Newsletter · 2026-10-06 · Ashutosh Agarwal

AI's Power Crunch Spreads to Engines and Compressors as Oracle Invokes Force Majeure - Powering AI Infrastructure - Week of October 6, 2026

The powering-AI-infrastructure weekly for the week of October 6, 2026 (podcasts published September 29 to October 6). Caterpillar is taking engine orders into 2029 and 2030, compressor lead times have reached about 200 weeks, Oracle is invoking force majeure on its Project Jupiter site in New Mexico, and New Jersey now requires data centers to bring their own power.

Powering AI Infrastructure

Week of October 6, 2026: AI's Power Crunch Spreads to Engines and Compressors as Oracle Invokes Force Majeure

For two years the shortage in AI power was big gas turbines. This week's podcasts made clear the shortage is spreading. Engines, compressors and the pipes that feed data-center sites are now sold out to the end of the decade. And for the first time, a large AI project, Oracle's Project Jupiter in New Mexico, is formally blaming a power-supply delay to push back its payments.

Meanwhile, the politics are hardening. New Jersey's governor laid out the rules data centers must now follow in her state, and the first of them is: bring your own power.

TL;DR

• The bottleneck has moved to engines and compressors. Caterpillar is taking orders into 2029–2030 and restarting an engine line it shut in 2022. INNIO says it is "more or less sold out for 2026 and 2027." Lead times for gas-compression equipment have gone from 40–60 weeks to about 200.

• Oracle's force majeure is the first real crack in AI project financing linked to power. It is invoking force majeure on Project Jupiter over pipeline and environmental delays, and podcast hosts say the project slips at least six months. Oracle's credit is already priced as the weakest of the hyperscalers.

• "Bring your own power" is becoming policy. New Jersey now requires data centers to supply their own energy, pay into the grid and sit in a separate rate class. Polling shows about 60% of US adults oppose new data centers, up 12 points this year. That favors on-site gas, engines and batteries over plans that rely on the shared grid.

What's new

1. Engines and compressors are the new turbines

The most useful podcast this week came from a gas-compression trade outlet. The October edition of GCM On The Go (Oct 5) read out what executives said on their Q2 2026 earnings calls about lead times. This is operator commentary, not pundit opinion, and it all points the same way.

• USA Compression (USAC), CEO Clint Green: "We ran 40-60 week delivery lead time on equipment for years, and now with the generator market growing like it has, it has driven out to 200 weeks. We are ordering equipment for 2028 and 2029, and we'll be looking at 2030 here pretty quick."

• Archrock (AROC), CEO Brad Childers: lead times are "right at 195 weeks... We do not see these long lead times abating or improving."

• Kodiak Gas Services (KGS), CEO Mickey McKee: Kodiak has locked in deliveries of large compressor packages for 2027, 2028 and 2029. Its target is a fleet of about 5.2 million horsepower by the end of 2030, adding roughly 150,000 horsepower a year.

Why does a compression company care about AI? Because both compete for the same engines. The feature put it plainly: demand from hyperscale data centers "has pushed engine manufacturers to full production capacity, limiting the availability of engines for the gas compression industry."

Then come the engine makers themselves.

"There is a lot of discussion around AI demand, and all I can tell you is no one is slowing down at the moment." Joe Creed, CEO of Caterpillar (CAT)

Caterpillar says it is "starting to take orders into 2029 and 2030." It has restarted production of its 10-megawatt medium-speed gas engine. It stopped making that engine in 2022 because demand was too thin. The restart brings back about 1.5 gigawatts of yearly capacity, with first shipments in Q4 2026. Caterpillar also turned a 250,000-square-foot Kansas plant into a turbine packaging and shipping site. Creed was clear that demand is not the limit: "It's just going to be a matter of how fast we can continue to increase production."

INNIO, the private maker of Jenbacher and Waukesha engines, gave the most striking numbers:

• a $6.6 billion backlog;

• $2.3 billion of equipment orders in Q2 2026, "fueled primarily by data center customers seeking behind-the-meter power solutions" (power made on site, not drawn from the utility grid);

• delivery slots running into 2030 and beyond.

CEO Olaf Berlin: "We are more or less sold out for 2026 and 2027, and we have many projects where customers are talking about 2030 and 2031." CFO Dennis Schulz says INNIO plans to lift capacity from 3.5 GW in 2025 to about 10 GW by 2030. It will do that by expanding existing plants, not building new ones.

