Newsletter · · Ashutosh Agarwal

Bring Your Own Power and GE Vernova's Turbine Peak - Powering AI: Grid, Gas, Generation & Nuclear - Week of July 31, 2026

Powering-AI newsletter for the week of July 31, 2026. A $100B NextEra and Brookfield 'bring your own power' campus in Paducah, Kentucky set the template for self-contained power-plus-compute, even as GE Vernova printed record orders yet guided gas-turbine orders to peak in 2026, PJM warned of involuntary data-center cutoffs by mid-2027, and the debate shifted from how big capex can get to who is on the hook for it.

Powering AI: Grid, Gas, Generation & Nuclear

Week of July 31, 2026: Bring Your Own Power and GE Vernova's Turbine Peak

This was the week the "who pays, and who builds the power" question stopped being theoretical. A $100 billion project in Kentucky put a name to the new bargain: data centers bring their own power plant, and everyone else's electric bill is supposed to be protected. At the same time, the single best-loved way to play this whole boom, GE Vernova's gas turbines, got a quiet warning from the company itself: orders are about to peak. Two stories, pulling in opposite directions, in the same seven days.

TL;DR

  • The template arrived. NextEra and Brookfield unveiled a ~$100B "bring your own power" campus in Paducah, Kentucky: the data center comes with its own gas plant (and a side of uranium enrichment). Expect copycats, and expect utilities to lean on this to keep residential bills from spiking.
  • Peak turbines? GE Vernova printed monster orders (+88% year over year, a $176B backlog), yet the company is telling people gas-turbine orders likely peak in 2026. The most crowded long in the space now has a date on the calendar to argue about.
  • The grid can't keep up, literally. PJM, the biggest US grid operator, warned that data centers without contracted power could face involuntary cutoffs as soon as mid-2027, and one utility says it will triple in five years the capacity it took 125 years to build.

What's New

1. "Bring your own power" got a $100 billion poster child: Paducah, Kentucky. This week on Bloomberg Talks, US Energy Secretary Chris Wright described the NextEra and Brookfield project on federal land in Paducah: a 1.2 gigawatt data center paired with 2 gigawatts of brand-new natural gas generation and an extra 800 megawatts pushed onto the local grid, with construction starting inside six months and phasing in over four years. His framing is the whole thesis in one line: the data center is "funding and underwriting the development of so much new power generation," and if AI ever over-builds, "we would love to have the extra power" for reshoring factories. (Worth noting: The TreppWire Podcast described the same campus a touch bigger: a 1.8 GW data center with roughly 2 GW of gas plus up to 2.6 GW of battery storage, targeting completion in 2031.) Either way, the shape is what matters: a self-contained power-plus-compute campus that skips the interconnection queue. Why it moves numbers: this is the model that lets NextEra (NEE) and its peers add rate base without triggering a consumer-bill backlash, and it drags gas turbines, gas supply, batteries and even enrichment along with it. The same episode confirmed a new company, General Matter, is building modern uranium enrichment on the very site that enriched uranium during the Cold War.

2. GE Vernova: the best number in the group, and the most uncomfortable one. On The Real Eisman Playbook, Steve Eisman walked through GE Vernova's quarter: EPS was actually a miss, but orders jumped 88% year over year to $24.2 billion, lifting the backlog to $176 billion. He reminded listeners there are only three companies on earth that build large gas turbines, GE Vernova, Mitsubishi and Siemens, and said that in a long-tail business like this, orders are the number that matters, so he's staying long "even at these nosebleed elevations." Then the uncomfortable part, from The Uptime Wind Energy Podcast: GE Vernova is guiding gas-turbine orders to peak in 2026 and then trend down, because the queue is already about five years out and few new buyers will line up to wait seven or eight. Why it moves numbers: the bull case (record orders, unmatchable backlog) and the bear case (management-flagged order peak, capacity it can't ramp) are now living inside the same company. If you own the turbine trade, the debate you have to win is whether a five-year backlog is a fortress or a warning.

