Newsletter · 2026-09-22 · Ashutosh Agarwal

Edison CEO Says a $1.4 Trillion Grid Build Can Cut Power Bills as AMD Tops $1 Trillion - The AI Capex Tracker - Week of September 22, 2026

The AI Capex Tracker for the week of September 22, 2026, synthesizing investor podcasts published September 20 and 21. This issue covers Edison International CEO Pedro Pizarro putting the electric industry's five-year investment at $1.4 trillion and arguing each extra gigawatt of data-center load can cut customer rates, Amazon's $2.4 billion Generac backup-generator deal with warrants, Oracle's 433 MW RWE power purchase agreement and equity-funded build, AMD briefly topping a $1 trillion market cap, and RockLinc's Jonathan Wellum warning of a debt bubble and coming write-offs.

The AI Capex Tracker

Week of September 22, 2026: Edison CEO Says a $1.4 Trillion Grid Build Can Cut Power Bills as AMD Tops $1 Trillion

Issue: Tuesday, September 22, 2026

(Covers podcasts published since the last issue: one dated Sunday, September 20 and two dated Monday, September 21. After a weekend that produced a blockbuster ($5 trillion capex numbers, a debt-and-hidden-leases scare, and a full-throated bull rebuttal), the new cycle was a coherent one, with every new episode pointing at the same thing: power, and who pays for it. A big-utility CEO put the electric industry's own five-year investment at "$1.4 trillion" and made the counterintuitive case that data centers can lower your power bill; Amazon and Oracle each spent real money locking up electricity and backup generation years ahead of needing it; and AMD briefly became a trillion-dollar company for the first time.)

TL;DR

The utility side of the AI build finally got a number and a twist. Edison International CEO Pedro Pizarro, who runs one of the country's largest electric utilities, said the industry is "collectively investing something like $1.4 trillion over the next five years," and argued that, done right, "every extra gigawatt" of data-center load can actually cut a utility's average customer rate "by a half to 1.2%," real money against the "data centers will spike your power bill" fear (Squawk on the Street, Sept 21).
Power is now the binding constraint, and the operators are acting like it. Amazon signed a "$2.4 billion" long-term deal for backup generators from Generac, and took warrants to buy up to "$340 million" of Generac stock, sending the shares up "more than 19%"; Oracle signed a "433-megawatt" power purchase agreement and its largest holder, Larry Ellison, called off a "$7.5 billion" stock sale, choosing not to lean on the one funding channel still open to it (Telltales, Sept 20).
AMD touched a trillion; a value investor waved the debt-bubble flag. AMD "briefly [topped] the trillion dollar market cap line for the first time," up about 9% on the day (Squawk on the Street, Sept 21), while RockLinc's Jonathan Wellum warned "we're in a debt bubble," and that the AI boom means "there's going to be overspending… overinvestment… write-offs… and some damage" (Wealthion, Sept 21).

What's new

A quick plain-English key, since a few new terms show up below. A power purchase agreement (PPA) is a long-term contract to buy a fixed amount of electricity, locking in supply and price years ahead. Warrants are the right (not the obligation) to buy a company's stock later at a set price; getting them for free as part of a deal is a way to grab a slice of the other side's upside. A gigawatt (GW) is a billion watts, very roughly the output of one large power plant, or enough to run a big cluster of data centers. Investment-grade is the credit-rating band that says a borrower is considered safe; the lowest rung of it is one notch above "junk," so a downgrade there raises borrowing costs sharply. Free cash flow is the cash a company has left after paying for its capital spending.

1. A utility CEO put a number on the grid's own bill, and argued data centers make power cheaper, not dearer. Squawk on the Street, Sept 21: Pedro Pizarro, CEO of Edison International (parent of Southern California Edison, which serves about 15 million people), interviewed live at Climate Week.

This is the freshest operator insight of the cycle, and it cuts against a fear that's been building in this letter for weeks. Pizarro framed the electric industry's response to AI load as an enormous capital program in its own right: "my colleagues and I are collectively investing something like $1.4 trillion over the next five years." That is the other half of the AI capex story (the wires, substations and generation that the chips and data centers plug into), and it rarely gets a headline number.

