Newsletter · · Ashutosh Agarwal

Chevron Becomes Microsoft's Landlord and Power's New AI Trade - Powering AI: Grid, Gas, Generation & Nuclear - Week of August 11, 2026

Powering-AI infrastructure newsletter for the week of August 11, 2026. Chevron's behind-the-meter Project Kilby reframes the power-for-AI trade around signed 20-year take-or-pay contracts, while Caterpillar and Cummins can't add engine capacity fast enough and a specific bear case questions whether the demand is real.

Powering AI: Grid, Gas, Generation & Nuclear

Week of August 11, 2026: Chevron Becomes Microsoft's Landlord and Power's New AI Trade

An oil major just told Wall Street it wants to be valued like a pipeline, not a driller. That single sentence is the most important thing the podcasts said this week, and it tells you where the whole "power for AI" trade is heading: away from speculative merchant bets and toward signed, decades-long contracts where somebody with a balance sheet agrees to keep the lights on for a data center and gets paid like infrastructure for doing it. Meanwhile the machinery that actually makes the electricity keeps selling out faster than anyone can build it.

TL;DR

  • Chevron's Project Kilby, a ~3-gigawatt, behind-the-meter power plant for a Microsoft data center on a 20-year take-or-pay contract, is the template everyone is now chasing. Signed demand, infrastructure economics, not a merchant gamble.
  • The gear is the bottleneck, and the numbers are getting silly: Caterpillar's backlog is up 92% to $72 billion, it's taking orders out to 2030, and both Cat and Cummins are racing to add engine and turbine capacity they still can't build fast enough.
  • The bear case got a real voice this week: that most of the AI "demand" funding all this is really just two cash-burning labs, and the 300-plus gigawatts of planned data centers may be building for customers who don't exist.

What's new

Chevron just reframed the entire trade: it wants to be your landlord, not your gas station. The most consequential discussion of the week came on Big Digital Energy, where the hosts dug into Chevron's Project Kilby. Here's the deal in plain terms: Chevron will build a roughly 3-gigawatt power plant sitting right next to a Microsoft data center ("behind the meter," meaning it bypasses the public grid entirely), burn its own cheap Permian gas to run it, and sell Microsoft the electricity under a 20-year take-or-pay contract at mid-teens returns, with a final go-ahead decision targeted for the end of this year. Take-or-pay means Microsoft pays whether it uses the power or not, and that's what turns a risky merchant plant into a bond-like annuity.

Why it matters for numbers: as one host put it, Chevron is effectively telling Wall Street, "we're not just an oil company anymore. We're an infrastructure company that happens to sit on top of the fuel." Pipelines get infrastructure multiples; oil wells get commodity multiples. If the market starts valuing Chevron's power business like the former, that's a real re-rating. And Kilby is tiny in gas terms, about 80 million cubic feet a day, roughly 0.3% of Permian production, so this isn't a supply story, it's a business-model story. The kicker: Chevron said it's in "advanced discussions with existing and prospective customers," which the hosts read as code for Google, Amazon, and Oracle. Every oil and gas company with stranded gas and good engineers is now asking whether it can sell electrons instead of molecules.

Caterpillar turned into an AI stock, and the backlog is the tell. Three separate podcasts hammered the same point from different angles. On Brew Markets, host Ann Berry walked through Cat's record quarter: $20.5 billion in revenue (the first time it has ever topped $20 billion), sales up 24%, adjusted earnings of $8.17 a share, nearly $2 above expectations, and a raised full-year outlook that CEO Joe Creed pinned on "continued strength in AI-related infrastructure investment." Its power and energy business, which makes the engines, generators and backup systems data centers need, jumped 17%.

The more actionable detail came on Squawk on the Street, where CNBC's Seema Mody reported that Cat used the call to directly rebut the growing worry that the AI build-out is stalling: its backlog is up 92% from a year ago, and the buyers are hyperscalers and utilities. Jefferies called it the biggest beat in its industrials coverage. And on Bloomberg Intelligence, machinery analyst Christopher Ciolino added the part that matters most for the whole supply chain: Cat is now taking orders out to 2030, lead times are extending, its big reciprocating natural-gas engines are moving from backup duty into full-time "prime power" for data centers, and Cat is trying to grow large-engine capacity roughly 3x and turbines 2.5x by the end of the decade, with Cummins doing the same. His summary: "It's a race to add capacity… the industry really can't add capacity fast enough." The honest bear note buried in the same conversation: this is still a cyclical company trading around 35x earnings, and the analyst's own open question was, "is this as good as it gets?"

