Newsletter · · Ashutosh Agarwal

Amazon Builds a 7.6 Gigawatt Private Gas Grid in West Texas - Powering AI: Grid, Gas, Generation & Nuclear - Week of August 18, 2026

Powering-AI podcast roundup for the week of August 18, 2026. Amazon is building a 7.65 gigawatt gas plant in Pecos County with no grid interconnection, the White House put its 100,000 megawatt benchmark behind a fast-track policy push, and the tape argued over whether the risk is demand or the borrowed money financing it.

Powering AI: Grid, Gas, Generation & Nuclear

Week of August 18, 2026: Amazon Builds a 7.6 Gigawatt Private Gas Grid in West Texas

The single loudest thing on the podcasts this week wasn't a utility earnings beat or a new reactor design. It was a hyperscaler quietly deciding it no longer needs the grid at all. Amazon is building a 7.6-gigawatt gas plant in the West Texas desert with no plan to plug into the public network, a private power station the size of a mid-sized country's entire draw, built to feed data centers directly. That's the story of the moment: the companies that need power are done waiting in line for it, and everyone from the White House to the engineers building HVDC converters is scrambling to keep up.

Underneath the noise, a clear picture emerged. Demand is real and enormous. Supply (turbines, transformers, linemen, interconnection queues) simply cannot move at the speed the money wants. And that gap is where the whole investment case lives.

TL;DR

  • The buildout has a number now. The White House is planning around 100,000 megawatts of new US generation by 2030, half of it for data centers, on top of $1 trillion-plus of grid spending. Contracted power assets are changing hands at 7-11x free cash flow for mid-teens returns, and a new gas turbine is a four-to-five-year wait.
  • Behind-the-meter gas just went from workaround to strategy. Amazon's 7.6 GW off-grid plant and EQT's "we're happy to go fully behind the meter" mark a real pivot: build your own power now, connect to the grid in a decade if ever.
  • The bear case is funding, not demand. CoreWeave is sitting on a ~$130 billion backlog while burning $9.4 billion of capex a quarter against $2.5 billion of revenue. The debate is no longer "is the demand real," it's "who pays if the financing music stops."

What's new

A hedge-fund PM and a White House power official laid out the whole thesis in one sitting. The most useful hour of the week came from Thematic Investors, where host Kieran Cavanna sat down with Ravi Beller, who runs the energy-focused hedge fund Ripple Effect, and Peter Lake, the White House's Senior Director of Power (and the former head of Texas's grid regulator). Lake put a hard number on it: "the benchmark we're using from the White House… is 100,000 megawatts of new power generation is needed by 2030," which he called "adding a California and a New York on top of our existing power grid." About half is data centers; the rest is factories, warehouses and general electrification. Beller's investing read is the part that moves numbers: generation portfolios (gas, renewables, batteries) are trading at "anywhere between 7 to 11 times free cash flow," with "50 to 100 percent of their cash flows contracted for an average of 12 to 20 years" at "mid-teens, unlevered returns." Why now, why durable: the hyperscalers, worth a combined $15 trillion-plus, "can pay 2 to 3x more" and will sign 20-year take-or-pay deals, so the cost of the build lands on them rather than on ordinary ratepayers. This is the clearest steel-man of the bull case anyone has voiced in weeks, and it comes from an operator and a policymaker, not a pundit.

Amazon showed what "just build it yourself" looks like at scale. On Big Digital Energy, the hosts walked through Amazon's Pecos County project: an 8,000-acre site, 35 gas turbines, 7.65 gigawatts, developed by Pacifico Energy, permitted to emit up to 33 million tons of CO2 a year, "roughly double the most polluting coal plant in the U.S." Crucially, Amazon has not applied for a grid interconnection. As one host put it: "This is no longer an alternative to the grid. It is the grid. It just happens to be a private grid owned by hyperscalers built on gas in West Texas." This matters for two reasons: it's a template others will copy in gas-rich ERCOT, and every molecule burned behind the meter is a molecule that doesn't flow through a pipeline, a live debate for gas-weighted names.

