Newsletter · · Ashutosh Agarwal

Power, Not Chips, Is the Binding AI Constraint as Operators Put Numbers on It - Powering AI: Grid, Gas, Generation & Nuclear - Week of September 22, 2026

The Powering AI: Grid, Gas, Generation & Nuclear brief for the week of September 22, 2026, synthesizing this week's energy and infrastructure investor podcasts. This issue leads with Edison International CEO Pedro Pizarro sizing the industry's grid spend at $1.4 trillion over five years and arguing data-center load can lower rates, then covers the 100 GW off-grid gas boom powering Meta's Prometheus with Williams, Oracle's 433 MW RWE power deal, Generac's $2.4 billion Amazon generator contract, Oklo's SMR timeline, the bull and bear debate on interconnection math, and why uranium's sharpest voices say the AI story is not what is moving the price.

Powering AI: Grid, Gas, Generation & Nuclear

Week of September 22, 2026: Power, Not Chips, Is the Binding AI Constraint as Operators Put Numbers on It

For most of this AI cycle, the argument that mattered was about chips. This week, the people who actually build and buy the infrastructure spent their airtime arguing about something else: electricity. A utility CEO put a $1.4 trillion number on the industry's five-year spend. A private-equity chief said the US grid has roughly three units of demand for every one unit of supply. And Oracle, in the same week its founder got subpoenaed, went out and locked up 433 megawatts of power years before it needs them. The tell of the week isn't a single stock. It's that "how do we power this" has quietly become the whole conversation.

The one-line version: power, not silicon, is now the binding constraint, and the companies feeding that constraint (grid, gas, nuclear, uranium, and the boxes in between) are where the operators are putting their money and their mouths.

TL;DR

  • A utility CEO said the quiet part with a number. Edison International's Pedro Pizarro pegged the industry's collective grid investment at ~$1.4 trillion over five years and argued a well-structured gigawatt of data-center load can actually lower everyone else's rates. That reframes data centers from a cost villain to a rate-base tailwind, if the contracts are written right.
  • Off-grid gas has gone from stunt to standard. More than 100 GW of behind-the-meter (self-powered) data centers have now been announced, with Meta's Ohio campus running on islanded gas built with Williams. Speed-to-power is worth $30-50B of revenue per gigawatt, so tenants are building their own power plants rather than wait 5-7 years for a grid connection.
  • The fuel cycle is on fire and the bulls and bears finally disagree. Uranium term price hit a nominal record (~$97/lb) with enrichment at all-time highs, but one respected resource investor flatly called the AI-uranium link overhyped. Real debate, at last.

What's new

1. The $1.4 trillion number came from a utility CEO, on camera. On Squawk on the Street (Sep 21), Edison International CEO Pedro Pizarro (who runs Southern California Edison, serving ~15 million people) gave the cleanest operator framing of the year:

"My colleagues and I are collectively investing something like $1.4 trillion over the next five years."

Then the part that matters for the thesis: every extra gigawatt of load, done right, "can actually reduce our system average rate by a half to 1.2%... well over a billion dollars in customer benefit over the lifetime of that investment." He reaffirmed Edison's 5-7% EPS growth through 2030 and backed the Ratepayer Protection Pledge so data centers pay for what they add. The catch he volunteered himself: SCE sits at the lowest investment-grade rating and Fitch just moved its outlook to negative on unresolved California wildfire liability. Rate-base growth is real; the balance sheet still has a tail.

2. Off-grid gas is now a 100-gigawatt category. The most useful hour of the week was energy researcher Michael Thomas on Inevitable (Sep 16), walking through the "Mad Max phase" of power development. The mechanics:

  • XAI's Colossus proved you can skip the utility queue by wheeling gas turbines in on the backs of semi-trucks. Dominion was quoting developers ~7 years to connect in Virginia; XAI got online in months.
  • The industry is now tracking more than 100 GW of these behind-the-meter projects, though only ~2 GW is operating today and ~5-10 GW is under construction. Most is still just announced.
  • Meta's Prometheus campus in New Albany, Ohio will exceed a gigawatt, powered by 200 MW of islanded natural gas built with Williams (WMB), with another 200 MW coming and the whole site scaling past 1 GW.

Why bother? Economics. "Every year that you can get online sooner is $10 to $50 billion of revenue," Thomas said. Data-center revenue per gigawatt has jumped from ~$10B a year ago to $30-50B on the latest SpaceX/XAI deals with Google and Anthropic. When speed is worth that much, tenants become power developers.