The PJM power auction went vertical, and politicians slammed a cap on it. If you want the cleanest read on how tight the grid has gotten, listen to Energy Gang. PJM, the grid operator for 13 states from New Jersey to Chicago, runs a yearly "capacity auction" that pays power plants just to be available. Exelon's guest explained that demand there is expected to rise 25% over four years, versus a historical norm of basically zero. In the latest auction, the Chicago (ComEd) zone would have cleared at $775 per megawatt-day without a cap, against about $28 just a few years ago. The 13 states' governors found that price shock politically unswallowable and capped it at $325. His most damning line: over the past two years PJM customers paid about $32 billion for capacity and got 1.2 gigawatts less power. Exelon's pitch is to let regulated utilities own generation again (not full re-regulation) because their lower cost of capital can be the "price to beat," and he pointed to Virginia, which builds and saw its capacity price rise only $2, versus Maryland, which doesn't and saw $18.

The counterpoint came from the merchant side on Energy Evolution, where LS Power CEO Paul Siegel called the cap "a short-term political win but damaging to confidence," because a price signal you're not allowed to act on doesn't attract investment. His bigger point is the one to hold onto: he expects most new supply to come through bilateral contracts (hyperscalers directly paying for their own generation on their own balance sheet), not the government-run auction, and that when they do, everyone else's grid costs should eventually fall because the same wires carry more volume. S&P Global's Tanya Peavy noted 23 gigawatts of capacity chose not to bid at all, some of it waiting for a separate "reliability backstop" that offers a higher $550 cap and possible 15-year contracts.

Texas is the whole thesis in one number: 900 gigawatts of requests, 90 gigawatts of grid. On Telltales, the hosts laid out that data-center developers have proposed roughly 900 gigawatts of projects in Texas, ten times the state's entire current capacity of about 90 gigawatts. Governor Abbott's response was a letter to regulators that, read closely, amounts to a roughly 12-month audit and delay, echoing New York's outright one-year hold on data centers. With combined-cycle plants too slow to build and turbines sold out, the hosts argued the only way to serve near-term compute demand is on-site turbines, which is exactly why they see gas-for-power demand resuming growth of about 1.5 billion cubic feet a day per year, on top of ~2 Bcf/d of LNG growth, enough to keep gas prices firmly in the $3.50 range and possibly toward $4.

The debate

The bull case ran through nearly every episode: this is a durable, multi-year supercycle that lifts grid, utilities, merchant generators, gas, and uranium together, because they're all links in the same chain and every link is short. Morgan Stanley's Thoughts on the Market put a figure on the constraint, about 38 gigawatts of incremental data-center power needed through 2028, and called power the primary bottleneck alongside politics and labor, with their base case a "conditional build-out" that keeps going past the midterms. The operator roundtable on Energy Crue sharpened it: Texas grid demand is seen jumping from 85 to 150 gigawatts by 2030, and gas is "the only viable bridge power for the next three to four years while nuclear technology tries to catch up." Their proof it can work is Ashburn, Virginia, the world's densest data-center hub, already running a pragmatic ~50% gas / ~40% nuclear mix at about 30% below the national average power cost.

The bear case finally got a serious, specific voice this week, and it's worth steel-manning because it threatens the entire complex at the root. On Better Offline, Ed Zitron laid out his "AI Demand Bubble" thesis, citing notes from UBS, Wells Fargo and Barclays: roughly 70% or more of Microsoft's, Google's and Amazon's AI revenue comes from just two customers, OpenAI and Anthropic, both of which lose enormous sums and survive on constant new fundraising. UBS reportedly estimates that OpenAI and Anthropic's compute spend will be about 48% of Google Cloud's revenue next year; Bloomberg pegs OpenAI at about $24.1 billion of Microsoft's fiscal-2026 revenue. Strip OpenAI out, Zitron argues, and Microsoft's growth was closer to 9% than the 17.7% it reported, even as it spent ~$115 billion in capex this year and plans ~$175 billion next year. His conclusion is the one every power bull has to answer: if there are 300-plus gigawatts of data centers in planning but the only real demand traces back to two companies burning cash, then a lot of the power being contracted is being built for customers who may not be there.