Why it matters: For investors, the shortage is now the story, not a side issue. On-site generation has moved from a stopgap to a full equipment market with its own multi-year backlog. That supports pricing for engine makers and for companies that lease out compressors, whose new equipment is now scarce.

2. Oracle's Project Jupiter: a power delay turns into a payment question

The Oct 2 issue covered the Energy Transfer pipeline to Oracle's (ORCL) Jupiter site in New Mexico, and the slip in its construction timeline. This week it escalated.

On The Enterprise AI Show, "AI News for End of September 2026" (Oct 4), the hosts reported that Oracle is citing force majeure over pipeline and environmental problems at Jupiter, and may delay payments as a result. Force majeure is the contract clause that excuses a party when events outside its control stop it from performing. One host called it "almost like the onset of a cold... Could be nothing." He also noted this is "how this stuff goes," and said he would watch for a second company doing the same.

The Canadian Investor, "The $1 Trillion AI Spending Boom & the Race to Make It Profitable with Braden Dennis" (Oct 3) explained how the deal works:

• Blue Owl owns the site as landlord. Oracle is the tenant, but Oracle is "also responsible to making sure that there's power that gets to the project."

• Oracle is invoking force majeure to "pay a lower lease until it becomes operational."

• Because of regulatory issues, the project is "being pushed back at least 6 months."

The same speaker called Oracle "the most vulnerable" hyperscaler. He pointed to credit default swaps, which are insurance contracts against a borrower failing to pay: "They're the highest of the hyperscalers. Like it's not even close."

The DHUnplugged hosts made a similar point on DHUnplugged #820: Gullible Markets (Sep 30). Oracle is "having a big problem with a couple of their deals," and builders are "running into the wall" on components.

Why it matters: This is the first time a power-supply delay has visibly changed who pays what in an AI lease. If force majeure becomes the standard way out when power arrives late, data-center landlords and their lenders take on power-delivery risk they thought belonged to the tenant. And a site with its own generation becomes worth more than one waiting on a pipeline.

3. New Jersey writes the rulebook: bring your own power

On Shift Key, "Governor Mikie Sherrill on Electricity Prices, Nuclear Power, and Building in New Jersey" (Oct 2), the governor talked with Heatmap's Matthew Zeitlin. Her 2025 campaign was dominated by power bills. Host Robinson Meyer said bills rose about $260 per household because of the AI boom's effect on PJM, the wholesale market covering New Jersey and 12 other states.

She froze rates "less than an hour into my administration," and says her measures will save ratepayers "over a billion dollars a year." She also lifted a 50-year nuclear moratorium and approved 18 solar and battery projects in six months. Her terms for data centers are blunt:

"They need to bring their own energy. They need to invest in our grid. They need to report their water and power usage... We've also put them in their own rate class so they are not harming other rate payers. And we mean business."

She pointed to an enforcement action against a data center in Vineland. And she told the industry: "you guys have been horrible at it... nobody knows what a data center is and you need to explain why it's even important."

Why it matters: New Jersey sits in PJM, where the price politics are hottest. A separate rate class means data centers pay their own way and other customers don't subsidize them. Together with bring-your-own-energy, this is a template other PJM states can copy. It moves value from adding load to the shared grid toward on-site generation and storage, which is the same direction as the engine story above.

4. The capex numbers keep going up, and so does the bill for borrowing

On The a16z Show, "The $1 Trillion AI Buildout | State of Markets" (Sep 30), the a16z partners put combined capital spending (capex) at Alphabet, Amazon, Meta, Microsoft and Oracle at about $780 billion in 2026, up from $416 billion in 2025. They said "all expectations point to them spending over a trillion dollars annually from 2027." They cited $1.7 trillion of combined cloud backlog at Microsoft, Google and Amazon. They also said that in "certain elements in the data center supply chain... you can't get access to materials or products until 2028." Their framing for this audience: "we should think of it as someone else's order book... a boon for chip orders, for power, for construction."

The Canadian Investor gave the long view. The same five companies spent $71 billion on capex in fiscal 2019 and $586 billion over the last 12 months, roughly 8x. At Amazon, capex has gone from 36% to 83% of operating cash flow.

The money is not free. On Odd Lots, "Everything in Markets Is Now Moving Incredibly Fast" (Oct 1), the guest cited Goldman Sachs: "about half of earnings growth this year is expected to be just driven by hyperscaler CapEx." He also said hyperscaler investment-grade bond issuance in September was, he believed, "a zero." One of the hosts added that the hyperscalers "got a little bit spooked with the CapEx response from the market." Separately, DHUnplugged reported that Goldman is now underweighting bonds from all the major AI hyperscalers, worried that the flood of new debt pushes prices down.

