Build and Serve

The power chain this section has built has a hole in the middle of it. Three thesis pieces covered what the AI buildout needs once the electricity exists: transformers to step it up and down (The Transformer Bottleneck), copper to carry it (The Copper Squeeze), and cooling to throw off the heat it becomes (The Heat Wall). Every one of them took the electrons as given. But the most binding near-term constraint on AI is not moving the power or cooling the chips, it is generating the power at all. For new dispatchable capacity that can be stood up on a data center's timeline, the answer through 2025 and 2026 has overwhelmingly been the natural-gas turbine. Nuclear is too slow to permit, renewables will not give you firm 24/7 output, and the grid interconnection queue is years deep. So the hyperscalers are buying gas. This piece is about the generator, and it turns out to be the cleanest case yet for the lens this section corrected itself onto six weeks ago.

The demand is not a forecast, it is already on the order books. In October 2025 OpenAI and Oracle placed the largest onsite-gas order ever, a 2.3 GW fleet for Stargate sites in Texas and New Mexico. xAI's Memphis supercomputer runs on dozens of gas turbines trucked onto the site. Exxon and Williams are standing up multi-gigawatt gas-for-data-center projects on ten-year contracts. GE Vernova now says 20% of its 100 GW gas backlog is explicitly tied to data centers. The framing is AI; the function underneath it, this cycle, is a gas-fired build-out, and it is large.

Three makers, sold out to the end of the decade

The supply side is an oligopoly of a cleanliness that almost never occurs. Only three firms build large heavy-duty (F, H, and HA-class) gas turbines at scale: GE Vernova, Siemens Energy, and Mitsubishi Power. Together they are more than 70% of production capacity and about two-thirds of the turbines in the gas plants currently under construction. The barriers are the realest kind there is, decades of combustion and metallurgy IP, guarded obsessively, that China has spent years and serious money failing to match at the 500-to-600 MW frontier. And right now all three are sold out. GE Vernova's gas backlog and slot reservations reached 100 GW in Q1 2026, it says no new unit ships before late 2028, and only about 10 GW of capacity is left across 2029 and 2030. Siemens Energy carries a record order backlog near €154 billion and just booked its highest-ever quarter of gas orders. Mitsubishi is sold out into 2028 and quoting waits as long as seven years. Lead times like that are the bottleneck in its purest, most investable-looking form.

Why I will not buy the bottleneck

And this is exactly where the section's corrected lens earns its keep, because the obvious trade (a sold-out oligopoly, so buy the makers) is the one the lens says to distrust. "What Stays Scarce" established that durable value goes not to the bottleneck but to durable, appropriable scarcity, and that most manufacturing capacity is the opposite of durable: it can be overbuilt, so it is cyclical. A gas-turbine order book is the textbook case, and we do not have to imagine the bust because the industry already lived it. The last time turbines were sold out, the merchant-power boom around 2000 and 2001, global orders peaked near 108 GW a year and GE alone shipped more than 60 GW in 2001. Then the forecast demand never arrived. The market value of gas-turbine production fell from roughly $32 billion in 2001 to almost nothing by 2003. Enron, NRG, Mirant, and Calpine went bankrupt; Calpine's plants ran at 45% capacity; the industry kept adding capacity into the glut. That is the fiber-2001 structure precisely, the same shape I flagged for transformers, and the makers themselves are visibly scarred by it. All three are expanding only cautiously, and Siemens Energy has said outright that it has no intention of building beyond its current footprint. The sold-out backlog everyone is excited about is the same object that nearly destroyed the industry a generation ago. I will not pay a premium for the order book.

What I will buy: the installed base, not the backlog

What survives a turbine bust is the thing the lens points at: the installed base. Once a turbine is spinning in a plant it needs the maker's parts and service for twenty to thirty years, under long-term service agreements that each run a decade or two, and that revenue does not care about the order cycle. The numbers are stark. Across 2024 to 2034 the industry is projected to spend about $261 billion on gas-turbine services against roughly $137 billion on new turbines: the aftermarket is nearly twice the equipment. The OEMs capture more than 62% of it. For GE Vernova, the Power segment already earns about half its revenue from service on a fleet of more than 7,000 turbines, at 17-to-19% segment margins, and the long-term service agreements throw off the majority of the segment's operating profit. The decisive detail is what happened in the last bust: when new orders fell 99% between 2001 and 2003, the installed base still had to be serviced, and aftermarket revenue is the part that held. So the durable, appropriable, oligopoly-protected scarcity here is not the ability to build a turbine. It is the lock on servicing the ones already built. The real gift the AI boom is handing these three companies is not the order spike that will probably overbuild. It is that the boom is bolting down a whole new generation of turbines whose service contracts will pay high-margin cash for two decades, straight through the next bust. Build-and-sell is the cyclical headline. Build-and-serve is the durable business, and it is the OEMs' own phrase for the shift.

The signals

That distinction sets the calls. GE Vernova (GEV), already a BUY in the transformer book at a roughly $250 billion cap, is reaffirmed here for the refined reason: it is the most service-heavy and most US-liquid way to own the annuity, with the grid-electrification franchise stacked on top. Own it for the installed base and the electrification leg, not for the equipment backlog, and accept that the stock is priced richly after its run. The overbuild risk is real, but it lands on the equipment line, which is the smaller and lower-margin half of Power.

Siemens Energy is the purest leverage to the gas-service build-out: the record backlog, the fastest gas-order growth of the three, net income that recently tripled, a roughly $177 billion cap. The catch is genuine. You also buy the troubled history of its Siemens Gamesa wind unit, and the cleanest US access is a thin OTC ADR (SMEGF, near $180); the liquid line is the Frankfurt-listed ENR.DE. It is a turnaround with an annuity engine inside it, and I rate it BUY with that ADR-liquidity caveat stated plainly.

Mitsubishi Heavy Industries is a HOLD. The turbine franchise is real and sold out into 2028, but turbines are one division inside a sprawling industrial-and-defense conglomerate of about $80 billion, so the exposure is the most diluted of the three, and the cleanest US access is again a thin ADR (MHVYF, near $23). The thesis is present; the vehicle is the muddiest.

And the thing to AVOID is the instinct this whole piece argues against: do not buy any of the three for the sold-out order book. The backlog is the 2001 setup wearing a 2026 suit, and the capacity all three are now adding lands in 2028 to 2031, exactly when data-center demand would have to keep compounding to absorb it. Own the service annuity. The backlog is a trap with a date on it.

What this completes

This is the fourth bottleneck and the last obvious one in the physical power chain: generation, then transmission step-up (transformers), then conduction (copper), then heat rejection (cooling). Read together, the lens now sorts the whole book the same way at every layer. The piece of the chain that looks scarcest in the moment (transformers, turbines) is usually the overbuild trap, and the durable money sits in the scarcity that cannot be competed away: copper's geology, Vertiv's engineering lead, and now the turbine makers' lock on servicing their own installed base. The framing always points at the shortage. The function is always one layer over, in whatever the shortage leaves behind once it clears. Here, what it leaves behind is seven thousand turbines that only three companies on earth are allowed to fix.