Newsletter · · Ashutosh Agarwal

Hyperscalers Build Their Own Power as the Grid Falls Behind Data Center Demand - Powering AI: Grid, Gas, Generation & Nuclear - Week of September 15, 2026

Powering-AI podcast roundup for the week of September 15, 2026. About 230 GW of data-center interconnection requests face roughly 93 GW of new power plants, pushing hyperscalers toward on-site gas, batteries and nuclear, while a voter backlash, a Texas hookup slowdown and signs of an Oracle capex peak sharpened the bear case.

Powering AI: Grid, Gas, Generation & Nuclear

Week of September 15, 2026: The Grid Can't Keep Up, So Hyperscalers Won't Wait

The AI power story stopped being a slide in a bull deck this week and became a fight over your electric bill. Politicians are running against data centers, Texas just hit the brakes on new hookups, and the companies that need the electrons have simply stopped waiting for the grid; they're building their own power plants in the parking lot. Underneath the noise, the numbers that matter for a book got bigger, and the timeline got more honest.

Here's the through-line: demand is real and enormous, but the grid physically cannot connect it fast enough, so capital is sprinting to whatever can deliver power now: gas turbines, batteries, on-site generation, and eventually nuclear. That's great for the equipment makers and the merchant generators. It's also exactly the setup that produces an air-pocket if the announced load turns out to be half real.

TL;DR

  • The mismatch is the trade. There are roughly 230 gigawatts of data-center interconnection requests against maybe 93 GW of all new power plants getting built over five years, a 2.5x gap that is pushing hyperscalers off the grid and into gas and batteries.
  • The capex wave is confirmed and accelerating. Utilities are planning $1.4 trillion of spending by 2030 (+21% year-over-year); utility capex already grew 22% last year to $212 billion. Equipment prices are up 2-3x across the board.
  • The bear case has a face now. Public opinion has flipped hard against data centers, Texas is auditing 250-300 projects, and Oracle's own numbers hint that hyperscaler capex peaks around 2027 and slows in 2028.

What's new this week

1. The interconnection math went mainstream, and it's brutal. On Interchange Recharged, Wood Mackenzie analyst Sylvia Leyva Martinez and clean-power developer Tim Montague laid out the single most important number in the sector: about 230 GW of data-center interconnection applications against a Bank of America estimate of only ~93 GW of new power-plant buildout over the next five years, solar, wind and gas combined. That's 2.5x more demand knocking than supply being built. Montague's own gloss, from his show Clean Power Hour: "there's at least two and a half times as much load growth as is going to get built." Why it matters: this gap is the entire investment thesis for anything that delivers power quickly, and the reason the grid queue itself has become the bottleneck.

2. The $1.4 trillion number is now the benchmark. On Open Circuit, a Power Lines analyst debuted research (picked up by the Wall Street Journal that morning) that US utilities plan to spend $1.4 trillion of capex by 2030, a 21% jump over last year's ~$1.1 trillion. For scale, they called it "three interstate highway systems" of spending. Utilities requested a record $31 billion in rate increases last year, and retail power prices are already up 40% in five years. This is the rate-base engine behind the regulated-utility EPS story, and the political powder keg behind point 4.

3. Equipment is the choke point, and pricing power is real. The best operator-grade detail came from Catalyst with Shayle Kann, where Kann (an investor at Energy Impact Partners) and his colleague walked the supply chain component by component. Their tally: conductor is roughly twice as expensive, transformers 2x-plus, switchgear 2x, and gas power plants 2-3x more than a few years ago. Gas turbines in particular keep setting records: they cited combined-cycle turbines quoted at $3,600 to $4,000 per kilowatt. And here's the kicker for the utilities: about two-thirds of all US utility spending is just maintenance and hardening (per the Edison Electric Institute), non-negotiable work that would happen anyway, and is also getting more expensive. Translation: backlog and pricing for the electrical-equipment names is being driven by far more than data centers alone.