3. The demand math keeps getting bigger, and more physical. On TraderMerlin, the host laid out the numbers cleanly: US data-center electricity use was 176 terawatt-hours in 2023 (4.4% of all US power) and is projected to hit 325–580 TWh by 2028 (6.7%–12%), the low end alone is a doubling. Utilities are forecasting roughly 166 gigawatts of peak-demand growth over five years, with data centers about 54% of it, and Goldman Sachs sees data-center power demand doubling by 2027. Make it concrete: on The MAD Podcast, Samsara CEO Sanjit Biswas said a large utility told him it will triple in the next five years the grid capacity it built over the prior 125 years, and 90% of that new demand is data-center related. He also flagged the bottleneck nobody models: there simply aren't enough electricians, which is why Meta is running programs to reskill people into the trade. Why it matters: the demand isn't the debate anymore. Turning it into delivered electrons (transformers, transmission, and warm bodies to wire it) is.

4. The grid is already saying "no." The TreppWire Podcast surfaced the sharpest near-term risk of the week: PJM Interconnection, the largest US grid operator, and the one that runs Northern Virginia's "data center alley," warned that data centers without sufficient contracted generation could face involuntary power cuts as early as mid-2027, after a capacity auction failed to secure enough supply to cover a projected shortfall of nearly 7 gigawatts starting in June 2028. Why it moves numbers: tightening capacity is exactly what re-rates merchant power and capacity payments, and it's the hard evidence behind why hyperscalers are willing to build their own plants rather than wait in line.

5. The backlash, and the utilities' answer to it. Two operators framed the politics. On Power & Politics, New York Governor Kathy Hochul explained her one-year moratorium on hyperscale data centers of 50 megawatts or more (each one draws as much power as 50,000 homes). She has 30 pending applications that together would outstrip available supply (in one county, five applications would eat a full nuclear reactor's worth of output), and her Public Service Commission is writing a premium rate structure so big consumers "bring their own power or pay a premium." Tellingly, she says the developers admitted to her they don't even need tax breaks: "We're making enormous sums of money." The industry's counter came on Power Lunch, where Duke Energy CEO Harry Sedaris said Duke's signed data-center contracts should deliver $3.6 billion of savings to North Carolina customers over 15 years by spreading that contracted revenue over the system's fixed costs, while Duke builds 14 gigawatts of generation over five years (mostly gas, plus solar and batteries) and can require data centers to curtail load 50–100 hours a year at peak. Why it matters: the "data centers raise my bill" story is now a live political risk, and the regulated utilities that can credibly answer it (DUK) get to grow rate base with the wind at their back; the ones that can't will get moratoriums.

The Debate

Bull: this is a multi-year supercycle, and the whole stack re-rates together. The demand is real and enormous (TraderMerlin, MAD); the buyers have effectively bottomless balance sheets and are willing to underwrite new generation directly (Bloomberg Talks); the grid is tight enough that capacity has pricing power (TreppWire); the turbine makers have order books measured in years (Real Eisman Playbook); and on Volts, David Roberts' guest called this the fastest US energy growth since World War II, with a second, even larger electrification wave (EVs, heat pumps) queued up behind the AI wave. If you believe that, gas, grid, utilities, nuclear and copper all move together for years.

Bear: a lot of this is front-loaded, and the podcasts flagged the cracks themselves. GE Vernova, the poster child, is telling you orders peak this year (Uptime Wind); a five-year backlog with a peak in it is the classic shape of a capex air-pocket. The nuclear leg is mostly hope: on Switched On, BloombergNEF's analysts said they track roughly 51 gigawatts of announced hyperscaler-nuclear deals, but almost none have binding offtake, financing, or even an NRC application, and of the ~50 SMR and advanced-reactor developers, they expect only two or three to survive. The near-term nuclear that's actually real is small: about 2.2 GW of restarts (Palisades, Constellation's Crane/former Three Mile Island with Microsoft, NextEra's Duane Arnold with Google) plus ~2.3 GW of uprates by 2036, and gas still carries the load through 2030. Then there's the money: on The Real Eisman Playbook, the story of the week was Google lifting 2026 AI capex from $190B to $205B and going free-cash-flow negative, with the market no longer cheering. On Squawk on the Street, the tell was that Goldman Sachs is now selling hedge funds a CDS basket to bet against the hyperscalers. And the politics are turning (Power & Politics). The power demand may be durable; the pace of the capex funding it may not be.

