Newsletter · · Ashutosh Agarwal

Amazon Locks In More Nuclear as Data-Center Backlash Spreads - Powering AI Infrastructure - Week of October 2, 2026

The powering-AI-infrastructure weekly for the week of October 2, 2026, synthesizing operator and investor podcasts on the power buildout. Amazon signed a 20-year deal for a 190 MW nuclear uprate at Constellation's Calvert Cliffs, local opposition blocked or delayed about 200 billion dollars of data-center projects in the first half of 2026, and uranium contract terms firmed even as the stocks sit well off their highs.

Powering AI Infrastructure

Week of October 2, 2026: Amazon Locks In More Nuclear as Data-Center Backlash Spreads

On Thursday, Amazon ($AMZN) signed up for more nuclear power. This time it is not a new reactor or a restart. It is a 190-megawatt "uprate", meaning more output squeezed from a plant that already exists: Constellation's ($CEG) Calvert Cliffs in Maryland.

That same week, a different set of podcasts kept a running count of data centers that never got built. Local opposition blocked or delayed $130 billion of projects in the first quarter alone.

So that is this week. Buyers are paying more than ever for power that already exists or can be added quickly, while the new projects run into permits, pipelines and angry neighbors. My view is that this split, more than any one turbine order or uranium price, is the story to watch into 2027.

TL;DR

  • Existing power is the scarce asset. The Amazon–Constellation deal adds 190 MW at Calvert Cliffs under a 20-year contract and unlocks more than $3B of investment. A podcast analyst estimates data centers would keep running even with power at $5,000+ per megawatt-hour. Owners of plants already on the grid hold the pricing power.

  • Getting new projects built is the bottleneck. About $200B of data-center projects were blocked or delayed in the first half of 2026. Oracle's flagship site is waiting on 17 miles of gas pipe. A Schneider Electric executive says 150 GW of planned data-center power is meant to run fully off the grid.

  • In uranium, contract terms matter more than the spot price. Spot sits around $89.50. Long-term contracts now carry ceilings of $140–160 a pound, and some have no ceiling at all. Uranium stocks, meanwhile, are 50–75% off their highs.

What's new

1. Amazon adds nuclear without a new reactor ($CEG, $AMZN)

On Schwab Network's opening-bell podcast, "10-Year Yield Tap 24-Year Highs; MU Earnings, AMZN & CEG Deal Power AI Stocks" (Oct 1), anchor Kevin Hanks walked through the deal:

  • It is a 20-year power purchase agreement (a PPA: a long-term contract to buy a plant's electricity at agreed terms).

  • It covers 190 megawatts of added nuclear capacity at Calvert Cliffs.

  • It "unlocks over $3 billion in investments."

  • Both stocks traded higher on the news.

Why it matters: this is the third kind of hyperscaler nuclear deal, after restarts (Crane/TMI) and new small reactors. Uprates are the fastest of the three. They use a site that already has a permit, a grid connection and a workforce, so there is nothing to fight over in a zoning meeting. For anyone who owns a nuclear fleet, every reactor that can be uprated is now an asset Big Tech wants to buy.

2. Data centers would pay almost any price for power, and that changes power markets

The sharpest idea this week came from Hans Royal of RP Energy, a market analyst, on NPM Interconnections, Episode 214 (Oct 1). Royal built a measure he calls the "compute heat rate." It is the electricity price at which turning power into AI output stops making economic sense.

His numbers:

  • Build threshold: about $5,631 per megawatt-hour. That is the power price at which "it not to make sense economically to convert electrons into intelligence anymore." US wholesale power typically trades at a small fraction of that.

  • Dispatch threshold: over $9,000/MWh. This is the break-even for a data center that is already built, once the construction cost is sunk.

  • Frontier models: $18,000+/MWh. Even the lowest-value AI workloads stay above $1,000/MWh.

  • The measure fell about 30% from Q2 to Q3 as AI model prices compete down. Royal still calls that "not a collapse," because it remains "100 times plus higher than the average wholesale price of power."

The part that matters for markets is what happens during a crunch:

"A lot of these workloads will actually not curtail when power prices go to 200, 300, $500 per megawatt hour, up to $1,000 plus per megawatt hour."

