Newsletter · 2026-07-28 · Ashutosh Agarwal

Natural Gas Supply Replaces the Grid as the Next AI Bottleneck - Powering AI: Grid, Gas, Generation & Nuclear - Week of July 28, 2026

Power and AI infrastructure newsletter for the week of July 28, 2026. The podcast tape moved the constraint from interconnection queues to natural gas molecules and then to financing, with GE Vernova's record backlog selling off, PJM failing a third capacity auction, and a modeller arguing US gas supply and demand stop matching around 2028.

Powering AI: Grid, Gas, Generation & Nuclear

Week of July 28, 2026: Natural Gas Supply Replaces the Grid as the Next AI Bottleneck

For two years the story of AI and electricity has been about wires: interconnection queues, transformer lead times, substations that take five years to energize. This week the podcasts moved the argument somewhere less comfortable. The wires are still a problem, but the new question is whether there will be enough natural gas to run everything we are busy plugging in, and whether the bond market and the voters will keep paying for it. Three separate conversations, none of them coordinated, landed on the same conclusion from different directions.

TL;DR

• The bottleneck is migrating from wires to molecules. On Invest Like the Best, Matthew Smith laid out a case that US gas supply and demand stop matching around 2028, storage drops below anything on record by 2029, and gas goes to $8-10 or worse, while the forward market still prices it flat in the mid-$3s. Same week, EQT's CEO on Closing Bell Overtime sized incremental power gas demand at 10-18 Bcf/d. Those two people disagree about almost nothing.

• GE Vernova printed the best backlog in the sector and the stock fell 6%. Backlog to $176 billion, free-cash-flow guidance nearly doubled, over half of 2031 turbine orders already under contract by year-end, and it wasn't enough (Squawk on the Street, Schwab Network). When great numbers stop working, that is information.

• Politics arrived. PJM failed its capacity auction for the third consecutive time, and senior White House officials told Bloomberg that breaking up the grid operator is on the table (Crain's Daily Gist). Duke Energy's CEO spent his week at the White House promising your power bill will go down because of data centers (Power Lunch).

What's New

1. The Gas Argument, Made Properly

The single best thing published on this theme all week was Matthew Smith's conversation with Patrick O'Shaughnessy on Invest Like the Best (July 21). It is worth understanding in detail, because it is the rare thesis that is both simple to state and hard to dismiss.

Start with the arithmetic he uses. The US produces roughly 110-112 Bcf/d of natural gas, billion cubic feet per day, the standard unit. About 15 Bcf/d of that currently leaves the country as LNG, or roughly 12-15% of daily supply. Projects already permitted, financed and under construction take export capacity to about 35 Bcf/d by the end of 2030. As Smith put it, "the die has been mostly cast," these are not proposals, they are steel in the ground with project finance attached.

Now the supply side. His firm modelled, in his words, "every producing gas well and the entire pipeline and processing and gathering system," and concluded the US can add about 20 Bcf/d of new production. So before AI enters the picture at all, sources and uses roughly balance: 20 Bcf/d of new gas against 20 Bcf/d of new export demand.

Then add compute. Smith's team assigned probabilities to every announced power project, and his base case counts only those with permits, an offtake contract and an interconnection agreement in hand. That base case is about 5 Bcf/d of credible new gas demand from AI. Loosen the screening and "that number can more than double and be, you know, 12 to 15 Bcf a day by the early 2030s if unmitigated."

There is no obvious release valve. You cannot simply switch off exports: "there's tens of billions of project financing and contracts attached," and the US will be roughly a third of global gas supply within a few years, with allies depending on it. So the pressure shows up in storage. US working gas storage is about 4 trillion cubic feet. Smith expects meaningful draws starting in 2028, and by 2029 "we drop below all known historical storage evidence."

What makes this actionable rather than merely alarming is the gap between his model and the price. The forward curve for 2029 and 2030 is flat in the mid-$3s. Nobody is positioned for this. "Gas has lulled everybody to sleep," he said, and he means it literally: EQT is currently shutting in gas because it thinks the molecules are worth more later, and the rig count shows no sign of the market catching on. Historical shortage episodes (Russia-Ukraine, the 2014 polar vortex, December 2022) took gas to $8, $9, $10. Those were transitory. His point is that this one is structural, and "the deficit gets really convex and unbounded."