The honest read this week: the demand side is genuinely one-sided (nobody credible argued US electricity load is about to fall). The disagreement is entirely about timing and financing, whether the equipment order cycle and the trillion-dollar capex behind it hold up long enough for the physical build to catch up.

The Names in Play

  • GE Vernova (GEV). The cleanest expression of the boom and now the cleanest expression of the debate. Bull: $24.2B orders (+88%), $176B backlog, one of only three large-turbine makers on earth. Bear: management's own 2026 order-peak guide. Next catalyst: any color on 2027 order intake and whether it can lift turbine production capacity. On Squawk on the Street, Jim Cramer's shorthand for powering OpenAI's proposed 10 GW Ohio complex was blunt: "get the GEV, get the GE Vernova turbines."
  • NextEra (NEE). Named owner of the Paducah "bring your own power" template (Bloomberg Talks) and a real near-term nuclear restart (Duane Arnold, with Google) (Switched On). Catalyst: proof the closed-loop campus model actually breaks ground on schedule.
  • Duke Energy (DUK). The utility making the affordability case out loud: $3.6B of customer savings, 14 GW of five-year build (Power Lunch). In a week of backlash headlines, that's the political cover a regulated utility needs to grow.
  • Constellation (CEG). Owns Crane (the former Three Mile Island), being restarted with Microsoft, one of the few nuclear stories with a real date attached (Switched On).

Read-throughs

  • Turbine OEMs to gas supply. Every "bring your own power" campus is a gas plant. Paducah alone is ~2 GW of new gas (Bloomberg Talks, and on Inside Data Centre, operators described going behind-the-meter with ~300–500 MW of on-site gas just to skip the queue. That's incremental, sticky gas demand, good for the gas-weighted E&Ps and midstream that feed it, and it runs straight through the same three turbine makers.
  • Copper and the wire. You cannot triple a grid without copper. Across Interchange Recharged (demand doubling by 2035), Anything But Typical (miner Robert Friedland's line that we must mine more copper in the next 15 years than in the last 10,000, with ~17-year lead times to open a mine), and The David Lin Report (AI data centers are ~6% copper by cost; a call for copper above $8/lb), the message was identical: structural deficit. It's the read-through to Freeport (FCX) and the cable/conductor names, even though no guest named a specific stock.
  • Uranium and enrichment. The Paducah project revives domestic enrichment via General Matter (Bloomberg Talks), and the restart-and-uprate wave (Switched On) is incremental pull on fuel, supportive of the converter/enricher and physical-uranium thesis (Cameco and peers), even as the big SMR offtake stories stay mostly on paper.
  • Who funds the deal. The hyperscaler customers are the balance sheet behind all of it, and it's getting levered. The TreppWire Podcast flagged BlackRock's $12B debt financing for a Meta data center in El Paso, and on Money For the Rest of Us, the estimate was $6–7 trillion of total data-center-plus-power-grid capex, with J.P. Morgan projecting 75% ($4 trillion) funded by debt. That's the tail risk under the whole power trade: if the financing wobbles, the demand for turbines, transformers and gas wobbles with it.

What Changed

The tone flipped from "how big can capex get" to "who's on the hook for it." A month ago the debate was demand size; this week it was financing (Google FCF-negative, Goldman selling hyperscaler CDS) and permission-to-build (Hochul's moratorium, Duke's affordability pledge, PJM's shortfall). And for the first time, the reigning turbine bull case has a company-issued expiration date on the order cycle. Nothing here breaks the multi-year thesis, but the market has started pricing the pace, not just the size.