Why it matters: The spending plans themselves aren't shrinking, which supports equipment backlogs. But the financing is getting pickier. Spending funded by cash flow (Microsoft, Alphabet) looks steadier than projects funded by leases and special-purpose vehicles (Oracle, Blue Owl structures).

5. Uranium: contract terms still tighten while the stocks slump

Justin Huhn of Uranium Insider did two podcasts on Oct 1. Last week's issue covered the headline numbers. The new detail this week is how contracts are being written.

On In it to Win it, "Justin Huhn Says Uranium Supply Jumped 50% And The Price Still Doubled", he put reactor demand at just over 200 million pounds this year against about 175 million pounds of mined supply. After secondary sources, that leaves a 15–20 million pound shortfall. Spot is about $89.35. Spot is the price for near-term delivery in a thin market. The published term price for long-term contracts is $96.50, but he says that figure is only the low offer. "We're already north of $100 a pound in every single forward market with the exception of spot." Taking the midpoint of today's contract floors and ceilings gives "$110, you know, $115 a pound."

On Palisades Gold Radio, "Justin Huhn: The Uranium Blow-Off Top, AI Hyper-Scalers & 'Huge Value' in The Miners", he said producers like Cameco (CCJ), Kazatomprom and Orano are signing contracts for delivery in the late 2020s and early 2030s with "floors in the 70s, ceilings at $140, $150, sometimes we're hearing $160." Some contracts have "floors and no ceilings." His wildcard:

"Somebody like Amazon, somebody like Meta actually signing a long-term uranium offtake or a long-term contract. We expect the price to go well north of $200 a pound."

He counts "almost 80 reactors under construction in the world right now," the most in decades.

On 10-Minute Contrarian, "Ep275: Uranium Is Cheap Again" (Oct 4), host VP built on Huhn's numbers. China is building 38 reactors, with 43 more planned and 142 proposed. India is building 8, with 14 planned and 26 proposed. Spot has "bounced between that $85 to $90 range" all year while "all uranium equities have been dropping hard." He thinks short-term buyers who came in on the AI story "bailed," and that miners are near their July–August lows. This is a retail-focused show, so treat it as sentiment, not fundamentals.

Why it matters: Physical uranium prices and contract terms keep improving while the share prices fall. That gap is the trade in this part of the chain. The catalyst to watch is the first direct uranium contract signed by a hyperscaler.

The debate

The bull case: one supercycle across the whole power chain. The order books are real and stretch out for years. Caterpillar is selling into 2030, INNIO is sold out through 2027, compression lead times are at 200 weeks, and a16z hears of data-center parts unavailable until 2028. Spending is heading past $1 trillion a year from 2027. The a16z partners also pushed back on the affordability fear. They cited a US study that "for every 10% increase in data center capacity, residential rates went down by 40 bps." Their logic is that a big, steady customer spreads the fixed cost of poles, wires and substations over more power sold. On Trader Talk, "The $3T AI spend, grid bottlenecks, and the indexing trap" (Sep 30), the hosts cited S&P figures: data centers use about 5% of the grid today, heading to 15–20% by 2030, on a grid "20 to 50 years old." Their conclusion was that this is "why natural gas stocks have done very well."

The bear case: politics, credit and the spare capacity already in the grid.

• The voters are turning. On Sustainability Currents, "Green Finance Backs Data Centers" (Sep 30), Crédit Agricole's sustainable-finance team cited an August University of Pennsylvania study. About 60% of US adults now oppose new data center development, "up 12 percentage points from beginning of the year." They expect pushback to rise in power-constrained markets and said "community opposition... can really delay projects by years." Sherrill's own account, of bills up by double digits, sits awkwardly next to a16z's 40-basis-point claim.

• The grid has more room than the buildout assumes. On Open Circuit, "Utilities and Big Tech lost trust. What will earn it back?" (Sep 29), the host noted the US grid "runs out at about half capacity." He cited a Brattle Group estimate of more than $100 billion in savings from using existing wires better. Guest Caroline, a former Google energy executive, said that four years ago "at Google, we started telling our regulators, there's not enough capacity." She described utilities' first response as trying to "rent-seek off the data centers at an astronomical rate." On Clean Power Hour, "The Battery Boom Is Moving Beyond EVs" (Sep 29), the hosts set 230 GW of data-center interconnection requests (requests to connect to the grid) against a US grid that runs at about 500 GW. They calculated that 20 million Tesla Powerwall 3s, home batteries in about one in five homes, would cover it. Both points challenge the "build everything new" story that supports regulated utility spending.