His winners and losers list is where it gets uncomfortable for the consensus trade. Winners: Expand Energy, which he says controls roughly 70% of the remaining core Haynesville locations and is "far and away the biggest winner," currently CEO-less after turnover earlier in the year, stock down sharply over six months, trading at four times EBITDA "on a forward curve where no one believes what I'm telling you to be the case." Then Range Resources, which he calls the highest-quality upstream company in Appalachia. And a second-order group most people miss: because gas sets the marginal price of power in most US markets, solar assets with contracts that reset to market get a windfall with zero incremental capital. He named XPLR Infrastructure (ticker XIFR, the former NextEra YieldCo) and Clearway Energy specifically.

The losers are the names everybody owns. "Some of the biggest winners so far, at least in the stock market, have been the manufacturers of gas turbines or distributed power gen sets." He compared it directly to the early 2000s gas-plant boom that ended in a decade of overcapacity, and pointed out that most of them are adding capacity again for 2028-29. Caterpillar, he said, is roughly doubling its solar turbine capacity between now and the end of 2029, "which I would judge is just at the exact wrong time when people may be questioning whether they even want to deploy those assets because the gas is much more expensive than they planned." On Bloom Energy: at 2 GW or more of annual manufacturing, he does not believe the fuel will physically be there. A 6-series Bloom fuel cell consumes about 150 million cubic feet a day of gas per gigawatt, and the market is underwriting a ramp from 2 GW toward 5 GW.

And then the line that should worry anyone modelling hyperscaler margins: energy is currently budgeted at roughly 10% of the cost of compute. If gas doubles or triples structurally, "it could end up being 20% or 30% of the cost of compute by 2029." One of O'Shaughnessy's recent guests, hearing the thesis, told Smith it "sounds like DRAM two years ago, slowly at first and then all at once."

2. The Gas Producer Says the Same Thing, From the Other Side of the Trade

On Closing Bell Overtime (July 21), EQT CEO Toby Rice put numbers on the demand side that are, if anything, larger than Smith's. "Over 10 to 18 BCF a day of more power natural gas is going to be needed to meet this power demand," he said. "When you pair that on top of the LNG demand that's coming, we're looking between 20 and 40 BCF a day of more natural gas demand."

The geographic concentration matters. EQT is tracking more than 45 projects in Appalachia totalling about 20 Bcf/d of potential demand, in a basin that produces roughly 35 Bcf/d today. Rice was careful: "Not all these will happen, but if just a small percentage of these do, it's going to be a tremendous impact." He also claimed EQT has created or participated in over 3.5 Bcf/d of that demand already, including the CPV Shea deal announced that day: over 300 million cubic feet a day feeding a two-gigawatt power plant.

Two other details from that interview are worth filing. EQT raised production guidance above 90 Bcf while lowering capex, the kind of combination that gets a gas producer re-rated if the gas call is right. And Rice offered a striking price comparison for anyone arguing exports raise domestic bills: international gas is "over $19" while US gas is "less than $3." His framing is that the spread proves exports have insulated Americans from geopolitical shocks. The bearish reading of the same fact is that a $16 arbitrage does not stay open forever.

This is an operator, not a pundit, so weight it accordingly, but note that a producer talking his book and an independent modeller with a short list of turbine-maker losers arrived at the same demand number in the same week.

3. GE Vernova Put Up a Superb Quarter and the Stock Fell 6%

On Schwab Network (July 22), Alex Coffey walked the print: backlog up to "a whopping $176 billion," a $13 billion sequential increase from equipment and services demand. Revenue guidance raised a full billion at both ends, to $45.5-46.5 billion. Adjusted free cash flow guidance almost doubled, midpoint from $7 billion to $12 billion. EBITDA margin maintained. Electrification, the highest-growth segment, beat. Even wind, the segment nobody wants, came in better than feared on revenue.

The stock fell about 6%.

On Squawk on the Street (July 22), Seema Mody relayed her interview with CEO Scott Strazik and explained the tension. Strazik said hyperscaler power demand is not slowing, but costs will stay elevated, with an up-to-$200 million tariff hit, and with the Middle East conflict back on he expects "the inflationary environment we're currently living in to continue." The company has already raised prices multiple times in the past year. The open question is how much further it can push.