The genuinely counterintuitive part is his math on rates. The political worry (loud enough that Pizarro noted "data center pushback gaining some steam in California with towns and residents worried about energy costs and pollution") is that AI will drive up everyone's power bill. His rebuttal: spread over a big enough customer base, new load lowers the average cost of running the system. "We estimate that done right, an extra gigawatt of load in our area… every extra gigawatt can actually reduce our system average rate by a half to 1.2%. That could mean well over a billion dollars in customer benefit over the lifetime of that investment." The catch is in "done right": he backs a "Ratepayer Protection Act" whose "core… is making sure that… when a data center comes in, it's welcomed… but it also bears the cost that it's creating on the system." Notably, Southern California itself is an inference hub, not a training hub ("we don't have training centers, but we still have [inference] centers"), meaning steadier, always-on load rather than the spikier demand of model training.

He was also blunt about the risk sitting on his own balance sheet, and it's a real one for anyone owning regulated utilities in fire country. After California "failed to pass any legislation that would cap [its] liabilities for future wildfires," Pizarro conceded that "Southern California Edison and EIX are at the lowest possible investment grade credit ratings," and "Fitch last week just changed our outlook to negative." His point is that this isn't only a shareholder problem: "that is driving the cost of debt up… and the cost of debt is something that is passed right through to our customers," potentially "hundreds of millions of dollars… in cost exposure." He nonetheless "recommitted… to our 5% to 7% EPS annual growth rate through 2030."

Why it matters: this reframes the utilities as an AI-capex beneficiary with a quantifiable rate tailwind, not just a bottleneck, which is bullish for regulated names that can add large load cleanly. But it also flags the specific tail risk (wildfire liability + a credit downgrade path) that makes the California utilities the risky way to play it. The actionable split: own the load-growth story where the liability tail is benign; treat EIX's own equity as the higher-risk expression, with Fitch's negative outlook as the thing to watch.

2. Amazon and Oracle each spent real money on power this week, the clearest sign yet that electricity, not chips, is the constraint. Telltales, Sept 20: an analytical markets round-up (note: this show is produced with AI-generated voices; treat it as commentary, not operator testimony, though its facts are specific and checkable).

Two deals, one message. First, Amazon–Generac: Amazon "signed a $2.4 billion long-term supply agreement for backup generators at its data centers," with "first delivery… expected in 2027 and 2028." The clever bit is what Amazon took alongside the order: "warrants, the right to buy up to $340 million of Generac stock." Generac's stock "ran more than 19%." The hosts' read is worth keeping: Generac "generat[es] about $600 million of free cash flow a year," so "$2.4 billion… is roughly four years of everything the company makes landing in one contract," and the warrant is "Amazon telling you it knows exactly how big that order is." The caution: Generac already trades at "29 times free cash flow," and it still "has to build capacity for an order that size… nobody has shown us what that costs yet." In plain terms, Amazon just secured years of on-site backup power and grabbed a piece of the supplier's upside: the behavior of a buyer that expects grid power to be scarce.

Second, Oracle's funding, told three ways in one week. Oracle printed on September 11: "revenue about $19 billion, up 30%" led by triple-digit cloud-infrastructure growth. What happened after the print is the story: "On Sunday, Larry Ellison called off a plan to sell up to 50 million shares, call it $7.5 billion. On Thursday, Oracle signed a 433-megawatt power purchase agreement with RWE. And on the same Thursday, the House Veterans Affairs Committee subpoenaed Ellison and [Oracle president Mike Sicilia] over the $27 billion VA health records contract." The hosts tie them together: all three are about "who pays for the rest of this build, and on what terms." Oracle "funded a big piece of last quarter with equity, not the bond market," and already carries "$126 billion of net debt… against a $561 billion enterprise value" ("roughly a fifth of what you're buying here is already borrowed"), with capital spending "running about $76 billion over the trailing 12 months." Ellison not selling into the equity channel says "the people closest to it are not leaning on it"; the 433-MW PPA is a company "lock[ing] up electricity years ahead of needing it," which "companies about to slow down do not" do. The wobble: the "$27 billion" VA contract "with a political counterparty" that "could actually move the numbers" against a revenue line "running near 46% growth"; the hosts put "60-40 that the growth holds through the next two prints."

Why it matters: strip away the personalities and you get the same signal twice: the giants are paying up, and structuring creatively, to secure power and backup capacity for 2027-2028. That is bullish for the constrained power/generation complex (Generac now joins Vertiv on the list) and confirms the weekend's funding thesis in miniature: the build is real, it leans on equity and creative contracts more than plain-vanilla bonds, and insiders aren't dumping stock into it.

3. AMD joined the trillion-dollar club for the first time; a value investor planted the flag on the other side. Squawk on the Street, Sept 21; Wealthion, Sept 21.