In the past, factories cut back when prices spiked, and that helped cap the spikes. Data centers will not do that. As they grow as a share of demand, peaker plants (gas units that only switch on when demand is highest) will set prices for more hours of the year. Royal says "we're starting to see these impacts clearly in PJM and in Dominion Hub especially." PJM cited his framework in its May reliability white paper, which argued that current price caps are too low to get data centers to cut usage voluntarily.

Why it matters: this is the strongest case I have heard for merchant generators like Vistra ($VST), Constellation ($CEG) and Talen ($TLN), the companies that sell power at market prices instead of regulated rates. A fast-growing customer that does not cut back when prices rise means more high-price hours each year.

3. 150 gigawatts planned off the grid

Greg Castle, who runs Schneider Electric's North American Power and Grid business, gave the operator's view on Becker's Healthcare Podcast, "Preparing Healthcare Infrastructure for AI and Rising Power Demand" (Sep 29):

  • 487 gigawatts of data-center projects have been announced over the past three years, all planned for the next five.

  • About 150 GW of that is planned to run fully off-grid ("islanded"), because of what he calls "the long five to seven year plus interconnect challenge."

  • That 150 GW is "the equivalent of France and the U.K. combined", or "25 percent of the entire U.S. peak demand today."

  • The same pressure is reaching hospitals. A 500-bed hospital uses 8–10 MW. Running AI across it means a 1–2 MW on-site data center, or 10–25% more energy. Castle's advice: plan for on-site generation, because "you may not have that luxury" of asking the utility for another megawatt.

Castle added that "even if half of that actually comes to fruition, it's quite astonishing."

Why it matters: 150 GW of self-built power would mean a lot of gas turbines, reciprocating engines, fuel cells and batteries. That points to Caterpillar ($CAT), Cummins ($CMI), Bloom ($BE) and the switchgear makers. The open question is how much of that power ever runs. Castle's hedge, "even if half," is doing a lot of work in that sentence.

4. Canada's largest gas producer bets on behind-the-meter power

On NGI's Hub & Flow, "LNG, Oilsands, AI: Canada's Top Natural Gas Producer Assesses What Comes Next" (Oct 1), Tourmaline's Jamie explained why Alberta may do data centers better than the US:

  • US backlash is worst where data centers were connected "on-grid first," competing with homes for power. He pointed to PJM capacity auctions that have "skyrocketed" to around $300.

  • Alberta had spare capacity, with power "under $40 a megawatt hour Canadian for much of this year." It gave 1.2 GW of its 2 GW surplus to data centers. Greenlight Energy Center has made its final investment decision, and the Borealis Energy Center (the former TransAlta Keephills site) is moving forward.

  • Every site after that has to bring behind-the-meter power, meaning generation built on site, away from homes. Tourmaline is developing its own at Malachite 1, south of Edson.

  • His estimate: on-grid and behind-the-meter sites together could reach "roughly a billion cubic feet a day of consumption in Alberta by the end of the decade."

He also gave a clear US gas call. As Permian pipeline capacity fills in, demand rises "through the back end of 27 into 29 with 28 being the real year of tightness." He expects "a pretty meaningful re-rate in NYMEX pricing" for 2028.

Why it matters: this is a producer, not a pundit, putting a number on data-center gas demand in one province. For scale, the US-wide figure people usually cite is about 6 Bcf/d by 2030. It also suggests a fix for the backlash problem: build new generation on site, far from homes, and sell surplus power back to the grid.

5. In uranium, the contracts look stronger than the stocks

Justin Huhn of Uranium Insider appeared on two podcasts on Oct 1. He is a newsletter writer, not an operator, but he follows contract terms closely.

On Palisades Gold Radio, "The Uranium Blow-Off Top, AI Hyper-Scalers & 'Huge Value' in The Miners":

  • Long-term contracts from Cameco ($CCJ), Kazatomprom and Orano for late-2020s and early-2030s delivery carry "floors in the 70s, ceilings at $140, $150, sometimes we're hearing $160." Some are being negotiated "with floors and no ceilings." (The floor protects the miner if prices crash. The ceiling caps what the buyer can be charged.)

