Newsletter · · Ashutosh Agarwal

AI's Power Buildout Meets Its First Real Math Test - Powering AI: Grid, Gas, Generation & Nuclear - Week of July 21, 2026

Powering AI infrastructure newsletter for the week of July 21, 2026. The builders admit they cannot build fast enough while Jim Chanos put a number on the doubt, and the week's podcasts turned from asking how much power AI needs to asking at what return, with the flags landing on the arms dealers, utilities, turbines, copper and nuclear.

Powering AI: Grid, Gas, Generation & Nuclear

Week of July 21, 2026: AI's Power Buildout Meets Its First Real Math Test

For two years, the only question anyone asked about AI and electricity was "how much?" This week the podcasts started asking a harder one: "at what return?" The demand story didn't crack. If anything, the people actually building admit they can't build fast enough. But the smart-money skeptics finally got specific about where and when the money stops making sense. That gap, between a physical buildout running flat-out and a financial buildout whose margins are quietly thinning, is the whole story right now.

TL;DR

  • The builders are blunt: OpenAI's own compute chief says demand "far outstrips" supply, the constraint is the physical world, and every new data center now comes bundled with new power generation, transmission and transformers the grid wouldn't otherwise have gotten.
  • The bears got precise: Jim Chanos says the hyperscalers' return on each incremental dollar of capex has fallen from about 40% to roughly 20% and is heading toward 10%, with a "reckoning" he pencils in for late 2026 into 2027.
  • Follow the money down the stack: the next great AI financing wave, per Morgan Stanley, is energy and power itself, and the picks-and-shovels names (electrical gear, gas turbines, copper, nuclear) are where guests kept planting flags.

What's new

The clearest "we cannot build fast enough" yet, straight from the buyer. On The MAD Podcast with Matt Turck, OpenAI's Head of Industrial Compute, Sachin Katti (previously Intel's CTO), described the buildout as "one of the largest things humanity has ever built." His line that matters for anyone long power: "Demand far outstrips compute supply today. So anything we can bring online, we consume immediately." The binding constraint isn't demand, it's that "the physical world does not move that fast." This is the single most valuable kind of source we get: not a pundit guessing, but the operator writing the checks.

And that buyer is now funding the grid itself. Katti spelled out something that changes how you think about who pays for generation. Whenever OpenAI builds a data center, he said, "we make it a hard commitment that we are not taking power away from the grid", and instead, "we are investing in the grid to generate new power." In practice that means funding new gas, solar or hydro generation, plus "transmission lines... transformers, substations." Where the grid is tapped out, they move behind the meter with on-site gas turbines, "the most dense, transportable form of energy", but there, he admitted, "we are bottlenecked" on turbine supply. On nuclear, asked if it interested him: "Absolutely. It can't come soon enough." That is a demand signal for turbine OEMs, transformer makers and reactor developers coming directly from the largest AI buyer.

The hard pipeline numbers, from someone who counts them for a living. On The Data Center Frontier Show, DC Byte's Colby Cox gave the kind of concrete data the headlines usually skip. Projects of 900 megawatts or larger that DC Byte counts as "committed" have gone from 3 in early 2023 to 17 today; early-stage projects went from 7 to 49, and 15 of those are above 2 gigawatts, "with a couple getting into the 10 range." He also flagged the bottleneck: about 20% more projects are now stuck in the committed and early-stage buckets, "a lot of that being driven by the power availability," which is pushing developers behind the meter. One more number that reframes the equipment thesis: rack density has gone from around 6 kilowatts when he started, to 30, to "serious planning around that 100 kW" and some designs "pushing 300 and beyond." Denser racks mean more copper, more switchgear, more cooling, per megawatt.

Wall Street's plumbing is being rebuilt around this. Thoughts on the Market put Morgan Stanley's debt-capital-markets chief Anish Shah alongside credit analyst Lindsay Tyler, and the framing was striking: hyperscaler bond issuance has gone "from less than 1% of the investment grade market to more than 10%," and AI-related funding "could top 15% of the total issuance across all credit products." They put a price on a gigawatt: "roughly $12 billion for the shell and often more than double that for chips and racks," with a handful of players adding 30-plus gigawatts over two years, around $2 trillion of capex. And the tell for our sector: asked what the next AI financing opportunity is, Shah said, "It most certainly is energy and power... a ton of capital being raised in utilities."

The skeptic put a number on the doubt. On RiskReversal Pod, veteran short-seller Jim Chanos gave the most quantified bear case of the week. He's watching the hyperscalers' return on incremental invested capital, the profit on each new dollar spent, and says it "has gone from, as a group, 40% a year and a half ago to about 20% today," and "if the spend keeps up at this kind of rate, it's going to be moving toward 10%." His core accusation is a duration mismatch: "people are committing long-term capital projects based on near-term spot pricing," building 20-year assets against one-to-two-year contracts. He reached for history: shale, and 19th-century railroads that "built double and triple capacity" before "freight rates collapsed and railroads went bankrupt." His predicted moment of truth: "late 26, 27."

The debate

This was a genuine two-sided week, so here is the steel-man of each.

The bull case: this is a real, physical, multi-year supercycle, and the whole stack re-rates together. The strongest version isn't from a promoter, it's from the buyers and the builders. Katti's "we consume it immediately" and his commitment to fund new generation means the demand pull reaches all the way down to turbines, transformers and reactors. Cox's pipeline numbers show the committed backlog is real and growing, not vaporware. And BlackRock's Larry Fink, relayed on Wall Street Unplugged, was, in the host's words, "jumping up and down" on CNBC about the need to build power, putting the all-in cost of one gigawatt at "$50 to $60 billion" and saying America needs not 100 but "over 200, 250" gigawatts. Even the bull's favorite hedge showed up via former Fed governor Kevin Warsh: if AI disappoints, "you're still going to have power... the infrastructure built that can be used for something else." In other words, the electrons are useful even if the chatbots aren't.