AMD finally moved. On Monday, CNBC flagged "shares of AMD seeing a big pop today and briefly topping the trillion dollar market cap line for the first time," trading just below that level "with a 9% rally," and noted the options market had been "pil[ing] into" AMD and Intel calls the week before. There was no new product disclosure attached (this was a market/positioning move, not a Helios or MI450X milestone), but crossing $1 trillion is a psychological marker for the only credible number-two in AI accelerators, and it says the "second-source to Nvidia" trade is being re-rated.

On the other side, RockLinc's Jonathan Wellum (a Canadian value investor, so a useful contrast to the momentum crowd) made the plain-spoken bear case. He's not calling a top on AI's usefulness ("we're going to have a real revolution in AI"), but on the money: "there's trillions of dollars being spent… it's very difficult to know the valuations and how much money is going to be made, what the margins are going to be, what the return on invested capital is going to be." His conclusion is the classic late-boom warning: "there's going to be overspending… overinvestment… write-offs… and some damage," and doing that "in the midst of interest rates going up and inflation running hotter… just puts more pressure on the whole system." With the 10-year "close to 5%" and the 30-year "5.3, 5.4%," his portfolio prescription is to "be careful about valuations," favor hard assets and companies with pricing power, and keep "a little powder dry." (Worth flagging: this interview was recorded on September 11 and published Sept 21, so it's a considered view rather than a reaction to this week's tape.)

Why it matters: AMD at a trillion is the clearest "risk-on" tell of the cycle: the market is broadening the AI-hardware trade beyond Nvidia. Wellum's warning is the standing counterweight: none of the ROI questions have been answered, and a higher-for-longer rate path is the mechanism that turns "overspending" into "write-offs." Both can be true; the tie-breaker is whether the revenue shows up before the rate pressure does.

The debate

Steel-manning both sides of the $700B-plus 2026 hyperscaler capex thesis. This cycle didn't move the headline capex number (the weekend's Catz "$5 trillion, 2025-2030" and "$830 billion this year → $1.1 trillion next" still stands as the anchor). What it did was sharpen the power leg of the argument, and power is where bull and bear now actually disagree.

Bull: the constraint is real, the operators are pre-committing capital to solve it, and it can even be win-win. The tell isn't rhetoric, it's spending: Amazon locking up "$2.4 billion" of backup generation and taking Generac warrants, Oracle signing a "433-megawatt" PPA and its founder declining to sell "$7.5 billion" of stock into the funding channel (Telltales, Sept 20). Companies planning to digest or pause do not contract power years out. And Pizarro's rate math reframes the political risk as an opportunity: an extra gigawatt of load "can actually reduce our system average rate by a half to 1.2%," turning data centers into a customer benefit if the cost allocation is fair (Squawk on the Street, Sept 21). AMD touching a trillion says the market is willing to broaden, not narrow, the trade.

Bear: the money is getting more expensive, the returns are still unproven, and the write-offs come with a lag. Wellum's frame is the whole bear case in one breath: "trillions… being spent," but nobody yet knows "the return on invested capital," so "overspending… overinvestment… write-offs… and some damage" are the base case of any boom, made worse "in the midst of interest rates going up," with the 10-year near 5% (Wealthion, Sept 21). The power theme cuts the bear way too: Oracle already carries "$126 billion of net debt" (a fifth of its enterprise value) and is funding on equity because the bond market is dearer; Generac has to build the capacity for a four-years-of-cash-flow order at a "29 times" multiple, and "nobody has shown us what that costs yet" (Telltales, Sept 20). And the political friction Pizarro acknowledged ("data center pushback… in California" plus a live Ratepayer Protection Act) is exactly the permitting-and-cost-allocation drag that slows the build (Squawk on the Street, Sept 21).

The synthesis. Both camps now agree power is the gating item, and both agree the operators are pre-committing to it. They diverge on the return: bulls read the pre-commitment as conviction (you don't lock up gigawatts to walk away), bears read it as the moment the overspend becomes contractually locked in, right as rates rise. The utility angle is the fresh wrinkle: it's the rare part of the build with a regulated return and a plausible win-win for customers, which is why the load-growth-plus-fair-cost-allocation utilities look like the cleanest risk/reward this cycle, with wildfire-exposed EIX the pointed exception.

