Newsletter · · Ashutosh Agarwal

Washington Puts a Number on AI's Power Problem at 100 Gigawatts by 2030 - Powering AI: Grid, Gas, Generation & Nuclear - Week of August 14, 2026

Powering-AI podcast roundup for the week of August 14, 2026. The White House sized the job at 100,000 megawatts of new generation by 2030, PJM's capacity auction hit its cap for a third straight time, and the tape split over whether this is a durable supercycle or a financing accident waiting to happen.

Powering AI: Grid, Gas, Generation & Nuclear

Week of August 14, 2026: Washington Puts a Number on AI's Power Problem at 100 Gigawatts by 2030

This was the week the AI power story stopped being a slide in a hyperscaler deck and started sounding like a plan. On one podcast, a senior White House energy official and an energy hedge-fund manager sat down and put hard numbers on how much electricity America has to build, how it gets paid for, and who wins. On another, the head of the country's main power-research institute bet a skeptic a curry over whether it's really 100 gigawatts or a third of that. And on a third, a couple of practitioners walked through the auction result that just told 67 million people their power bills are going up.

Put together, the week did something useful: it turned a vibe into a spreadsheet. Here's what mattered, why it moves numbers, and where the argument is still very much alive.

TL;DR

  • The government finally sized the job. The White House is now working to a benchmark of 100,000 megawatts of new generation by 2030 (think a whole second Texas grid) with roughly half driven by data centers and the rest by reshoring and electrification. That's the anchor for a $1.4-trillion-ish grid capex decade.
  • The grid is already sending the invoice. PJM's latest capacity auction hit its price cap for the third straight time, and the annual bill for keeping the lights on jumped seven to eight times versus a couple of years ago. New data centers over 50 megawatts are being told to bring their own power or curtail.
  • The bull case is a financing case, and so is the bear case. The optimists point to 15-to-20-year take-or-pay contracts from the richest companies on earth. The pessimists point to the same balance sheets, and to a nuclear timeline that doesn't arrive until the 2030s.

What's new

Washington put a number on it, and a financing model behind it. The single most useful hour of the week was on Thematic Investors, "Fast-Tracking the US Grid – Policy, Power Demand, and Infrastructure", where host Kieran Cavanna sat down with Peter Lake, the White House Senior Director of Power and former head of Texas's grid regulator, and Ravi Beller, who runs the energy-focused hedge fund Ripple Effect after stints at Goldman, BlackRock and Point State. Lake said the administration is now working to a Department of Energy benchmark of "100,000 megawatts of new power generation needed by 2030," which, in his words, is like "adding a California and a New York on top of our existing power grid." About 50,000 of those megawatts come from data centers; the other half is ordinary economic growth, electrification, and bringing manufacturing home. Beller sized the money around it: $1 to $1.5 trillion on transmission and distribution over five years (the poles and wires, mostly billed back to customers by regulated utilities), plus a new $250-350 billion on actual power plants. This matters because it converts a slogan into rate base and contracted cash flow, the two things that drive utility and independent-power-producer earnings.

The bills are already arriving, and the grid is short. On Renewable Rides, "Ep 119: How to Buy an Onsite Energy System Without Overpaying for It", VECKTA's Gareth Evans and Dan Roberts walked through PJM's July 14 capacity auction, the yearly process where the largest US grid operator buys a promise that enough power will show up when demand peaks. It cleared the federally approved cap of $325 per megawatt-day for the third consecutive time, and even at that capped price PJM landed a 14.4% reserve margin against a 20% target, roughly 7 gigawatts short. The dollar figure is the headline: the annual capacity bill went from about $2.2 billion to $14.7 billion for the year that began June 2025, and sat near $16.4 billion at the latest auction, a seven-to-eight-fold jump. Their punchline for investors: starting June 2027, new data centers above 50 megawatts in PJM territory have to curtail during shortages, which is why a proposed $100 billion AI campus in Kentucky is skipping the grid entirely and self-building 2 gigawatts of gas generation plus 2.5 gigawatts of batteries.

