Kedrosky Verbatim.

Stranded Assets and the AI-Driven Gas Turbine Renaissance.

---

What follows (see below) was written by Paul Kedrosky, a fellow blogger back in the days when people blogged on Blogger.

Paul is an investor in private and public companies, as well as a writer and researcher. Originally trained as an engineer, he went on to do a Ph.D. at the University of Western Ontario in Canada, where he researched aspects of the economics of technology—specifically, the role of path dependency and network effects in risk & complexity.

Paul regularly speaks at private and public events across the U.S. and around the world, usually on topics related to risk finance, economics, the future of work, and artificial intelligence. He is currently a research fellow at MIT’s Institute for the Digital Economy, where he is studying artificial intelligence, economic disruption, and the future of work.

You can find his daily note at paulkedrosky.com. We strongly suggest you subscribe. It’s always worth reading.

---

What’s Happening:

AI has flipped the global gas-turbine market from slack to locked-in:

• Lead times: Now 5–7 years for large turbines.

• Order books: OEMs (Mitsubishi, GE, Siemens) say they are fully committed to ~2030–2032.

• Prices: Turbine costs are up 2x in some categories.

• Driver: AI/data centers projected to take ~12% of U.S. power demand by 2028 vs ~4% in 2023.

• Customer mix: Hyperscalers are crowding out utilities and emerging-market buyers for the same hardware.

The key point: this is forward-committed demand—capacity pre-sold years ahead based on today’s AI-energy nexus narrative.

---

Why It Matters
This “locked-in” status is a risk amplifier.

• Path dependency:

  • OEM capacity is now structurally pointed at AI-centric U.S. projects, not system-wide decarbonization.

  • Even if sentiment about AI moderates, the hardware pipeline is already committed; reallocating is slow and costly.

• Global crowd-out:

  • Southeast Asia: Coal-to-gas transitions are delayed because they are behind U.S. tech in the queue.

  • Europe: Projects need state aid + regulatory approval, then still hit multi-year OEM bottlenecks.

  • Result: coal and older thermal stay on longer because replacement hardware simply isn’t available.

• Long-lived lock-in:

  • These plants run 20–30 years.

  • Decisions made in a 2024–2026 AI enthusiasm window hard-wire gas capacity and emissions into grids well into the 2040s.

    ---

Yes, But
There are caveats—but they mostly sharpen the risk.

• “Hydrogen-ready” is mostly optionality, not reality

  • Claims of 30–50% H₂ blends now, 100% by 2030 are technically plausible but commercially constrained by fuel availability and cost.

  • In practice, these assets are gas plants with a marketing story, not guaranteed future-zero-carbon infrastructure.

• Sentiment vs fundamentals

  • The AI build-out is being treated by buyers as a straight-line growth story; that’s how you justify ordering turbines for 2031–2032 delivery.

  • If AI power demand plateaus or shifts (efficiency gains, different architectures, better load shifting) even slightly, the system is left with overbuilt, CO2-emitting, geographically-misaligned gas capacity.

• Regulation and politics still bite

  • Especially in Europe, permitting + state-aid rules could delay or reshape projects, increasing the risk of stranded or underutilized assets relative to today’s assumptions.

    ---

What It All Means
Energy risk in AI is real, and getting worse in unexpected ways.

• Risk is being front-loaded and then frozen into hardware.

  • OEM backlogs into the 2030s are sentiment-contingent commitments being sold as inevitabilities.

  • If the AI arc bends differently, you don’t mark-to-market easily; you live with 30 years of capacity and emissions.

• Decarbonization is being reordered by who can pre-book steel.

  • The U.S. AI sector gets firm power.

  • Emerging markets get delayed transitions.

  • Europe gets a harder reliability problem with slower firm-capacity additions.

• The signal is:

  • The constraint is now OEM and engineering capacity, not capital.

  • The biggest mispricing may be how much optionality we’ve already given up by turning an AI capex hype window into decades of committed thermal infrastructure.