Sell signals to watch:

Utility credit outlooks turning negative. Fitch just moved EIX/Southern California Edison to a negative outlook at the lowest investment-grade rung; more of this across load-growth utilities would say the wildfire/liability and cost-of-debt drag is outrunning the load-growth benefit (Squawk on the Street, Sept 21).
Data-center power pushback hardening into moratoriums. California "pushback… gaining some steam" over "energy costs and pollution," on top of the Texas grid-connection moratorium flagged last issue, is the permitting friction that turns 30%-growth plans into "town meeting" reality (Squawk on the Street, Sept 21).
Hyperscalers leaning harder on equity than bonds to fund the build. Oracle funded "a big piece of last quarter with equity, not the bond market"; if that becomes the pattern across names, it signals the debt market is charging too much, confirming the weekend's "money got more expensive" thesis made visible (Telltales, Sept 20).
Suppliers unable to show the cost of scaling into mega-orders. Generac's "$2.4 billion" order is four years of its cash flow; if it (or its peers) can't fund the capacity build cleanly, the "shovel sellers get paid" thesis has a margin problem hiding in it (Telltales, Sept 20).

Stocks in play

NVDA. Bull: still the default long; the AMD re-rating below is a read-through that the whole AI-accelerator complex is being valued higher, and Nvidia remains the volume leader. Bear: the rate/ROI worries in "The debate" hit the priciest names first. Next catalyst: the next hyperscaler capex guidance against the weekend's escalated "$830 billion this year → $1.1 trillion next" path; custom-silicon share.

AVGO. Bull: the custom-ASIC arms dealer that gets paid regardless of which model wins. Bear: sits inside the same rate-and-ROI math as the rest. Next catalyst: the next hyperscaler custom-ASIC order update.

AMD. Bull: the market just crowned it, with shares "briefly topping the trillion dollar market cap line for the first time," up ~9%, with heavy call buying, as investors broaden the AI-hardware trade beyond Nvidia (Squawk on the Street, Sept 21). Bear: the pop came with no new product disclosure (it's positioning and momentum, not a fresh MI450X/Helios milestone), so it's the higher-beta name most exposed to a rate shock or air pocket. Next catalyst: a company-issued MI450X/Helios deployment update and a named frontier anchor customer.

MSFT. Bull: a core "build-it-to-rent-it" landlord with the balance sheet to carry the debt. Bear: last issue's depreciation stretch (server life 15→25 years) still flatters EPS as capex peaks. Next catalyst: whether peers copy the depreciation change; Azure growth against the capex line.

GOOGL. Bull: a cash-machine landlord that "owns the stack," and it just dodged the worst antitrust outcome: a federal judge "declined to order a breakup" of its ad-tech business (Telltales, Sept 20). Bear: the remedy still bites: Google must appoint an antitrust compliance officer for six years, "expose pre-bid equivalent interfaces… to rival publisher ad servers," share bid data, and end AdWords direct bidding into its own ad server, with "60 days to comply" against a rival's estimate of "12 to 15 months to build the thing that complies" (Telltales, Sept 20). Next catalyst: the ad-tech remedy's implementation timeline; free-cash-flow trajectory against a rising capex line.

AMZN. Bull: AWS is the offtake engine, and Amazon is behaving like it believes its own demand: a "$2.4 billion" backup-generator deal with Generac plus warrants for "$340 million" of stock, securing power for 2027-2028 deliveries (Telltales, Sept 20). Bear: the heaviest absolute capex to justify (last issue's "$220 billion" full-year plan, group free cash flow already negative). Next catalyst: proof AWS demand is broad beyond the frontier labs.

META. Bull: an ad engine that insulates holders from a factory write-down; the lightest capex of the big four on the weekend's Catz math ("$140 billion"). Bear: still the poster child for the off-balance-sheet build (the "$27 billion" Hyperion JV bond from last issue). Next catalyst: whether ad growth keeps outrunning the capex line; how much more of the build moves off the books.

Also in play: GNRC (Generac): up "more than 19%" on the "$2.4 billion" Amazon backup-generator deal and warrant; the read-through is that on-site power/backup is now a contracted infrastructure category, not a storm-demand business, though it trades at "29 times free cash flow" and has to fund the capacity build (Telltales, Sept 20). ORCL (Oracle): funding via equity not bonds, "$126 billion" net debt on a "$561 billion" EV, a new "433-MW" RWE PPA, Ellison pulling a "$7.5 billion" stock sale, and a "$27 billion" VA-contract subpoena as the wildcard (Telltales, Sept 20). EIX (Edison International): reaffirmed "5% to 7% EPS growth through 2030" and a data-center load tailwind, but at "the lowest possible investment grade" rating with a fresh Fitch negative outlook after California failed to cap wildfire liability (Squawk on the Street, Sept 21). MU (Micron): the next hard catalyst, as it "reports Wednesday the 30th, about $51 billion consensus revenue," trading at "24 times free cash flow" going in (Telltales, Sept 20).

