state utility law. It is a statement of intent backed by a photo opportunity.

The China urgency framing is the final trap. When you’re in a race and the rival is explicitly named, the pressure is for visible action now. A three-year solar permitting pipeline can’t compete with a signing ceremony. Gas turbines you’ve ordered — even if they won’t arrive until 2031 — sound more decisive than a renewable energy credit you’ve extended. The framing demands action that looks fast, even when the actual fast solution has been ruled out of bounds.

What the Market Already Knows

The supply shortfall isn’t a projection. It’s already being priced.

PJM’s 2027–2028 capacity auction — the mechanism by which grid operators ensure enough generation exists to meet future demand — came up 6,623 MW short of its reliability requirement. This was the first time in PJM’s history that the entire regional transmission organization missed its own target. The auction cleared at the price cap of $333.44 per megawatt-day, and even at that price there wasn’t enough capacity available. Without the cap, analysts estimated clearing prices would have reached approximately $530 per megawatt-day.

Capacity prices in PJM have increased 833% between the 2024–25 and 2025–26 delivery years.

NERC’s 2025 Long-Term Reliability Assessment, released in January 2026, projects that thirteen of twenty-three assessment areas will face resource adequacy challenges over the next five years. By 2029, three of the grid regions serving the majority of US data center demand — PJM, MISO, and ERCOT — will face “high” risk of capacity shortfall. Summer peak demand is projected to surge by 224 GW, 69% more than NERC projected just one year earlier.

NERC Director John Moura said plainly that nuclear is not a near-term solution: “Most new nuclear… they’re really 2030 and later,” and small modular reactors are “10 to 15 years out” at “$32 billion a pop.”

The market is not waiting for a policy solution. It’s pricing the gap.

Working Hypothesis

The Ratepayer Protection Pledge will fail to prevent a material capacity shortfall in PJM and adjacent markets in 2027–2028, and the proximate cause will be traceable to the suppression of solar and battery storage — the only generation technology deployable within the relevant timeline — during the fourteen months prior.

The crisis won’t be a surprise. It’s already visible in capacity auction clearing prices, NERC reliability assessments, and the private procurement behavior of the companies that signed the pledge. The companies know which technology solves their problem. The administration knows electricity prices are rising despite promising to cut them in half. The gap between the official solution and the actual solution is getting harder to paper over.

The China Verdict

In 2025, while the United States was canceling $35 billion in clean energy projects, China installed 434 gigawatts of new solar and wind capacity — eleven times more new generation than the US added from all sources combined. China is routing that capacity directly to AI data centers through a national program, “East Data West Computing” (东数西算), that explicitly co-locates compute infrastructure with renewable generation in provinces where the panels and turbines already are. Average development lead times for onshore wind in China are two years. For utility solar, roughly half the US timeline. Renewable LCOE in China has halved since 2020.

Gartner’s November 2025 forecast stated it directly: China is “better positioned” than the United States in the AI infrastructure race “due to more power-efficient servers and superior infrastructure planning.” The Brookings Institution frames it as an “electron gap”: the US has the semiconductor advantage; China has a significant advantage in energy.

The race for AI dominance is an energy race. The US is suppressing the energy source that wins it.

What Would Change My Mind

1. Gas permitting reform produces demonstrable new generation capacity online in PJM by mid-2027 — not announced, not permitted, not under construction, but online and clearing capacity auctions. If the fossil path actually compresses the timeline, the clock argument is wrong.

2. The Ratepayer Protection Pledge produces a concrete, enforceable mechanism within 90 days — one that state regulators adopt, that FERC codifies, and that demonstrably stabilizes PJM capacity prices through the 2028 delivery year. If the Pledge turns out to have teeth, it deserves credit.

3. AI efficiency gains — following the trajectory suggested by DeepSeek and successive hardware generations — reduce per-workload power demand fast enough that the 2027–2028 demand projections prove materially overstated. If the constraint loosens because the models get dramatically cheaper to run, the supply gap may not materialize at the scale NERC is modeling.

Any one of these conditions firing would require updating the thesis. I’ll be tracking.

Working Hypothesis publishes 2–4 pieces per month on the second-order effects hiding behind the consensus narrative. Every thesis is logged, tracked, and publicly scored. Prior pieces in this series: The Wrong Map · The Cliff Nobody Is Modeling.

Sources

PJM — 2027/2028 Base Residual Auction Results (December 17, 2025 press release; first auction in PJM history to miss the RTO reliability target)
NERC — 2025 Long-Term Reliability Assessment (January 2026; 224 GW demand surge projection; PJM/MISO/ERCOT high-risk timeline)
Wood Mackenzie — US gas turbine order backlog data, 2025 (paywalled; 44 GW estimate, sevenfold increase year-over-year)
GE Vernova — FY2025 earnings and backlog disclosures (46 GW → 83 GW → 100+ GW on contract)
Siemens Energy — FY2025 annual results (194 units sold vs. 100 in FY2024; €131 billion record backlog)
IEEFA — gas turbine planning timeline analysis, 2025 (seven-to-eight year advised planning horizon)
EPRI — Bobby Noble; large gas turbine lead times exceeding five years (cited in industry reporting)
RMI — interconnection queue timeline analysis (under two years in 2008 to over eight years in 2025)
BloombergNEF — combined-cycle gas plant capital cost escalation data, 2025 (49% cost increase; paywalled)
UBS — AI data center power research, 2025 (solar+storage “essentially the sole solution”; paywalled)
White House — Ratepayer Protection Pledge, March 4, 2026 (seven-company signing; no enforcement mechanism)
Virginia State Corporation Commission — GS-5 rate class order, November 2025 (25 MW+ / 75%+ load factor threshold; 15.8% cost increase for data center class)
Gartner — data center electricity demand forecast, November 2025 (China “better positioned due to more power-efficient servers and superior infrastructure planning”)
Brookings Institution — “electron gap” framing; US semiconductor advantage vs. China energy advantage
China National Energy Administration / State Grid — East Data West Computing program; CNY 5 trillion grid investment through 2030