3. Oracle locked up power in the middle of a bad week, which is the point. On Telltales (Sep 20), hosts Marcus Graham and Ava Cabot made the sharpest read of the week: Oracle signed a 433 MW power purchase agreement with RWE on the same Thursday its founder was subpoenaed and days after he called off a ~$7.5B stock sale. Their line:

"Power is the binding constraint on this build-out, not chips, and companies about to slow down do not go and lock up electricity years ahead of needing it."

Set against ~$76B of trailing capex and ~$126B of net debt, the power deal is Oracle "behaving like a company that believes its own order book." Whatever you think of the stock, the behavior is the signal.

4. Generac just turned a storm-backup business into a data-center supplier. Same Telltales episode, and echoed by trader Joel Elconin on The KE Report (Sep 18): Generac (GNRC) signed a ~$2.4 billion long-term supply deal with Amazon for backup generators at its data centers, first delivery 2027-2028. Amazon also took warrants to buy up to $340 million of Generac stock, and the shares ran +19%. The number to sit with: that order is roughly four years of Generac's ~$600M annual free cash flow landing in one contract, and the stock was already at 29x FCF before Amazon walked in. As Telltales put it, you're now being asked to underwrite whether "a storm demand business with lumpy revenue turns into a contracted infrastructure supplier," because those two things don't trade at the same multiple. The 19% pop started that argument; it didn't settle it.

5. Oklo's CFO gave the most concrete SMR timeline we've heard. On Energy Evolution (Sep 15), Oklo CFO Craig Bellmer confirmed the 75-megawatt Aurora reactor went critical in August under the DOE pilot program and put hard dates on the rest:

"Our first powerhouse in Idaho, we are targeting for that asset to be operational and generating power in 2028. I'm always clear to note that that is an aggressive timeline."

The commercial anchor is a 1.2 GW project in Ohio with Meta, live in the 2030s. Oklo has raised ~$2.5 billion since going public, close to 30% of all SMR-developer fundraising S&P tracks. Two things a book should note: NuScale's 77 MW design was the first to win NRC standard design approval (the regulatory path of least resistance), and X-Energy has raised ~$2.8B this year, with Holtec and Westinghouse expected to list soon. This is becoming a crowded, well-capitalized field, which is bullish for the theme and dangerous for any single name.

The debate

For once, the week gave us both sides with conviction.

Bull: the whole stack re-rates together, and the operators are pre-committing capital. The through-line across the operator interviews is that this is a multi-year, self-reinforcing build. EQT's infrastructure chief Erwin Thompson on Dakota Live! (Sep 16) sized it bluntly:

"We're adding upwards of 15 to 20 gigawatts per annum of data center power at least... anywhere from a billion to five billion per gigawatt... you have 15 of demand with five of supply."

He sees hyperscaler capex heading toward $4-5 trillion by 2030 and thinks the risk is that demand surprises up, not down. Microsoft's Satya Nadella (All-In, Sep 15) put ~$80B behind Azure this year. Apollo's Jim Zelter (Bloomberg Talks, Sep 16) said the build will consume "any and all precincts" of capital. When utilities, private equity, and hyperscalers are all pre-committing, the bull case is simply: believe the contracts.

Bear: front-loaded bookings, a real air-pocket risk, and timelines that don't line up. The skeptics were louder this week than usual, and more credible for it.

  • The interconnection math doesn't clear. On Clean Power Hour (Sep 17), hosts cited ex-Microsoft executive Ramez Naam: 230 GW of data-center interconnection applications over five years against only ~93 GW of expected new power plants (BofA), a 2.5x gap. Not every announced gigawatt gets built; the host's own utility-scale pipeline gets discounted from ~90-100 GW to 30-40 GW in practice. "Is there a bubble? For sure," he said. AI demand is durable; the announcements are not all real.
  • The ROI test is getting stricter. On The Intrinsic Value Podcast (Sep 20), analysts argued Alphabet's ~$200B of AI capex needs to throw off $70-100B of new annual revenue (roughly $0.40 per dollar spent) with uncertain durability, even as Nebius and CoreWeave raise prices 100% and 25%. If the returns wobble, the power orders behind them wobble too.
  • The nuclear names already lived a hype cycle. Motley Fool Hidden Gems (Sep 17) used Oklo as the cautionary tale: down ~80% from a 2025 high near $200 to around $40, even with the long-term trend intact. The trend can be right and the entry point still wrong.