GE Vernova's own CTO walked through the equipment bottleneck. This is insider commentary worth its weight: on Redefining Energy, host Michael Barnard interviewed Cornelis Plett, CTO of GE Vernova's grid-integration arm. Plett described high-voltage direct-current (HVDC, the long-distance backbone that moves bulk power efficiently) converging on a "2 gigawatt standard design at 525 kV" in Europe, with "3 gigawatt" designs now appearing in the US and India, and noted Dominion has already approved HVDC links specifically to serve data-center demand. The quiet tell for margins: GE Vernova owns the "actual factory for the special type of transformers that are needed for HVDC converter stations," which lets it "control the factory slots and the capacity." When the scarce input is factory time, the company that owns the factory sets the price.

The funding-risk case got its poster child. On Market Maker (pundit analysis, to be clear), the hosts dissected CoreWeave's numbers: an order backlog of "104 billion… up 246% year over year," plus another $25 billion landed after quarter-end for roughly $130 billion, about "10 times their forward-looking revenue." But capex hit "$9.4 billion in Q2 alone" against ~$2.5 billion of revenue, with a $567 million adjusted loss and quarterly interest expense that "surged to $640 million." Their verdict: "it's a funding risk above everything else." The same week, NVIDIA lined up a $500 billion private-capital partnership (Apollo, BlackRock, Blackstone, Brookfield, Goldman, KKR) to underwrite AI infrastructure, with Jensen Huang reframing chips as "a new class of productive, investable infrastructure." Read together: the demand is being financed with an enormous amount of borrowed money, and that's the fault line.

Flexibility quietly became the cheapest new "generator" on the grid. Two shows landed on the same insight. On Moonshots with Peter Diamandis, investor Ramez Naam cited Tyler Norris's research that being flexible just "100 hours out of the year… unlocks 100 gigawatts of capacity on the grid," roughly $5 trillion of data-center capacity, without building a single new plant. And on Cleaning Up, EPRI President Arshad Mansoor described a new industry standard for load flexibility ("DCFlex") already embraced by "more than 60 organizations, including Google." The takeaway for the supercycle bulls: if flexibility lets 100 gigawatts of demand ride on 50 gigawatts of new hardware, some of the equipment order book everyone is extrapolating may never get placed.

The debate

This week the bull and bear cases were both argued well, by credible people, so let's steel-man each.

Bull: this is a multi-year, multi-asset supercycle, and it re-rates the whole stack together. Beller's framing on Thematic Investors is the tightest version: demand is growing at "3% to 4% annualized", the fastest in 50 years, and "even if you took AI out of the picture, power demand… would still be growing about 1% per year." Supply can't respond because "you're going to be waiting… four or five years to get a new turbine if you're ordering it today," and there's "limited labor… limited equipment… limited land… limited permitting." Add hyperscaler take-or-pay contracts and a new 15-year price signal out of PJM, and you get durable, contracted, mid-teens-return assets across gas, grid equipment, and IPPs. Lake's policy tailwind is real: fast-track interconnection can shrink a curtailable data center's wait from "six or eight years" to "60 or 90 days." On The KE Report, commentator Joel Elconin put the pundit's version bluntly: nuclear and "the energy trade" are the picks-and-shovels on AI capex.

Bear: the demand number is inflated, the timelines are fantasy, and the financing is the weak link. On Cleaning Up, host Michael Liebreich pushed back hard on Mansoor's 100 GW: his number is "30 gigawatts by 2030," aligned with Jigar Shah's ~34 GW, and he noted only "two and a half gigawatts" of AI compute were actually added last year. Getting to 100 GW by 2032, he argued, requires "an improbable step change," and the "chance of there not being a correction between now and 2032… is as close to zero as I can imagine", with a bond-market correction the truly dangerous one. Both men agreed nuclear is not the rescue: Mansoor himself said advanced reactors and SMRs are "not the technology that enters the equation" before 2035, likely just "500 megawatt, one gigawatt" by 2032, because "you need to build at least 10… before you could build 100." On Moonshots, Naam made the same point about cost curves: nuclear is "ruinously expensive" precisely "because we don't do a lot of it." And the Market Maker funding-risk argument caps it: bookings are being pulled forward on borrowed money, and lead times are near their peak.

The honest middle, voiced by Mansoor: flexibility might mean "100 gigawatt doesn't need 100 gigawatt of infrastructure investment. Maybe it needs 50." Both sides could be partly right: huge demand, but met with less new steel than the order books imply.