The one genuinely new disclosure, and the most useful number of the week for anyone modelling this sector: over 50% of GE Vernova's 2031 gas turbine orders will be in contract by year-end. Five years of forward book, half sold, before 2027 begins. Strazik also said nuclear should start mattering meaningfully in about four years, and that he has had productive conversations with Energy Secretary Wright and the Saudis about deploying GE Vernova nuclear technology with US allies.

Melius Research called the guidance raise "very modest." Losses tied to wind remain a drag. And the stock had doubled last year and was still up 84% over twelve months coming into the print. That combination, flawless order book, high bar, muted reaction, is exactly what a late-cycle setup looks like. It does not mean the cycle is over. It does mean the easy money in owning the backlog has been made.

4. PJM Failed a Third Auction, and Washington Is Now Openly Discussing a Breakup

This is the development with the longest tail, and it barely registered outside the trade press. Per Crain's Daily Gist (July 22), citing Bloomberg reporting: a PJM capacity auction, the mechanism by which the largest US grid operator pays generators to guarantee they will be available years from now, failed to secure enough supply commitments for the third time in a row.

The Federal Energy Regulatory Commission convened a one-day conference on PJM's operations. Senior White House officials, speaking anonymously, said all options should be on the table, "including breaking up PJM, letting some states withdraw from it, or restructuring the organization and its governance," and that change is inevitable. The Energy Department and the National Energy Dominance Council are participating, alongside representatives from Pennsylvania and Indiana, plus Google, Constellation Energy and Vistra.

The political mechanics are straightforward and worth stating plainly: rising electricity bills have become a live cost-of-living issue heading into November midterms. The administration wants data centers built to win the AI race with China, and simultaneously wants consumer power bills to stop rising. Pennsylvania Governor Josh Shapiro has said his state will leave PJM's markets absent reform. American Electric Power, one of the largest US utilities, has also threatened to end its membership. White House officials said PJM is losing the confidence of stakeholders, including, pointedly, the data center developers themselves, that it can guarantee future supply.

For anyone with capacity-price exposure in PJM, the market design you underwrote is now a political variable. Nobody on any podcast this week could tell you which way it resolves. That is precisely the point.

5. Duke's CEO Is Now Arguing Data Centers Will Lower Your Bill

The corporate response to that politics showed up the same week. On Power Lunch (July 24), Duke Energy President and CEO Harry Sedaris appeared straight from a White House event on power affordability, and made a specific, falsifiable claim: "In North Carolina, we're estimating $3.6 billion of savings over the data center contracts that we have signed to date over the next 15 years."

The mechanism, in his words: data centers sign fixed-rate 15-year contracts, "and then that revenue is spread over the fixed costs of the system, which lowers the rates to the rest of the customers." It also spreads storm recovery costs and fuel costs across a bigger base. He acknowledged the credibility problem head-on: "the math is complicated. We're trying to simplify that for folks with this pledge so that they can trust that we're doing everything that we can."

Two operational details matter more than the pledge. First, reliability: asked whether a gigawatt data center should be curtailed before residential customers, Sedaris confirmed Duke's data center contracts include the right to reduce their load "for 50 to 100 hours a year at those critical times." The data centers can meet that either by running their own backup generators, with no lost productivity, or by rescheduling work, since "they can run training or inference, which uses different amount of powers at those times." That is load flexibility written into a regulated utility contract, and it is a template others will copy.

Second, the build: Duke is putting up 14 gigawatts over the next five years. Mostly gas, but including solar and batteries "because those are quick resources that can be built and brought online in less than a year," plus uprates to squeeze more from existing nuclear and gas plants. In Florida, 300 megawatts of solar a year, with lower fuel costs feeding a roughly $50 a month residential bill reduction.

North Carolina's governor has asked Duke to make the affordability commitment legally binding. Sedaris did not agree to that, noting the federal pledge was voluntary and that "our regulators hold us accountable," but said Duke is "open to having discussions of how we can make that more clear." Watch that thread. A voluntary pledge that becomes an enforceable rate-case condition is a different animal for a regulated utility's earned return.

6. What On-Site Gas Actually Costs, From the People Who Model It

Bloomberg NEF's Switched On (July 23) is the most useful single episode if you want hard numbers on behind-the-meter power, that is, generation a data center builds on its own site instead of buying from the grid. BNEF analyst Mushfiqa Mishi walked host Tom Rowlands-Rees through her research.