4. The politics turned, fast. On Growth Stories with IBD, Michael Thomas, founder of CleanView, which tracks US data-center and power projects, noted that three-quarters of Americans now say they'd oppose a data center in their neighborhood, versus an even split just a year ago. Meanwhile, on The Neuron, the hosts flagged that Texas Governor Abbott ordered a slowdown of new data-center hookups through ERCOT, with 250-300 projects to be audited, against a staggering 474 GW of proposed new load, about five times the state's record peak demand of ~94 GW. And a research group, Data Center Watch, found grassroots opposition blocked or delayed at least 75 projects worth ~$130 billion in just the first three months of 2026. This is now a bipartisan campaign issue heading into the midterms.

5. Speed-to-power is the new religion, and it's bullish for gas. Thomas's central point on IBD: hyperscalers have flipped their priority from efficiency to speed. Google got a Texas data center online in 18 months by co-locating solar, wind and batteries with a partner (Intersect Power), versus a typical five-year ERCOT wait. Meta is signing 10-year natural-gas deals with Williams and dropping billions of dollars of chips into 12-month "tents" instead of its efficient 24-month buildings. xAI's Colossus ran on mobile gas turbines, then shifted them to a second site when the first got a grid connection. As Thomas put it, hyperscalers are "essentially becoming utilities."

The debate

This week the podcasts leaned bullish on demand but unusually candid on risk. Both cases were voiced.

The bull case: a multi-year supercycle across the whole stack. The demand is a genuine shock: the GridPulse guests, Carnegie Mellon's Paulina Jaramillo and Johns Hopkins' Abe Silverman (a former FERC attorney), called it "tectonic," the largest since the 1960s-70s. Silverman's framing sticks: PJM has to add "a Philadelphia's worth of electricity" (6-8 GW) every single year through the early 2030s just to serve data centers. Lawrence Berkeley National Lab estimates data centers could reach ~50% of annual US electricity demand, from essentially nothing five years ago. When the grid can't keep up, the market pays up: PJM's capacity auction hit its ~$350/MW-day cap and still didn't clear all its needs, and the market monitor pins ~$30 billion of higher capacity costs across the last three auctions on data centers. That is pure tailwind for merchant generators and anything that makes an electron. On Capital Decanted, the hosts made the picks-and-shovels version of the argument: power is now "the top priority for GPs," utility capex grew 22% year-over-year to $212 billion in 2025, triple the prior decade's 7.6% pace, and the "adjacency" plays (power, cooling, grid gear) let you bet on the buildout without picking which AI model wins.

The bear case: front-loaded bookings, then an air-pocket. The same Capital Decanted hosts wouldn't let the bull case off the hook: hyperscaler capex is heading to 45-50% of revenue by 2026, they cited a view that the "ROI math is fundamentally broken," and warned of stranded-asset risk if the boom cools, invoking the railroad and shipbuilding busts. The GridPulse academics sharpened it: data centers are "a four-year-old industry," and we're building trillion-dollar forecasts "as if they're going to meet all of their wildest aspirations", while grandma absorbs the risk if they don't. And you can see the taper in a real P&L: on Chip Stock Investor, the Rossolillos noted Oracle's capex hit $28.5 billion in a single quarter, up from $8.5 billion a year earlier, pushing spending above revenue and to 1.6x operating cash flow, with negative free cash flow and $125 billion of debt. Their call: peak capex around fiscal 2027, a slowing growth rate in 2027, and a more meaningful slowdown in 2028. If the biggest buyers are within 18 months of their capex peak, the downstream power-demand curve isn't a straight line up.

And a frame-breaker worth hearing. On Healthy Spaces, a European data-center operator pushed back on the whole premise: data centers, he argued, are "less than 9%" of the grid growth actually needed: transportation, heating and cooling, and heavy industry each require roughly 3x more. "We are not the cause of a problem," he said. "We are a symptom of an underinvestment in the grid", one stretching back 40-50 years, requiring grids to grow 2-3x faster than they're planned to. If he's right, the capex supercycle is even bigger and more durable than the AI-only story implies, because AI is just the trigger, not the whole load.