"On-site power is no longer an alternative for all these major new loads. Control over their supply is becoming a mandatory table stake." (Dan Roberts, Renewable Rides)

The $500 billion machine that funds the load. On Market Maker, "The $500 Billion AI Debt Machine & The Great CapEx Test", the hosts unpacked how the demand actually gets financed. The AI cloud provider CoreWeave now runs 1.5 gigawatts of active power with contracts to reach 4.2 gigawatts (about the electricity of three million homes) against a revenue backlog that has swollen to roughly $130 billion, up 246% year over year. The plumbing behind it: NVIDIA's new partnership with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500 billion to underwrite AI infrastructure, and a CoreWeave $8.5 billion bond rated A3 by Moody's, the first GPU-backed loan to reach investment grade, on NVIDIA's implicit backstop. Why an energy reader should care: this is the money that ultimately signs the power contracts. It's also the week's cleanest bear tell: CoreWeave spent $9.4 billion of capex in a single quarter on about $2.5 billion of revenue, with quarterly interest expense of $640 million. As the hosts put it, the whole thing works "as long as the hyperscalers keep their commitments" and wobbles the moment spending merely decelerates.

The politics turned, fast and bipartisan. On Open Circuit, "Data centers have a Texas-sized energy problem", former Department of Energy loan chief Jigar Shah and his co-hosts flagged what they called "the most active political story in the US right now": Texas Governor Greg Abbott ordered regulators to freeze new data-center grid connections pending an audit of water, power and tax breaks, days after New York Governor Kathy Hochul paused approvals of data centers over 50 megawatts for a year. The constructive nugget for utility investors was buried underneath the noise: Virginia's regulator issued a cost-causation ruling that pushed most of a transmission upgrade onto the data-center tariff rather than households, cutting a residential increase from about $3 a month to $0.94. That's the template that lets the buildout proceed without a voter revolt, and it's spreading (Senator Heinrich's proposed Grid Savings Act, plus rules moving in New Jersey, Pennsylvania, Virginia and Oregon).

The debate

The tape this week was unusually two-sided, so let's steel-man both.

The bull case: a durable supercycle that re-rates the whole stack together. Beller's framing on Thematic Investors is the cleanest version. He argued US power demand is set to grow 3-4% a year for five years, the fastest in half a century, and that generation portfolios of gas, renewables and batteries are trading at just 7 to 11 times free cash flow with 50-100% of cash flows already contracted for 12 to 20 years. New investments are earning mid-teens unlevered returns, 30-50% levered, better than prior cycles. Crucially, he's not worried about the usual infrastructure killer, oversupply: "there's limited labor, there's limited equipment, there's limited land, there's limited permitting," so the biggest existing developers of gas, renewables, storage and nuclear are "probably going to be the long-term winners here over the next decade." Lake's policy scaffolding reinforces it: a fast-track interconnection that connects curtailable loads in 60-90 days instead of six to eight years, a PJM 15-year price signal for long-dated assets, and a take-or-pay "ratepayer protection pledge" that pushes hyperscalers to build, bring or buy their own power. Hyperscalers, Beller noted, can "pay 40 to 50% more for 15, 20 years at a time" and keep the capex off consumers' backs.

The bear case: front-loaded bookings, a nuclear mirage, and a bond-market accident. The counterweight came from Cleaning Up, "AI Has An Electricity Problem", where host Michael Liebreich pushed back hard on EPRI chief Arshad Mansoor (more on their bet below). His three worries travel straight to the power complex. First, timing: only about 2.5 gigawatts of AI compute actually got built in 2025, so the jump to Mansoor's 100 gigawatts by 2032 is "an improbable step change." Second, nuclear is not the escape hatch this decade: Mansoor himself conceded advanced reactors and small modular reactors (SMRs, factory-built mini-plants) won't be a meaningful contributor before 2032, reaching real scale only "2035 and beyond," because you have to build ten of a design before you can build a hundred. Third, and most unsettling, the financing risk: the hyperscalers are betting on the order of $2 trillion on data centers and short-lived GPUs, and "if it goes wrong, it wipes out really the entire tentpole that's holding up the US economy... What happens when it's a bond market correction?" The Market Maker circular-financing thread is the same fear wearing a different hat.