Read-throughs

Power / grid: this cycle's whole story, and it's now a stock-picking problem, not a macro theme. The industry is spending "$1.4 trillion over five years" on the grid, and the operators (Amazon, Oracle) are contracting power and backup years ahead, which is bullish for the constrained-equipment complex, where Generac now joins Vertiv as a named beneficiary. The actionable nuance is Pizarro's rate math: load growth is a genuine tailwind for regulated utilities where the wildfire/liability tail is benign, which is exactly why EIX is the wrong expression of a right idea: you want the load growth without the "lowest investment grade + negative outlook" credit overhang (Squawk on the Street, Sept 21; Telltales, Sept 20).
Utilities / nuclear: a regulated return is the rare clean way to own the build. Oracle's "433-megawatt" PPA with RWE is the concrete power-procurement datapoint this cycle; the direction of travel (giants signing multi-hundred-MW contracts) is the same setup that has driven nuclear PPA names like Vistra, Constellation and Talen. Watch for the next hyperscaler PPA as the tell on how tight generation really is (Telltales, Sept 20).
Memory / networking: the calendar is the catalyst. The event that matters is Micron's print on September 30 (~"$51 billion" consensus, "24 times free cash flow" going in) is the next read on whether the memory tightness the bulls love is holding or, as the Hedgeye bear argued last issue, has "passed the peak" (Telltales, Sept 20).
Rates / credit: the plumbing keeps confirming the weekend's thesis. Wellum's "debt bubble" with the 10-year "close to 5%" reinforces rather than advances the story. The concrete tell is Oracle funding on equity instead of bonds while its founder declines to sell: the clearest sign the debt channel is dearer than the headlines admit, exactly the point last week's credit voices were making (Wealthion, Sept 21; Telltales, Sept 20).

What changed vs last issue

Last issue (Monday, Sept 21, "Catz says $5 trillion. The debt's off the books.") swept the weekend blockbuster: Oracle CEO Safra Catz's "$5 trillion, 2025-2030" capex figure, the railroad-boom debt analogy from a financial historian and a $1.2-trillion credit investor, the Tom Lee / Dan Ives bull rebuttal ("13-to-1" demand), Dan Niles's "speed bump, not top," and Microsoft's 15-to-25-year depreciation stretch. This edition covers the Sunday-and-Monday gap (Sept 20-21), a cycle whose common thread was power and funding.

The grid got its own capex number, and a bullish rate twist. New this cycle: Edison's Pizarro put the electric industry's investment at "$1.4 trillion over the next five years" and argued each extra gigawatt of load can cut customer rates "by a half to 1.2%," the first operator framing of data centers as a rate benefit, not just a cost (Squawk on the Street, Sept 21).
The weekend's funding thesis got concrete corporate examples. Last issue was about the system (bonds, hidden leases). New this cycle: specific moves: Amazon's "$2.4 billion" Generac deal with warrants, Oracle's equity-not-bonds funding, Ellison pulling a "$7.5 billion" sale, and a "433-MW" PPA (Telltales, Sept 20).
AMD finally showed a signal. This cycle it "briefly [topped] the trillion dollar market cap line for the first time," +9% (Squawk on the Street, Sept 21).
A new bear voice, but no new bear numbers. RockLinc's Wellum added a value-investor's "debt bubble / overspending / write-offs" warning, reinforcing (not advancing) the weekend's rate-and-ROI case (Wealthion, Sept 21).
A fresh GOOGL datapoint. A judge declined to break up Google's ad-tech business but imposed sharp behavioral remedies on a "60 days to comply" clock, a modest positive on the tail risk, with an implementation headache attached (Telltales, Sept 20).
Still open: whether utility credit outlooks keep deteriorating; whether data-center pushback hardens into moratoriums; whether hyperscalers keep funding on equity over bonds; and Micron's print as the memory verdict.

Next catalysts to watch: Micron earnings (Wed, Sept 30, ~$51B consensus) as the memory read; Aurora Innovation's investor day this week (a soft autonomy/AI-adjacent read); the path of the 10-year Treasury; the next hyperscaler PPA or power deal as the tell on how tight generation is; and any move on EIX's credit outlook or California's wildfire-liability framework.