The scale first: BNEF tracks about 141 gigawatts of announced on-site gas generation globally, roughly 127 gigawatts of it in the US, with the honest caveat that "not all of these announcements will pan out." The reason is timing, not preference: "you can't get grid connected until 2030, 2031. Whereas if you want to get a gas turbine or a gas engine or even fuel cells powered by natural gas, you can get that in a year, two years or even three years."

Then the sizing trap, which catches a lot of models. A data center's stated capacity is its IT load. Add cooling and facility operations, about 20% more. Then, because behind-the-meter power lacks the grid's reliability, overbuild another 10-15%. So a "500 megawatt" data center actually needs 660-700 megawatts of generation. Every announced megawatt of IT load implies roughly 1.3-1.4 megawatts of real capacity.

On cost, meaning levelized cost of electricity, the all-in cost per unit of power over the asset's life, gas engines come out cheapest at about $103 per megawatt hour, with solid oxide fuel cells the most expensive at about $140 per megawatt hour, modelled over 30 years at 60% average utilization. That is a wide spread, and it explains a lot of procurement behaviour: engines are cheaper per kilowatt installed ($2,200-$2,400) than fuel cells ($3,000-$4,000), and both follow variable load better than a big heavy-duty turbine, which prefers a steady profile.

The most investable detail was about how turbine makers are now rationing. These are relationship businesses, Mishi explained: "they're not going to prioritize new customers if old customers have already put in orders," meaning existing IPP and utility relationships come before a new data center developer. And there are gates. A developer may have to "put down like 20% of your order in prepayment," or demonstrate permits, land and financing, before a manufacturer will even commit to build a turbine for 2030 delivery. Rowlands-Rees had the right image: a crowded bar where the bartender serves the regulars first.

The strategic question the episode leaves open is the one that should keep turbine bulls honest. Each solution gets used up in turn: grid connection, then heavy-duty turbines, then aeroderivative turbines (essentially repurposed jet engines), then engines and fuel cells. Every wave shortens the lead-time advantage and lengthens the queue behind it. And Iron Mountain's Chris Pennington, speaking on Inside Data Centre (July 24) from Datacloud in Cannes, flagged the commitment problem on the other side: on-site plants of "300, 400, 500 megawatts is not an uncommon size," but the contracts needed to make the economics work run "10, 15, maybe 20 years in length." As he put it, "as a data center operator, we don't naturally go out and invest in power generation plants," and "ultimately, our customers, they all want to be connected to the grid." He hopes behind-the-meter gas is "a relatively short-lived phenomenon." Anyone underwriting a 20-year on-site gas contract as permanent should sit with that sentence.

The Debate

Both sides got voiced properly this week, which is rare. Here is the honest version of each.

The Bull Case: A Decade of Visibility, and It Is Not Priced

The strongest bull articulation came from Baird's Ben Kallo on The Real Eisman Playbook (July 21), Episode 69. Kallo covers the power supply chain and has a buy on GE Vernova, and his framing was that estimates of what the US needs are still wildly unsettled: "the range of estimates could be 100 gigawatts of new capacity needed by 2035, which would up to 350," the top of which he described as more than doubling US generation. Against that, the country is adding roughly 30 gigawatts a year over the next five years. And crucially, this is base load, power available 24/7, which is what data centers actually need, with peak capacity required on top.

The lead times are the moat. When GE Vernova books a turbine order today, "that turbine is not going to actually be built and then put into a utility until 2030, 2031." Gas turbines are about 90% of Power segment equipment sales. And the company is increasing content per project by cross-selling electrification equipment, big transformers and grid interconnection hardware, alongside the turbine, and getting price on both. Kallo noted the CEO describes it as a supercycle with the company sold out to at least 2035, and added: "We've done our work and talked to developers. We have a lot of visibility here for the next decade."

There is a technology tailwind too. NVIDIA's move to 800-volt data center architecture requires more power electronics on both sides of the fence, which is more content for a GE Vernova.