  • The new buyer: hyperscaler representatives have been "poking around the fuel cycle," and NextGen is getting inbound interest. "We don't think it's far-fetched to see somebody like Amazon, somebody like Meta, actually signing a long-term uranium offtake."

  • His call: "We expect the price to go well north of $200 a pound."

  • On small modular reactors, he says GE Vernova's ($GEV) BWRX-300 leads. The NRC had just approved early site permitting for TVA's first US unit. Hyperscaler representatives visit the Darlington site in Ontario every week. A GE Vernova representative said that once the first unit is built, "we're going to have hundreds in the order book."

  • The odd part is that Huhn's own supply-and-demand model assumes zero SMR uranium demand until 2031. Hyperscaler demand for SMRs is upside he has not counted.

  • On the stocks: some "are off 50% to 75% from their highs of last year," including suppliers like BWX Technologies ($BWXT).

On In it to Win it, "Justin Huhn Says Uranium Supply Jumped 50% And The Price Still Doubled", he made the supply argument:

"Supply has increased by almost 50%. Price has doubled. That's all you really need to know."

  • Mine supply went from about 115 to 175 million pounds a year, through restarts at McArthur River, Langer Heinrich and Honeymoon, more Kazakh and Uzbek output, and 3.5–4 Mlb from the US.

  • Spot is "89 50 kind of mid market." Below $88, utilities step in to buy.

  • US utilities hold "just over two years" of contracted fuel, against 3.5–4 years in the EU.

  • NextGen's Rook I could produce up to 29 Mlb a year starting 2031, but Huhn expects real output in 2032. He argues it will not depress prices because it is sold through long-term contracts, not into the spot market.

A natural-resources fund manager on Barron's Live, "The Outlook for Energy and Natural Resources Stocks" (Sep 29) supported the price argument. Uranium is now "essentially $100 a pound," but the price needed to justify a new mine has risen to "probably $150 or $160," so "I don't even think we're at the point where you could bring on new capacity yet."

Why it matters: term prices, contract structures and the incentive price for new mines all point up, yet the stocks are pricing in "the present market." Investors who trust the contracts and doubt the stock prices have something to work with.

The bigger risk: projects that never get built

This week the bear case was not that demand is fading. It was that projects cannot get built.

The data on backlash. On InvestTalk, "Is the Data Center Backlash a Real Investment Risk for AI Infrastructure?" (Oct 2), host Luke Guerrero cited tracker Data Center Watch and polling:

  • Q1 2026: local opposition blocked or delayed 75 projects worth $130 billion, as many projects as in all of 2025.

  • Q2 2026: 45 more across 27 states, worth $68 billion.

  • 20 projects cancelled outright, worth nearly $42 billion.

  • A March Gallup poll found 70% of Americans oppose a data center where they live, and 55% strongly oppose one. 142 organized groups in 24 states are fighting new construction, and 30 states have introduced legislation.

"What we're starting to see is this physical expansion hitting a wall of this organized opposition that is not getting discouraged because they've seen success. They're in fact getting empowered."

Oracle's 17 miles of pipe. On Morning Brew Daily (Sep 25), Toby Howell and Neil Freiman explained what is holding up Oracle's ($ORCL) $165 billion Project Jupiter in New Mexico. Energy Transfer needs to build a 17-mile gas pipeline to feed it, and the pipeline is "stuck in permitting purgatory." Pipeline construction slipped from August to February 2027, which puts the 2028 opening at risk. Oracle has $18 billion of debt tied to the project, now trading at 89–91 cents on the dollar. Howell summed it up: "hundreds of billions of dollars of cutting edge computing waiting on 17 miles of pipe."

Who pays for the grid. On Factor This (Sep 28), Chris Root, the former COO of Vermont transmission company VELCO, laid out the timing problem. A single 300 MW data center would be "the biggest load in all of New England." It can be built in 24 months, takes a year to study, and the transmission to serve it "could take seven years." Utilities have a legal obligation to serve, so they "can't say no." On cost: in Virginia, grid costs were passed to customers, and "I don't think that would fly now… you're going to have to allocate it back to those large customers." Co-guest Damir Novosel noted that NERC (the grid reliability regulator) is moving to make data centers registered entities. That would make large loads directly responsible for grid reliability.