• The returns don't add up yet. On Odd Lots, the hosts cited an economist's estimate that the spending would need revenue "like 9% of GDP" to justify it. The Canadian Investor hosts noted hyperscaler returns on invested capital are falling, and that management teams "do not give you an answer on where they think ROIC is going." Then there is Oracle.

My view: the bear case this week is not "demand disappears." It is "demand moves." Sherrill's rules, Oracle's force majeure and the engine backlogs all point the same way, toward power that data centers build or buy for themselves. That is bullish for engines, compression and on-site gas. It is more mixed for regulated utilities whose growth plans assume data-center load gets added to the shared grid.

The names in play

• Caterpillar (CAT). The clearest operator signal of the week. It is restarting an engine line it shut for lack of demand, taking orders into 2029–2030 and converting plants to add capacity. Its own CEO says production, not demand, is the limit. The next thing to watch is whether the restarted medium-speed engine ships on time in Q4 2026.

• USA Compression (USAC), Archrock (AROC), Kodiak (KGS). These are the overlooked winners. When new compressors take four years to arrive, the existing fleet is worth more and pricing power goes to the owner. Kodiak has delivery slots booked through 2029.

• Oracle (ORCL). The bear case has a specific document now. Watch whether a second developer or tenant uses the same clause, which the Enterprise AI Show hosts called the line between "a one-off" and "a trend."

• Cameco (CCJ). Huhn says contract terms favor producers more than at any point in this cycle, while the shares trade near their summer lows.

• Bloom Energy (BE). On Odd Lots, the guest named it among stocks priced on three-year forward sales estimates. In other words, its valuation depends on the hardware boom continuing.

Read-throughs

• Gensets and engines → compression and gas producers. Engine makers are giving capacity to data centers first, so gas producers and pipeline companies wait longer for compression. That could slow production growth in the gas fields that supply both LNG exports and power plants. It also supports the companies that lease compressors.

• Pipelines → data-center developers. Jupiter shows that one gas pipeline can hold up a multi-billion-dollar campus. Sites with gas already secured, or with generation on site, should be valued at a premium.

• Copper and specialty metals. On The Bid, "AI Beyond Tech: How Artificial Intelligence Is Reshaping the Economy" (Oct 2), BlackRock estimated about 121 GW of data-center power will be needed by 2030, at $40–50 billion per gigawatt. It called copper "a common denominator" across generation and distribution. It also flagged indium, a byproduct of zinc mining that goes into optical networking, as a less obvious choke point. Read-through: Freeport (FCX) and the cable and wire makers.

• Hyperscaler credit → everything downstream. Goldman underweighting hyperscaler bonds and the September pause in new issues are worth watching. Every PPA (power purchase agreement), uprate and pipeline in this newsletter ultimately depends on these companies' balance sheets and their willingness to borrow.

• The cost per token. On Monetary Matters, "Token Bill Dwarfs CPU Bill | Ben Pouladian on Meta's Muse, Anthropic S1, and Why All Roads Still Lead to NVIDIA" (Sep 30), Ben Pouladian argued the key metric is becoming "what does it cost to extract a billion tokens per megawatt." He said "energy is a big input," and that the stack has to be optimized "all the way from the power that's going into the data center." Once investors start pricing AI companies on cost per token, cheap, firm power becomes a competitive edge and not just a line item.

What changed since last issue

• Oracle Jupiter went from schedule slip to force majeure. Last issue it was a pipeline construction date sliding to February 2027. Now Oracle is using the delay to try to cut its lease payments, and the project is pushed back at least six months.

• The equipment shortage widened beyond turbines. Last issue's skeptics said new turbine makers would arrive around 2030. This week, engine and compressor makers put their own sold-out dates out to the same year.

• Backlash moved from polls to policy. Last issue had project cancellation counts and Gallup polling. This week a sitting governor in PJM described a separate rate class, bring-your-own-energy rules and an enforcement action.

• Uranium is unchanged at about $89–90 spot. The new detail is in contract design: no-ceiling contracts and an implied forward price of $110–115.