The demand-side evidence supports the shape of this. Limitless (July 22) noted US data center power demand roughly doubling from 31 to 66 gigawatts in 24 months, taking data centers from about 1% to 3% of total US electricity, and GE Vernova's 2025 orders doubling year over year to $7.1 billion, with roughly $7 billion of Microsoft deals and OpenAI as a primary customer. The hosts called GE Vernova "the TSMC of power," which is a stretch, but the pricing power comparison is not absurd. Energy News Beat (July 21) cited Lawrence Berkeley, EPRI and Rhodium work putting data centers at up to 20% of US electricity by 2035 in aggressive scenarios, and flagged the transmission gap starkly: the US needs roughly 5,000 miles of new high-voltage line a year and added 392 miles over the last two years.

And on Super-Spiked (July 25), Arjun Murti's sector screens had LNG names, independent power producers and midstream all sitting in the favourable quadrants on growth and returns, while shale gas producers have badly lagged. If Smith's gas call is right, that relative performance gap is the trade.

The Bear Case: Everyone Is Ordering at the Top

The bears were not shy either, and they attacked from four separate directions.

The demand numbers are fiction. Joshua Rhodes, who models the Texas grid at UT Austin and sits as a utility commissioner for Austin Energy, gave the most quotable version on Renewable Rides (July 21). Data centers have now requested 435 gigawatts of interconnection in Texas, 90% of the queue, on a grid that has never delivered more than 85 gigawatts at once. He called it a bubble in 2025 and the bubble got bigger: "100 years to get to where we are today, and then in five years, we want to 5X it." Why it cannot happen: "everybody in the world who builds power plants would have to come to Texas and do nothing but build power plants in Texas."

Kallo, from the bull side, conceded the same mechanic: a developer permits ten sites, "in the end, it might not be ten data centers. There might be only two. So there is some double counting going on." He also pointed out that as of this week, "the only real gigawatt-scale data center that's been built is by SpaceX." Everything else is in progress.

The cost of building at the top of the cycle never goes away. Rhodes' inflation numbers are the most useful thing published on equipment pricing this week. Against consumer inflation of maybe 10-20% on goods people complain about, "transformers, it's like 200%. Like wires, it's like 180%. Like, you know, switch gears too, you know, two to three X what they were a couple of years ago." That is great for the sellers today. His analogy for what comes next: "when you buy a house, if you buy a house at the top of the market and then the market corrects, your lender doesn't give you a break on the principal down the road. And so if we buy a bunch of infrastructure right now at premium prices, and then things cool down later, we're still going to be paying off the mortgage on that high-priced infrastructure." Ratepayers eat that. Which loops directly back to the PJM politics.

Labour, not equipment, may be the real ceiling. Kallo, again a bull, named it: "labor is going to be an area which is going to become more and more of a bottleneck out there, whether it's for data centers or just energy overall." Asked what kind: "Electricians. Everything." You can raise turbine capacity with capital. You cannot raise the number of licensed electricians with capital, at least not on a five-year view.

The money is getting more expensive, fast. This was the newest strand this week. On The Exchange (July 24), Seema Mody reported that Google, Amazon and Meta are all seeing credit spreads widen, with bond investors demanding more compensation to lend to companies "on track to collectively spend more on CapEx than they generate in free cash flow by next year." The concrete data point: the Financial Times reported bond investors requiring a 7.5% rate to finance Meta's one-gigawatt Texas data center, above what previous projects paid. And Barclays notes Oracle's five-year credit default swap, insurance against default, is no longer trading on company fundamentals but as "a liquid hedge on concerns tied to CapEx spending." The group is on pace for $800 billion of infrastructure spend this year and $1.2 trillion next.

Big Technology (July 24) took the financing question further, with Alex Kantrowitz and Ranjan Roy working through Ed Zitron's research. Bloomberg estimates over $500 billion of outstanding AI data center debt, at least $200 billion of it held by private credit, roughly 8% of all outstanding private credit loans. Nikkei Asia puts Meta, Google, Amazon, Microsoft and Oracle at around $1.65 trillion of debt accrued over five years, plus hundreds of billions off balance sheet. The structural detail that matters: money routed through special purpose vehicles does not show up as capital expenditure. Meta reported $88.6 billion of capex, a figure that excludes the Hyperion SPV's own $46 billion of exposure.