How it gets paid for. Two finance podcasts looked at the money behind the buildout:

  • On Unhedged, "Hyperscale my debt!" (Oct 1), the FT's Emily Herbert said investors have lent AI-related companies about $500 billion this year, $200 billion of it to hyperscalers. The FT expects about $1 trillion of hyperscaler debt by 2030. The extra interest these companies pay over government bonds has "really blown out," and Herbert says that is not about credit quality: "every quarter these companies come to the market and they revise up the number for how much they expect to spend."

  • On The Intrinsic Value Podcast (Sep 27), Shawn O'Malley cited Wall Street Journal reporting. Meta ($META) carries close to $700 billion in off-balance-sheet commitments (leases not yet started plus purchase commitments, energy included), and Alphabet about $1 trillion. Both are up about 800% in a year. The model is Meta's Louisiana campus: a special-purpose company owned roughly 80/20 by Blue Owl ($OWL) and Meta issued about $27 billion of bonds, with Meta signing a long-term lease. Co-host Daniel Mahncke's point was that this is "how pipelines, how power plants and telecom networks have been funded for decades." The hyperscalers now borrow like utilities.

The debate

Bull: the whole power chain re-rates together.

  • Demand is real and long-lasting. A board member of the Electric Power Research Institute (EPRI), the utilities' research arm, said on Climate Rising (Sep 30) that US demand is now projected to grow 3–5% a year this decade, "a rate we haven't seen since the 1970s." Utility CEOs who used to treat 50 MW as a big hookup now get requests for 5,000 MW.

  • BlackRock's guest on The Bid (Oct 2) estimates ~121 GW of data-center power needed by end of decade, at $40–50 billion of investment per gigawatt.

  • Data centers will pay far above today's prices (Royal's $5,000+ break-even), and they want speed more than anything. The EPRI board member put it this way: "If we make them choose between speed to power and clean, they'll all choose speed to power."

  • Gas demand from AI is already showing up in the data. The Barron's Live fund manager: "Every cold day, you're withdrawing even more than you would expect. So it's not weather alone." He called $3 gas "far too cheap."

  • Equipment is sold out. On Frictionless (Sep 25), P Equity Research said a gas turbine ordered today arrives around 2030, and GE Vernova's backlog extends beyond that.

Bear: plenty of demand, but too much financing, opposition and supply coming.

  • Financing depends on a few companies. The $1 trillion of debt, the $700 billion of off-balance-sheet commitments and Oracle's bonds at 89 cents all rest on hyperscaler capex plans that, as Unhedged noted, keep getting revised up. O'Malley's warning: in two years "we'll be talking about how there's so much more supply than demand."

  • Opposition is growing and winning. About $200 billion was blocked or delayed in six months.

  • The LNG side looks oversupplied. On Columbia Energy Exchange (Sep 29), Ira Joseph cited Total's forecast of LNG supply +50% from 2026 to 2029 and +75% by 2031, and called coal plus renewables with batteries "a really, really bad story for LNG demand growth on the power side." If prices fall far enough, US exporters, as the marginal supplier, are the ones who get shut in.

  • Batteries are crashing in Texas. On Catalyst with Shayle Kann (Oct 1), Kann said ERCOT went from 200 MW to 16.5 GW of grid batteries this decade. The average battery earned 84% less in 2025 than in 2023, and 13.7 GW withdrew from the queue in the first half of 2026. Modo Energy's Brant Vermillion does not see price swings returning in a big way until 2029–2030. That is the bear case in miniature: a market that will need a lot more capacity eventually, overbuilt in the meantime.

  • Copper deficit charts have a poor record. A mining analyst on Rock Stock Channel (Sep 25) said the supply-gap charts "have been wrong every year since 2014 to 2025," and exchange inventories are "still at quite elevated levels."

My read: both sides are right about different time frames. Through 2027 the bears have the stronger case on project completions. Some announced capacity will be delayed or cancelled, and the companies that raised capital to build it will feel that. On power prices, though, the bulls look better positioned. Every cancelled project means less new supply, while the data centers that do connect barely cut usage when prices spike. That favors whoever already owns generation and grid connections.

Batteries: the quiet fix for flexibility

Last week the question was whether data centers would cut usage when the grid is stressed. This week's answer: mostly not. Batteries are increasingly doing that job instead.