The bear case: this is front-loaded spending into thinning returns, and the "scarcity" that justifies the multiples is already leaking. Chanos supplies the math (returns halving, then halving again). And a second RiskReversal Pod episode, pointedly titled "The AI Scarcity Myth Is Breaking," pressed the softest spot in the bull thesis: if compute is so scarce, why did Meta bounce this week on news it will rent out excess compute? As the host put it, that "blows a hole in the whole bull thesis of scarcity." The much-hyped compute deals Anthropic and Google struck with Elon Musk's XAI, roughly $1 billion each, turned out to be three-month contracts, not the durable long-term offtake the story implies. On the demand-durability side, Brett Rentmeester of Windrock Wealth on Wealthion laid out the spend curve, hyperscaler capex from $156 billion in 2023 to $443 billion in 2025 to an estimated $920 billion by the end of 2026, and named the fear plainly: "you could get to a point... where the buildout gets ahead of the demand."

Where they actually agree. Notice the bears aren't shorting the power. Chanos's worry is capital-intensity and financing (especially off-balance-sheet), not that data centers won't need electricity. Warsh's hedge and Chanos's warning are two sides of the same coin: the electrons and steel likely get built and stay useful; it's the equity multiples and the leverage stacked on top that carry the risk. For a power book, that distinction is the whole game: it argues for the arms dealers over the spenders.

The names in play

Electrical equipment and contractors. On CNBC's Fast Money, TCW portfolio manager Eli Horton made the cleanest single-stock case of the week: the power-infrastructure layer is constrained, and he pointed to Powell Industries (POWL) and Bloom Energy (BE) as "picks-and-shovels" ways to play rising electricity demand from data centers, reshoring and electrification. He also flagged that Quanta Services (PWR) benefits from the skilled-labor bottleneck on these projects, a reminder that the scarce input isn't always the equipment; sometimes it's the electricians.

Nuclear, blue chips and moonshots. Stock Club gave the tidiest map. The hosts cited Bank of America's view that nuclear is a "$10 trillion industry" with demand tripling by 2050, and the administration's goal to quadruple US nuclear capacity from 100 to 400 gigawatts. They split the field into "blue chip" operators, Constellation (CEG) and Vistra (VST), with Microsoft backing the Three Mile Island restart for Constellation, and "moonshot" small modular reactors: NuScale (SMR), targeting a reactor to market around 2030, and Oklo (OKLO), the Sam Altman-backed name that sells directly to data-center operators and is targeting its first unit in 2027. Their sober caveat is worth repeating: only two SMRs run commercially today, in Russia and China, and both came in over budget and behind schedule. On Strategy Sunday, Armando Pantoja made the bull frame explicit, "intelligence cannot scale without power," so Oklo and the SMRs are the energy layer of AI, while noting Oklo's stock drew down this week on broad market volatility, that it holds about $2.5 billion in liquidity, and that it "could be a year or so" before the next big move. Useful expectation-setting for anyone tempted to chase.

Read-throughs

  • The turbine bottleneck is the tell. Katti naming on-site gas turbines as the behind-the-meter answer, and then admitting they're supply-constrained, reads straight through to the turbine and genset makers. When the largest AI buyer says the scarce item is turbines, that is a demand statement for GE Vernova (GEV) and the on-site power names like Cummins (CMI) and Caterpillar (CAT), even if this week's podcasts didn't name them directly.
  • Denser racks mean more copper. Cox's 6kW-to-100kW-and-beyond density curve is a per-megawatt intensity story. Equity Mates put a figure on it: Microsoft's $500 million Chicago data center used 2,177 tons of copper, global data-center energy demand grew 17% in 2025 and is expected to double by 2030, and Freeport-McMoRan (FCX) sits alongside BHP as the largest or second-largest copper producer in the world. More density, more electrification, more copper, a clean read-through to FCX and the conductor/cable chain.
  • Utilities are the next debt story, not just the next demand story. Morgan Stanley's "energy and power is the next financing wave," plus junior subordinated debt out of the utilities, says the capital that funded the shells is about to fund the grid, a tailwind for rate-base-driven names and a reason to watch balance-sheet capacity, not just load-growth headlines.
  • Behind-the-meter is becoming the default bridge. Between Cox ("behind the meter is going to work really well... a necessity for a little while"), Katti's on-site turbines, and the demand-response tricks Eric Olson described on The Dynamo Show, including Google funding its own wind-plus-storage in Minnesota and developers curtailing load to jump the interconnection queue, the message is that the grid can't move fast enough, so buyers are routing around it. That's bullish for on-site generation and storage, and a caution flag for anyone assuming every gigawatt flows through a regulated utility's meter.

One thing worth flagging

Olson, a former Department of Energy hand now consulting on the regulated market, gave the most honest description of the risk nobody prices: utilities are making "crazy" and wildly divergent 20-to-30-year load assumptions, and if the data centers don't fully show up, "maybe it's only 70... maybe it's only actually 30", someone is left holding a power plant nobody needs. His memorable line for how creative this is getting: in some places "it's easier to give like 20,000 people heat pumps" to free up 100 megawatts "than to deal with the interconnection queues." That's the overbuild-and-stranded-asset risk the bulls wave away and the bears can't quite pin down, and it's the thread to pull on next week.