The honest read: no one this week argued the demand isn't coming. The bear case is about timing and price: that bookings are being pulled forward, lead times are near peak, and a lot of announced gigawatts quietly evaporate. That's a more dangerous bear than "it's all fake," because it can be true and the theme still wins over five years.

The names in play

  • Williams (WMB) graduated from pipeline operator to data-center power partner: it built Meta's 200 MW of islanded gas in Ohio, with more coming (Inevitable). Next catalyst: whether the second 200 MW phase lands on time and whether other tenants copy the structure.
  • Generac (GNRC) is the week's cleanest re-rating candidate and its cleanest trap: a genuine ~$2.4B Amazon contract, but a 29x-FCF starting multiple and unproven capacity to fill an order that size (Telltales). Watch the 2027 delivery ramp and what the capacity build costs.
  • Eaton (ETN), Vertiv (VRT), Quanta (PWR) are the "picks and shovels," per Elconin on The KE Report: Eaton for electrical gear, Vertiv for liquid cooling (he flagged it as a laggard worth a look), Quanta for substations. Pundit view, not an operator, but a clean map of where the spend lands.
  • Oklo (OKLO) has the most concrete SMR roadmap (Idaho 2028, Meta 1.2 GW in Ohio in the 2030s) and the most crowded, richly-valued setup (Energy Evolution). Catalyst: taking the Idaho powerhouse vertical, and NuScale/Holtec/Westinghouse listings that reprice the whole group.
  • Cameco (CCJ) sits at the center of the fuel-cycle squeeze: named as one of only three reliable long-term contracting producers, all with 10-15 year pipeline problems (Going Nuclear), and signing US utility contracts at higher prices (Other People's Money).
  • Edison International (EIX) is the rate-base-growth story with a wildfire asterisk you can't ignore (Squawk on the Street).

Read-throughs

How a move in one link ripples to the others:

  • Gensets → Cummins (CMI), Caterpillar (CAT). The Generac/Amazon deal is a template. If hyperscalers are contracting years of backup-generation capacity, the read-through runs to every engine and turbine OEM that can deliver on-site power, not just the one that signed this week.
  • Gas-for-power → Williams (WMB), Kinder Morgan (KMI), ONEOK (OKE), and gas-weighted E&Ps. NGI's Hub & Flow (Sep 17) put US power-sector gas demand up 27-28% by 2052, needing ~25,000 miles of new pipe; ARC Energy Ideas (Sep 15) was structurally bullish gas on the same data-center + LNG pull. The off-grid boom is a direct new take-or-pay customer base for midstream.
  • Copper & electrical steel → Freeport (FCX). Aurelion's Jeremie Boyer on Other People's Money (Sep 20) sees copper at $6.50-7.00/lb by year-end and notes the risk now shows up in AI companies' own disclosures (he cited Lumentum), though he'd avoid Freeport itself as lower-quality at ~36x. The tell: copper is a "right now" story, not a 10-year-deficit story.
  • Uranium & enrichment → Cameco (CCJ), physical vehicles, SWU names. Term price ~$97/lb (nominal record), SWU at all-time highs, floor-to-ceiling contract spreads blown out to $70-80, and China/Russia/Japan all buying (Going Nuclear). The read-through is the whole fuel cycle, not just miners.
  • The customers funding it all → Meta, Microsoft, Amazon, Alphabet, Oracle. Every power deal above traces back to a hyperscaler balance sheet. Naveen Rao on All-In (Sep 21) made the constraint vivid: global data-center energy is under 100 GW, ~12 GW of it going to Google alone, and on current math "exhausted in three years," with energy now ~50% of the cost of serving a token. When power is half your unit cost, power contracts become the strategy.

The one genuinely contrarian voice

Worth flagging on its own: veteran resource investor Rick Rule on The Competent Investor (Sep 17) is bullish uranium, but explicitly downplayed the AI/data-center angle, calling that demand 6-9 years out and less important than Japanese restarts and enrichment expansion. Justin Huhn made the same structural point from the other direction on Going Nuclear: the AI narrative "has had very little impact on the price of uranium." If you're long uranium for the AI story, two of the sharpest voices in the space just told you your reason is wrong even if your position is right. The commodity works on supply-demand and sovereign buying; the AI headline is sentiment, not fundamentals. Useful to know which one you're actually betting on.