There's a second, quieter bear thread, that the trade is front-loaded. The Bloomberg Intelligence analyst's "is this as good as it gets?" on Caterpillar, and this week's Motley Fool Hidden Gems discussion, both land in the same place: the equipment names are great businesses, but the good news may already be in the price.

The names in play

GE Vernova and Quanta Services got named directly on Motley Fool Hidden Gems, where a host called both well-positioned for the data-center power and cooling build-out, but said the valuations are unattractive today and he'd rather wait for a better entry than own them now. That "right company, wrong price" caution is a useful tell that at least some of the buy-side is starting to flinch at the multiples.

Caterpillar and Cummins are the clearest operator-confirmed winners of the week (see above): the backlog, the extended lead times, and the shift into prime power are the real signal, not the stock pop. Chevron is the new wildcard: if Kilby gets its final investment decision and the follow-on deals with Google, Amazon and Oracle materialize, the story stops being about one project and starts being about a re-rating.

On the fuel cycle, In it to Win it had Rick Rule on Cameco: he says the recently announced Westinghouse IPO doesn't change his long-term view, because Cameco simply can't fund the projected US nuclear build-out on its own balance sheet and needs outside capital, and he expects the plan for ten 1-gigawatt Westinghouse reactors to proceed, which would be positive across Cameco's fuel-supply and processing chain. He's been buying uranium equities on the recent pullback ("I'm only interested in down moves"). Same episode, on gas: he thinks years of underinvestment "will," not "could," cause a supply shortfall by 2029-2030, and he owns EQT and Devon to play it.

Read-throughs

  • Gensets (CAT, CMI): the capacity race is the constraint, so whoever has product availability wins the contract. Both are tripling-ish engine capacity and still can't keep up, so watch for lead times to keep extending, which is bullish for pricing and bearish for the hyperscalers trying to hit compute timelines.
  • Copper and the wire (FCX read-through): on The David Lin Report, Copper Giant CEO Ian Harris made the blunt case: "there is no electrification without copper," Chile's output is down ~15% despite record prices, world inventories are historically about 15 days, and, the scary part, "the deposits don't exist" to meet demand, because a wire plant takes six months to build but a mine takes twenty years. Every transformer, cable and switchgear order downstream inherits that squeeze.
  • Gas-weighted producers and turbine makers: the Telltales and Energy Crue math (gas demand +1.5 Bcf/d a year, turbines sold out) reads straight through to Permian associated-gas producers and the turbine OEMs, and, per Clean Power Hour, to GE Vernova's and Mitsubishi Power's multi-year gas-turbine backlogs.
  • Uranium and the enrichers: if the Westinghouse-reactor plan is real, the read-through runs from Cameco through fuel processing and enrichment, the part of the chain that can't be stood up quickly.
  • The customers funding it all: every deal above ultimately rests on hyperscaler balance sheets. That's the bull's strength (Microsoft's signature on a 20-year contract) and the bear's whole argument (that the demand behind those signatures is two loss-making labs). You can't hold a view on the power complex without holding a view on that.

What changed

Two shifts stood out this week. First, the contract structure moved to center stage, the conversation is no longer "how much power will AI need" but "who signs the 20-year take-or-pay, and do they get a pipeline multiple for it." Chevron's Kilby made that concrete. Second, the regulatory brake got real: Texas's audit of its 900 gigawatts of requests, New York's one-year hold, and the political cap on PJM's auction all point the same way, the bottleneck is increasingly permission and hardware, not ambition. And on AI-Curious, Harvard's Ari Peskoe put a face on the political risk: his paper "Extracting Profits from the Public" argues ordinary ratepayers are subsidizing the build-out (PJM's market has cost ~$63 billion recently, with data centers responsible for ~$30 billion but not paying it), and he notes a single Meta facility in Louisiana will draw 5 gigawatts, as much as five nuclear reactors, four to five times the peak demand of New Orleans. That backlash is the risk that turns "supercycle" into "capex air-pocket."

What to watch into next week: Chevron's Kilby decision and any named follow-on deals; whether Texas or New York softens its data-center holds; and any hyperscaler capex commentary that either confirms or undercuts the demand the bears are questioning.