Roy, who worked on a trading floor through 2008, was careful not to overdraw the analogy, "this is going to only affect a class of capital holders," not everyone with a mortgage, but he did not soften the dependency: roughly 70% of Microsoft, Amazon and Google's AI capacity is contracted to OpenAI and Anthropic, two unprofitable companies whose combined revenue is a little over $120 billion against commitments in the hundreds of billions. His verdict on the growth was the sharpest line of the week: "You have unprecedented growth, earth-shattering growth, and it's still not enough."

Steve Eisman, on Squawk Pod (July 27), reduced it to the correlation risk. "It's all one trade. It's literally one." Even a 60/40 portfolio is not diversified, he argued, because more than half the equity side is tech and AI-related and "most of the new issuance of bonds is AI related." He sold his Alphabet position, which he had held for years, to reduce AI exposure, and has not replaced it, because "people either want to buy AI or they don't want to buy AI, but they don't want to shift out of it to buy Clorox." He is sitting in cash and explicitly not short. Asked what happens if AI disappoints: "I think we have a big correction." Asked how big: "that's a hard question. I don't know."

Where We Come Out

The bull case on the physical need for power is close to unassailable, and nobody credible argued against it this week. The bear case is not about whether the electrons are needed. It is about three things: how much of the announced pipeline converts, whether the equipment ordered at 2026 prices earns a return in 2032, and whether the financing chain holds long enough to find out.

Note the asymmetry in who is bullish. The operators (GE Vernova, EQT, Duke) are telling you demand is real and their books are full, and they are right. The people modelling the system rather than a single company (Smith on gas, Rhodes on Texas, BNEF on on-site economics) are all telling you the same buildout runs into a wall it has not priced. Both can be true. They resolve at different dates. The operators are describing 2027-2029. The system modellers are describing 2029-2032. If you own the equipment names, your job this year is to figure out which of those two windows your holding period sits in.

Read-Throughs

Gas turbines to gas producers, and back. This is the loop to internalise. Every genset, fuel cell and turbine sold behind the meter is a claim on the same molecules the LNG terminals have already contracted. Smith's phrasing was blunt: "Bring Your Own Generator," the arrangement where hyperscalers are asked to supply their own power to site a data center somewhere, "means more gas, not less. And so every time you read a press release from Bloom or from Google, think more gas." The read-through runs turbine order to gas demand to gas price to the cost of the turbine's own output. The equipment makers are, in effect, short the fuel their customers need.

Supporting evidence appeared elsewhere in the week. The Morning Market Briefing (July 24) noted SLB has an agreement with Meta to supply drilling equipment for gas that will power data-center turbines, and that SLB's data center business has reached a $1 billion revenue run rate in nine to twelve months. The oilfield service names are now AI infrastructure names, which nobody had on their 2024 bingo card.

Copper is the constraint behind the constraint. Two separate podcasts on the industrial supply chain landed on it. Around the Horn in Wholesale Distribution (July 24) cited a 10 million metric tonne US copper shortfall to supply data-center components, with 2 million tonnes of new demand by 2040 and up to 30 years from discovery to production for a new mine, noting that wiring harnesses, motors and drives are all copper-heavy, and "a copper shortage will prevent completion" of projects. On U.S. Manufacturing Today (July 21), Matt Bedingfield of Mint Innovation put a single data center at 50,000-200,000 tonnes of copper, a global shortfall of 6-10 million tonnes by 2035, and pointed out the US exports about a million tonnes of copper scrap while importing about a million tonnes of refined copper. A self-inflicted deficit. Freeport is the obvious liquid expression; the more interesting question is which cable and conductor manufacturers have contracted metal and which are buying spot.

Structural steel is quietly ripping. Brew Markets (July 23) flagged Steel Dynamics reporting a 45% year-over-year increase in steel fabrication order backlog driven by data center demand. Fabricated steel is not a glamorous AI derivative, and that is rather the point: it is a cleaner read on shovels actually going into ground than any announced-megawatt tally.

There are 100 gigawatts of generators already sitting idle. The most underdiscussed number of the week came from Bill Kaewert of Stored Energy Systems on The Industrial Talk Podcast Network (July 22): roughly 100 gigawatts of dispatchable backup generation already exists at US data centers, mostly doing nothing, with regulatory barriers to dispatching it before an actual outage occurs. Pair that with Duke's 50-100 hour curtailment clauses and Iron Mountain's work on batteries that can carry an entire data center for hours to unlock "stranded capacity" on the grid, and a picture forms: a meaningful slice of the capacity problem is a permissions problem. If regulators unlock existing backup fleets as a grid resource, that is bearish for incremental peaker orders and bullish for the controls, switchgear and software layer. Watch this thread, it is early and it is cheap.