And there's a quieter bear argument that caps the upside without any crash at all: efficiency and flexibility. On Catalyst, "Is now the time for DERs to scale?", Energy Impact Partners' Shayle Kann made the case that demand response and distributed resources (today around 20-25 gigawatts and, he thinks, capable of doubling in three to five years) are getting cheaper relative to ever-pricier gas turbines and retail power. Mansoor made the sharper version on Cleaning Up: if data centers agree to flex during a handful of stressed hours a year, "100 gigawatt doesn't need 100 gigawatt of infrastructure investment. Maybe it needs 50 gigawatt." Every gigawatt that flexes is a gigawatt of turbines, transformers and copper that never gets ordered.

The names in play

A few stocks actually had a thesis moved on them this week.

  • Southern Company got a genuine on-air upgrade of narrative. On Thematic Investors, Lake singled it out for adapting to the new nuclear paradigm and pointed to a $26 billion Department of Energy loan, "they're turning dirt now." He also credited Meta and Google (in Georgia) for the long-term off-take deals that are making plant restarts and new build financeable in the first place.
  • NVIDIA is quietly the largest energy-finance story on the board. As Market Maker laid out, its balance sheet is the implicit backstop that turns GPU-backed debt investment grade, which is exactly what lets the neoclouds sign the power contracts. The bull reads it as demand certainty; the bear reads it as circularity.
  • Nuclear and uranium got a reality check. Beller was blunt: "We are long gas. We are short uranium." Mansoor's SMR timeline on Cleaning Up says the same thing in slower motion: restarts and uprates are real now, but the SMR volume story is a post-2035 event. For anyone long the pure-play SMR names or the uranium trade on a 2026-28 view, that's the bear whispering.
  • Gas-genset makers were the week's sleeper. Both Catalyst (Kann calling a "flexible distributed natural-gas genset" the closest thing to a perfect grid resource, and flagging portfolio names Enchanted Rock and GridBeyond) and Cleaning Up (Mansoor building microgrids with Caterpillar and Wärtsilä units) point the same way: when the grid can't deliver, the load buys its own engines.

Read-throughs

Here's how one link pulls the others.

  • Gensets (CAT, and by extension CMI/Cummins): The Kentucky campus self-building 2 gigawatts of gas on Renewable Rides, plus Mansoor's Caterpillar/Wärtsilä microgrids and Kann's genset endorsement, all say behind-the-meter engines are becoming a table stake. That's an order book for on-site power equipment that grows faster the more the grid disappoints.
  • Copper, transformers and electrical steel (FCX and the cable/conductor names): Beller's "limited equipment" line and Mansoor's constraint list (transformers, electrical equipment, generators) are the same message from two directions: the bottleneck is physical gear, which is pricing power for whoever makes the metal and the boxes. The flip side: if flexibility trims 100 gigawatts to 50, some of that order book never lands.
  • Gas-weighted E&Ps and turbine OEMs: "We are long gas" plus a 4-to-5-year turbine lead time (per Beller) is a durable demand pull for dispatchable gas and for anyone with turbine slots to sell. Incremental data-center gas demand is the read-through that ties midstream (Williams, Kinder Morgan, ONEOK) to the power theme.
  • Enrichers, converters and physical uranium: the bear read-through. If SMRs are a 2035 story and a serious investor is openly short uranium, the near-term fuel-cycle trade is leaning on restarts and uprates, not new reactors.
  • Hyperscaler credit (MSFT, GOOGL, META, AMZN, ORCL): On Getting Credit, "The Cost of Computation", the team noted hyperscaler debt has now overtaken banks as a share of the US investment-grade bond market (the big five borrowing on the order of $100-200 billion a year) with Microsoft rated AAA (above US Treasuries) and Oracle down at BBB-, a notch above junk. These are the customers funding every power deal in this newsletter; their cost of capital is the sector's cost of capital.

Worth a curry

The best illustration of the whole debate was a friendly wager. On Cleaning Up, EPRI's Arshad Mansoor bet host Michael Liebreich a dinner at "the best Indian restaurant in London" that the US adds 100 gigawatts of AI compute by 2032; Liebreich is holding to roughly 30 gigawatts by 2030, in line with Jigar Shah's 34. To put 100 gigawatts in perspective, Mansoor noted US peak demand is about 450 gigawatts, so he's forecasting "two and a half new UKs" bolted onto the American grid in six years. Whoever's right, the sizing gap between them (call it 3x) is the entire investable question for this sector. It's the difference between a supercycle and an air-pocket.