Newsletter · · Ashutosh Agarwal

Data Centers Stop Waiting on the Grid and Build Their Own Power - Powering AI: Grid, Gas, Generation & Nuclear - Week of July 3, 2026

Hyperscalers and developers stopped waiting in the interconnection queue and started building their own power, mostly gas, at a roughly 80% premium, while NERC flagged data centers spontaneously severing from the grid and Deloitte pegged utility capex at $1.4 trillion through 2030. Our synthesis for the week of July 3, 2026.

Powering AI: Grid, Gas, Generation & Nuclear

Week of July 3, 2026: Data Centers Stop Waiting on the Grid and Build Their Own Power

Powering AI: Grid, Gas, Generation & Nuclear, Friday, July 3, 2026

For two decades the American grid grew at under 1% a year, and utilities got comfortable being boring. That era ended somewhere around 2023, and this week the tape finally said the quiet part out loud: the data centers are done waiting in line. They're building their own power. And the grid operators are starting to sound genuinely alarmed about what happens next.

TL;DR

  • The pods converged on one story: power is now the critical path, so hyperscalers and developers are bringing their own generation, mostly gas, at a roughly 80% premium to grid power just to turn the chips on. Bullish gas gensets, turbines and fuel cells; a warning shot for the "utilities capture all of it" thesis.
  • Deloitte's Tom Keefe put the number on the supercycle at EEI: $215B of utility capex in 2025, $1.4 trillion through 2030, but the money only flows if "forecasted" load becomes "contracted" load.
  • Uranium didn't move on genuinely bullish news. That's the tell of the week, and not a happy one for the bulls.

What's new

The grid is starting to break, and NERC said so. The most striking development this week came from InvestTalk (July 3): NERC issued a rare alert in May, only the third in its history, after clusters of data centers spontaneously severed their grid connections during minor line faults and flipped to on-site backup, threatening cascading failures. Data centers now swing 50+ megawatts within minutes; PJM has already been handed emergency DOE orders letting it curtail them as a last resort. The hosts' line: "the AI trade can't exist without the power trade."

So developers stopped waiting. On Open Circuit (July 2), Digital Realty's global head of energy Ian Black described the death of the old "find land, get in line for power" playbook. Power is now sequenced with everything else, and he's signed an energy-services agreement in 20 days built entirely on bring-your-own generation, while a separate ERCOT study has kept him waiting two years. The economics are staggering: data centers run $10-13 a watt versus ~$2 for solar-plus-storage, so a gas BYO plant turns $1B of renewable-equivalent spend into roughly $15B. That number permeates the whole value chain.

The bring-your-own-power wave, quantified. EnergyCents (July 2) gave the cleanest framing. Behind-the-meter used to mean a few megawatts; the average newly-proposed data-center project hit ~2 GW in Q1 2026 and keeps climbing. It's overwhelmingly gas, because the big-three heavy-duty turbines (GE, Siemens, Mitsubishi) are sold out, which opens the door for reciprocating engines from Caterpillar, Cummins and Wartsila, plus Baker Hughes and fuel cells. The kicker: this "speed-to-power" premium runs ~80% over retail, $140-150/MWh versus ~$80, and developers are paying it anyway, as a ten-year bridge until the grid shows up.

"Electricity, not AI models, will decide the winners of the AI race."

KR Sridhar, Bloom Energy

That was Bloom CEO KR Sridhar on 20VC (June 29), who dropped the stat of the week: Bloom put 50+ MW of fuel cells into Oracle's Utah data center in 55 days. A 2 GW campus takes 12-18 months to build; the power is the bottleneck, not the concrete.

And the utility bull case got its anchor number. At EEI 2026, Deloitte's Tom Keefe told Electric Perspectives (June 29) that investor-owned utilities spent $215B in 2025 and are on track for $1.4 trillion through 2030, reframing the group from defensive-dividend to long-duration infrastructure growth. The catch: two-thirds of the pipeline is still early-stage, so the winners are the ones converting forecasted load into contracted load via minimum bills, collateral and large-load tariffs (now 75+ across 36 states). The companion Alliant/QTS episode (June 30) showed the template live: a 500 MW-1 GW Iowa load where the data center funds the incremental build, and existing customers get a five-year rate freeze.

The debate

Bull: A durable, multi-year re-rating that lifts grid, utilities, equipment, gas and eventually nuclear together. Wood Mackenzie is tracking 220 GW of US data-center pipeline with 183 GW already committed, ~22% of US peak demand, per Interchange Recharged (June 30), which also flagged PJM commitments running ~3x accredited generation. Columbia Energy Exchange (June 30) counted 1,200+ proposed data centers demanding 100-300 GW, a third of today's US fleet. When even the skeptics argue about how to build it, not whether, the demand signal is real.

Bear: The tape is planting the seeds of doubt. If developers are paying an 80% premium for temporary gas because the grid is a decade behind, that's front-loaded bridge economics, not a permanent annuity for utilities whose pipeline may never fully get built. Equipment lead times are near peak (more below), which is usually late-cycle. And the on-site pivot means some of this load may never touch the rate base the utility bulls are underwriting. Notably, no one this week made a clean bear case on the merchant IPPs or restarts, that thread was quiet, so treat its absence as absence, not confirmation.

The names in play

GE Vernova was the one stock guests actually put on the table, though as chart, not fundamentals. On The Real Eisman Playbook (June 29), Strategas chartist Todd Sohn called GEV "one of my favorite stocks," a long-time hold that's overbought and consolidating but still "buyable," with the next leg keyed to more capex or earnings. Stock Market Today With IBD (June 30) echoed the "AI-energy play with a massive backlog" framing and flagged earnings in a few weeks as the catalyst, while warning GEV likes to gap up then base. Two technicians liking a chart is a setup, not a thesis.

Read-throughs

The gas-bridge story is where the money reads through. Caterpillar and Cummins are the direct beneficiaries: on Pitch The PM (June 30), Jarrett Harris of Iron Advisor Insights said CAT's power-systems segment is now driven by data centers, not fracking, and that CAT and Cummins fill the gap precisely because the big-three turbines are "years and years away." The second-order tell: yellow-iron lead times have stretched to 9-15 months and pricing is finally lifting after a long standoff, good for machinery now, but the kind of lead-time peak that eventually mean-reverts.

On the fuel cycle, the read-through is a caution. Uranium refused to rally on real news. On Mining Stock Daily (June 26), Sam Broom noted that $17.5B in US reactor loans (including full-size AP1000s) plus Canada's nuclear-renaissance push did essentially nothing to move Cameco or the uranium names, a macro-headwind sentiment tell despite tightening fundamentals. And ARC Energy Ideas (June 30) put a finger on the structural bottleneck: Canada can make its own CANDU fuel but has no licensed enrichment, still importing SWU from France and the US. The enrichment squeeze is real; the equity move just isn't there yet.

What we're watching into next week

The heat wave is the live stress test: PJM summer peaks will tell us whether the reliability worries NERC flagged are theoretical or imminent. And with utility earnings and GEV's print a few weeks out, the question is simple: does "contracted load" keep converting, or does the air-pocket the bears keep describing start showing up in the bookings?