  • On the Energy Capital Podcast (Sep 30), a Texas home-battery developer cited research putting the cost of curtailing a data center at "upwards of $40,000 a megawatt hour, in some cases, 6 digits." His alternative: a 100 MW data center at the Burleson Switch substation near Dallas can connect safely with only 80 MW of home batteries placed near the strained lines. The batteries can be built "in less than 6 months," while new 765-kV lines will take "10 years at best." The data center pays, and nearby homes get backup power.

  • On The Data Center Frontier Show (Sep 29), ON.energy's VP of Engineering Asser Elsamahy said large battery systems at data centers can cut a 1 GW load to about 800 MW at peak. He is selling to hyperscalers. Grid operators, ERCOT in particular, are tightening rules on how fast these loads can ramp and how they handle voltage dips, and he expects those rules "will only get stricter."

  • Vermillion of Modo adds a twist: as grid hookups get scarcer later this decade, data centers will build more of their own generation. His rule of thumb is that only 50–70% of data-center nameplate ends up as actual grid demand.

Why it matters: batteries placed at data centers or near the grid are becoming the way large loads get connected sooner. That is a new kind of demand for storage, separate from the merchant battery trading that is losing money in ERCOT right now. Grid-side electrical equipment suppliers such as Eaton ($ETN) and Vertiv ($VRT) are positioned to benefit as large-load connection rules tighten.

Read-throughs

  • Gensets and fuel cells ($CAT, $CMI, $BE): Schneider's 150 GW off-grid figure and Tourmaline's behind-the-meter plans are direct demand for on-site generation. Discount for completion risk, since Castle himself only counted "half."

  • Gas producers and midstream ($WMB, $KMI, $OKE): Tourmaline's call that 2028 is the tight year and the Barron's Live comment on storage draws both point to rising domestic gas demand. Oracle's stalled Energy Transfer lateral is the risk: pipelines feeding data centers face the same permitting fights as the data centers.

  • LNG ($LNG): US exporters are earning $40–50 million per cargo on a $17 spread between US and overseas prices, according to Ira Joseph on Columbia Energy Exchange. The Total supply wave is the threat to that spread after 2028.

  • Uranium and enrichment ($CCJ, $LEU): Centrus has a $4.5 billion backlog and revenue of $176 million, up 14%. The stock is still down 57% over 52 weeks and trades at about 50x forward earnings, per InvestTalk. Strong contracts, falling stock prices: the same split Huhn describes across the sector.

  • Copper ($FCX): On The KE Report (Oct 2), Sean Brodrick said Amazon, Google and other hyperscalers are "going around, knocking on the doors of copper companies saying, how much copper can you give us?" He expects them to help fund new mines through offtake deals. Set that against the inventory skeptic above.

  • Hyperscaler balance sheets ($META, $GOOGL, $AMZN, $ORCL): they now borrow like utilities, through bonds in multiple currencies and project-level special-purpose companies. Wider credit spreads are an early warning for every supplier further down the chain.

What changed since Tuesday

  • Oracle: last issue covered the force majeure notice (a contract clause excusing delays outside the company's control). This week brought the details: a 17-mile Energy Transfer pipeline, construction pushed to February 2027, and project debt at 89–91 cents. It has turned from a headline into a case study in how one permit can delay a multibillion-dollar project.

  • On-site power got bigger: last issue it was Bloom's grid cost figures and PowerSecure's 35–40 GW of distributed generation. This week Schneider put the planned off-grid total at 150 GW.

  • Uranium shifted from supply to contracts: last issue, Sprott's 175 vs. 200 Mlb supply gap. This week the focus moved to contract terms (ceilings of $140–160, some with none), spot around $89.50, and stocks at 52-week lows.

  • Flexibility: last issue, Oracle and Duke's 0.5% flexibility idea suggested about 100 GW of headroom. This week, compute heat rate economics and curtailment costs of $40,000+/MWh suggest the servers themselves will rarely be what flexes. Batteries will do more of it.

  • New this week: nuclear uprates as a hyperscaler deal type (Calvert Cliffs), a measured count of the backlash ($198B in two quarters), and a gas producer putting a 1 Bcf/d figure on Alberta data-center demand.