Nuclear is being sold as a gas hedge, not a climate product. This is a genuine framing shift, and Blue Energy's CEO articulated it best on Inevitable (July 21): hyperscalers want nuclear "isn't because it's clean. It is clean and that's nice... but it's because it's firm capacity that hedges their gas risk. They're building a ton of gas turbines right now to feed the AI growth. And you have to kind of recognize you're taking on a lot of gas commodity risk, a lot of gas supply risk. Next time there's a polar vortex that rips through the Midwest and takes out all of our gas supply, that's going to take out gigawatts of compute." Add emissions-regulation risk, since administrations change and turbines can get ratcheted down before the end of their useful life, and nuclear becomes an insurance policy rather than an ESG line item. If that framing sticks, it changes how nuclear offtake gets priced: as an option on gas, not as a premium for clean power.

Blue Energy's own build is the most interesting structure described in a while. Rather than building gas now and nuclear next door later, they are building half a nuclear plant on day one and splicing in H-class gas turbines to energize it early, then switching the shared steam turbine over to nuclear steam. Their site is the Port of Victoria on the Texas Gulf, next to a Crusoe data center project with up to 1.5 gigawatts of potential load, and they intend to build more generation than that, because the Texas coast needed power for desalination, battery plants and ports before AI showed up. Two facts from that conversation belong in any nuclear model: 70% of a nuclear plant's capex is EPC scope, not the reactor vendor's scope, which is why cost overruns keep happening in construction, not design, and every nuclear plant on earth has been financed by taxpayers or ratepayers, never project-financed. Blue Energy used equipment-collateralised debt to put deposits on long-lead gas turbines, which is a clever way through the funding valley of death.

The efficiency counterargument, and why it keeps failing. Worth having an answer ready, because clients ask. The Around the Horn hosts worked through it: if chips get dramatically more efficient, doesn't power demand fall? Their conclusion is that "any sort of efficiencies will be eaten up by the demand aspect of it," because we have barely scratched the use cases, and cheaper inference means more of it. Sage Geosystems CEO Cindy Taff made the complementary point on The Data Center Frontier Show (July 21): US electricity demand was flat for decades, AI changed that "overnight," and the industry's priority has shifted from decarbonising power to simply getting enough of it.

What Changed This Week

Three things genuinely moved, and one thing that had been noisy went quiet.

Affordability went from background grumble to policy driver. A month ago the political risk in this trade was abstract. This week a grid operator failed a third auction, the White House floated breaking it up, a governor threatened to withdraw a state, one of the largest US utilities threatened to withdraw too, and the CEO of Duke Energy went on television to argue data centers will lower your bill. Alfonso Salema of Exus Digital made the parallel point about Europe on Inside Data Centre: data centers "are perceived as pushing the affordability issue, raising power prices to that consumer," and unless the industry helps grid operators use existing capacity better, that perception hardens into policy. Model political risk explicitly now, especially on anything with PJM capacity exposure.

The constraint narrative moved from wires to molecules to money. In that order, within one week. Interconnection queues are now table stakes, everyone knows about them. The live arguments this week were about whether the gas exists and whether the debt clears.

Great numbers stopped being enough. GE Vernova beat on backlog and free cash flow and fell 6%. Alphabet, per The Exchange, "beat nearly every measure of the return on its AI spending and still had its worst day in three years." When the market stops paying for good results in a theme, the marginal buyer has left. That is not a thesis-killer. It is a positioning signal, and it argues for owning the parts of this trade that are not consensus, which, this week, means the gas resource rather than the equipment that burns it.

And a thread that was quiet: no episode this window brought news on merchant nuclear power purchase agreements, or on the Crane and Three Mile Island restart and uprate milestones. Nuclear discussion this week skewed to advanced designs and to nuclear-as-gas-hedge framing rather than to the existing fleet's deals. Nobody made a new argument on uranium spot or term pricing, or on enrichment bottlenecks, either, with Cameco coming up only in passing on general-market shows. Absence of news is not news, but if you are waiting on a merchant nuclear catalyst, it did not arrive.