The names in play

GE Vernova (GEV) got the most direct, insider read of the week via its own CTO: HVDC standardizing at 2-3 GW blocks, and captive transformer-factory capacity as a genuine scarcity moat. Barnard also noted GE and Hitachi are "building new assembly lines" for large gas turbines that are "sold out for something like seven years" (per Naam on Moonshots), a backlog that's bullish for pricing and bearish for anyone hoping to get powered quickly.

Caterpillar (CAT) and Wärtsilä came up as the on-site workhorses. Mansoor said EPRI is "working with companies like Caterpillar, Wärtsilä" on cleaner backup generators, and is building an Arizona data center microgrid with "Caterpillar, Wärtsilä units… configured as a microgrid that can isolate from the grid." Every islanded, behind-the-meter campus is genset demand.

Southern Company (SO) got a specific, favorable mention from the White House's Lake: it is "building out what I believe is the single largest power generation development in American history", up to 16,000 MW, while "freezing rates for the residential customers in Georgia and Alabama." That's the "done right" regulated template that keeps the political backlash at bay.

On the merchant/IPP and pure-nuclear names the podcasts leaned thematic this week. The bull case for contracted generation was made in the abstract (7-11x FCF, mid-teens returns) rather than ticker by ticker, so treat that as sector tailwind rather than a specific call.

Read-throughs

  • Gensets (CMI, CAT): the behind-the-meter pivot, Amazon's 35-turbine plant, EQT's gas-plus-battery bridges, EPRI's microgrids, is a direct read-through to on-site generation and backup power. If "build your own" is the new default in ERCOT, engine and turbine demand is stickier than the grid-connection narrative suggests.
  • Copper, transformers and cable: The Dynamo Show (Aug 12) hammered that transformers and switchgear are "critical things that you actually need in order to turn on a data center," and floated a "Foxconn for electrical equipment" to commoditize them. Until that exists, incumbents with capacity (see GE Vernova's captive transformer factory) hold pricing power, bullish for the whole electrical-steel and conductor chain.
  • Gas-weighted E&Ps and turbine OEMs: behind-the-meter gas is a double-edged read. On Big Digital Energy, the hosts flagged that islanded plants mean "more gas not flowing into the pipelines", good for wellhead and basin demand near West Texas, more complicated for long-haul midstream volumes. Turbine OEMs (GE, Hitachi) simply can't lose in the near term with seven-year backlogs.
  • Uranium and the fuel cycle: the podcasts were a splash of cold water. The consensus from Mansoor, Liebreich and Naam is that nuclear is a 2035-and-beyond story, not a this-cycle catalyst. Beller's line, "we are long gas. We are short uranium", captures the near-term tilt: gas and grid do the heavy lifting now; uranium is the long-duration option.
  • Renewables, storage and the "where can I even connect" problem: on Factor This, OCI Energy's Sabah pointed out that ERCOT just set a 91-gigawatt peak on July 22 (up ~6 GW from the prior record) yet cleared "less than $1,000" per megawatt-hour against a $5,000 ceiling, because "close to 10 to 11 gigawatt capacity of storage" and new solar absorbed the load. Solar PPAs run "50 versus $90" in ERCOT vs PJM. The kicker is interconnection: "there is ERCOT and there is everybody else," with MISO and PJM queues still stuck on 2018-2020 applications, which is exactly why data centers keep flooding back to Texas. Read-through: near-term supply is solar-plus-storage and gas, not nuclear, and Texas keeps winning the load.
  • Hyperscaler load customers (AMZN, META, MSFT): they are funding the deals, and increasingly building the power themselves. That's what makes the contracted-asset returns real, and, per Market Maker, what makes the neocloud layer (CoreWeave) a leveraged bet on that same demand holding up.

What changed

The tone on behind-the-meter shifted this week. As EQT's Jan noted on The Infrastructure Investor Podcast, flexibility and off-grid power used to be "almost a little bit taboo," something nobody wanted to talk about. Now it's the headline strategy, "we're happy to go fully behind the meter. The solutions work", with grid connections treated as a nice-to-have that might arrive "eight, nine, ten years out." Between Amazon's 7.6 GW plant and EQT's stance, the center of gravity has moved from "wait for the utility" to "power yourself, and let the grid catch up." That's the thread to watch into next week, especially whether ERCOT and regulators try to pull those private islands back onto the public system.