Last week we covered photonics testing. The week before, photonics testing. This week the bottleneck is about gas turbines.
It is well known now that power infrastructure is struggling to keep up with this massive AI buildout.
U.S. electricity demand is growing for the first time in two decades, and the grid is the slow part. Interconnection queues now run years, and new high-voltage capacity is its own backlog.
So the hyperscalers stopped waiting. Told to "bring your own power," they are increasingly pairing data centers with on-site generation, behind the meter, sized to the campus instead of the queue.
One of the fastest way to put firm megawatts on-site is a gas turbine, which is the whole problem. Every hyperscaler reached that conclusion at the same time, so the machine that was meant to be the quick fix has become the scarce one.
Inside the gas turbine
At its core, a heavy-duty gas turbine does something simple- it burns natural gas to spin a generator.
Air is pulled in and compressed, mixed with fuel, and ignited, and the hot exhaust rushes through rows of blades that turn a shaft, which drives the generator. The best plants go one step further and capture the leftover exhaust heat to raise steam for a second, steam-driven turbine, which is the combined-cycle design that pushes efficiency past 60%.
The result is a machine the size of a locomotive, weighing around 500,000 pounds, that can power roughly half a million homes.
A large frame gas turbine (Source: artofit.org)
It is the only dispatchable source you can build on a two-to-three-year clock. Nuclear is at least 5 years out. Renewables can't supply firm baseload alone. Gas is the bridge.
And only three companies build the big turbines required for data centers:
GE Vernova
Siemens Energy
Mitsubishi
They account for about three-quarters of the world's large-frame turbines. The roster has only shrunk. This is a business everyone pronounced dead a few years ago.
Demand is off the charts
The order books have gone vertical, and the OEMs are saying so on every call.
The backlog runs years deep: GE Vernova alone closed 2025 with roughly 80 GW of gas turbines booked or reserved, enough to keep its lines full into 2029, with management expecting every open slot to be spoken for through 2030 by the end of this year. Siemens Energy is sitting on the largest order book in its history.
And it's still accelerating: This isn't a one-time spike. GE's order intake climbed by about a third last year, with quarterly unit bookings running well above the year before. Siemens nearly doubled the turbines it sold in a single year and almost all of it, it says, pulled by data-center power.
Scarcity is showing up as price: A gas plant costs far more to build than it did two years ago with turbine equipment alone up something like 195% over 2019, and the turbine can be nearly a third of the whole plant's cost. When buyers are this desperate, the seller sets the number.
And the expansion doesn't fix it: GE Vernova is spending heavily to lift heavy-duty output from around 50 units a year toward 80, and Siemens is adding U.S. capacity too. It sounds enormous until you set it against the book. Flat out, GE will ship about 20 GW a year but the backlog is four times that and still growing. The line barely shortens.
Gas Turbine constraint is scored as Tight on Tessara with no near term relief.
The blade is the upstream constraint
Those expansions are almost all at the assembly end, meaning bigger halls, more cranes, more test bays.
But a turbine can only be finished as fast as its hardest single component shows up, and that component is the first row of blades, the ones sitting directly behind the flame.
Those blades have to survive gas hotter than the metal would normally tolerate, and the way engineers manage it is remarkable. Each blade is not forged but grown as a single, unbroken crystal of nickel superalloy.
The molten metal is cooled so precisely that the entire part hardens as one continuous grain, leaving no internal boundaries for heat and stress to crack along. From there, each blade is threaded with hair-fine cooling channels and sealed under a ceramic thermal-barrier coat.
The work has more in common with growing the silicon crystals that become computer chips than with ordinary metalwork. And it shows up in the bill, where the airfoils alone run about 35% of the engine's raw-material cost.
The OEMs mainly rely on a few specialist firms that make this. Two American companies dominate the market:
Precision Castparts (privately owned by Berkshire Hathaway)
Howmet Aerospace (publicly traded)
Between them, they hold an estimated 70–80% of the world's high-end single-crystal blades.
Why they stay constrained
The blade makers are expanding but slowly, and on their own terms.
They build slowly, and only against signed contracts: Howmet expects its gas-turbine business to grow from roughly $1 billion in 2025 to more than $2 billion over the next three to five years. That is rapid growth in industrial terms, but it is still a multi-year buildout. The company has been signing investment agreements with turbine OEMs before committing additional capacity.
Aero wants the same capacity: The same specialist companies that serves land-based turbines also serves jet engines, and commercial aerospace is already running through a strong aftermarket cycle. Every incremental power-turbine order therefore competes for finite high-spec casting capacity, tooling, labour and machine time.
The physics caps the pace: Single-crystal yields fall as blades get bigger, and power-turbine blades are the biggest there are. More scrap, longer runs, a batch that can take over a year. And any new line has to be proven out on a part that must survive a flame for decades, which itself takes years. You can't buy your way around the calendar.
Howmet is the largest turbine-blade maker in the world, with more than half the global market and it has told investors plainly that demand has now run past even its recent capacity additions. When the company at the center of the gate says it can't keep up, the gate is real.
The shortage itself is largely priced in. Howmet has run up about 30% this year to roughly 57x forward P/E, and blades are barely a tenth of its revenue, so the stock mostly tracks the aerospace cycle. If data-center demand keeps climbing, whoever holds the casting layer keeps the pricing power.
Who pays, who captures
Pays rent: Power developers building for hyperscalers and the utilities and merchant operators in between. Plus the three turbine makers themselves, all of whom need blades.
Captures rent: Precision Castparts, locked inside Berkshire, and Howmet in the US. Between them, they make the large majority of the world's single-crystal blades.
What to watch
Howmet Q2 Earnings: New capacity, and contract conversions talks, the read on whether the casting gate is widening or holding.
OEM next reports: Whether shipments actually hit their ramp targets. GE toward ~20 GW a year by 2027. If they lag their own guidance, the blade might be the bind.
To go deeper on power, this is a strong read on how behind-the-meter generation is reshaping the AI race.
The short version: the U.S. grid cannot keep up with AI’s power demand so the buildout is moving behind the meter.
The Week Ahead
A few important earnings this week. Read our post and pre-call briefs here and stay prepared.
Earnings Tab on Tessara
Wednesday, June 24
MU (Micron). We just published our forecast for Micron’s Q3 FY2026 earnings and how Tessara came to it. We’re showing the method, making forward calls, and letting our record accumulate in public. Worth a quick read before the earnings call this evening.
Want the live data behind The Chokepoint? Tessara is the research terminal for the AI buildout. We track what's binding in the supply chain and what it means for what you own. 400+ companies across compute, memory, foundry, networking, and power.
See you next week,
Teng & Arvind
This article is for informational and research purposes only. It is not financial advice, investment advice, or a recommendation to buy or sell any security. Tessara Research does not publish price targets. The views expressed here reflect our analysis at the time of publication and may change as new evidence arrives. Readers should do their own research and consult a qualified financial adviser before making investment decisions.