The names in play

Merchant and nuclear generators (VST, CEG, TLN). The read-through from PJM's capacity crunch (auctions hitting the price cap, $30 billion of attributable cost) flows straight to the merchant IPPs that own dispatchable and baseload capacity. Nobody laid out a fresh company thesis on Vistra or Talen this week, but the price signal behind them got louder. Constellation got a specific mention on The 7investing Podcast, where analyst Heather Horton said she owns it for the Three Mile Island restart serving Microsoft: "a really interesting AI play behind what's getting so powerful."

Small modular nuclear (SMR, OKLO, GEV). That same 7investing episode was the week's clearest-eyed nuclear discussion, and it was as much caution as enthusiasm. The economics still don't work yet: last year's levelized cost of energy (excluding China) ran about $70/MWh for hydro, $55 for solar, $51 for onshore wind, and SMR demonstration projects are claiming $80-130/MWh, which the hosts think is really $150-plus once reality intrudes. The bull path is a second generation getting to $40-90/MWh as the technology scales, plus build times of 2.5-3 years instead of a decade. NuScale is the furthest along commercially (a Romania utility project, Utah municipal work, and big optionality with the Tennessee Valley Authority); GE Vernova is iterating its proven Hitachi-partnered BWRX-300; Oklo is earliest-stage. The honest framing from the hosts: "we're still in the first, second innings," and the switch hasn't been flipped on real power yet. Own these as multi-year options, not this-year cash flows.

Hyperscaler buyers (GOOGL, META, AMZN, ORCL). The IBD conversation had the sharpest company color. Google is the consensus energy leader: it relicenses hydro, funds nuclear uprates and enhanced geothermal (it announced a very large geothermal project this week), and cracked the ERCOT queue with on-site renewables. Meta and Amazon are leaning into gas and modular "tents" to buy speed. Oracle is the cautionary tale: it's on the hook for much of Stargate's elevated capex, and, per Chip Stock Investor, that's where the strain shows up first.

On-site power and gensets (CMI, CAT). On Power Onward, Cummins insiders were blunt: "the utility cannot keep up with the growth of the industry," so customers "don't have a choice" but to build their own generation. The pairing that matters is genset-plus-battery: batteries handle the peaks (reacting in milliseconds versus a ~10-second genset start) and let the engine run at its efficient sweet spot, saving fuel and extending life. That's a structural demand story for prime-power gensets, not just backup.

Read-throughs

  • Off-grid gas is the fastest-growing demand pool. Every speed-to-power decision (Meta's 10-year Williams deals, xAI's mobile turbines, the "tents") is incremental gas burn happening outside the utility. Bullish for gas turbine OEMs and gas-weighted suppliers; watch turbine lead times, which the Catalyst crew says keep hitting new price highs.
  • Batteries are the shock absorber for the whole stack. From Critical Loop's microgrids (Clean Power Hour; CEO Bala Ramamurthy is targeting ~100 MWh deployed by early next year and 500 MW contracted by year-end) to TVA substation batteries to Cummins' BESS-plus-genset, storage is the bridge that lets a site with 7 MW of utility power run a 20 MW load today.
  • Copper and the raw-material upstream (FCX). The Capital Decanted hosts pushed the picks-and-shovels logic all the way up the chain to critical minerals (gallium, lithium and copper) as the deepest chokepoints. Conductor being "twice as expensive" (Catalyst) is the same story from the transformer end.
  • Uranium's term price keeps grinding toward $100. The only uranium data this week came from a technical trader on In it to Win it (treat it as chart commentary, not fundamentals), but the level is worth noting: long-term uranium contracts at $96.50/lb, with spot about $6.50 below, and Sprott's physical trust holding ~81.7 million pounds at a ~10% discount to NAV. No operator or Cameco commentary surfaced to corroborate.

What we're watching into Friday

  • Any operator commentary on gas turbine and transformer lead times with hard backlog numbers: the pricing is confirmed; the order books are the missing piece.
  • Whether the Texas/ERCOT audit and the midterm politics start showing up in project cancellations that matter to the merchant and equipment names.
  • A real Cameco or enrichment voice on the fuel cycle.