You have probably seen this name on your feed for a few weeks now. Peter Thiel bought it. The CEO bought it. Energy is the bottleneck. AI needs power. Nuclear is scarce. The stock is down anyway, which somehow proves the point.

All of that is true, and none of it is an explanation.

The Thiel part is real and pretty interesting. Thiel Macro reported zero US equity holdings for two straight quarters, then filed a 13F on August 14th showing an eight-name, $419 million book. Seven of the eight are power, utilities, or energy. Vistra is $59.1 million of it, about 14% of the fund. The one exception is Amazon, which is most widely reported as the counterparty on Vistra’s largest nuclear contract, though Vistra has never confirmed it.

The CEO part is also real, but smaller than it sounds. Jim Burke bought 2,000 shares on the open market on August 24th at $135, about $270,000. A real buy rather than an option exercise, which counts for something. It is also 2,000 shares against roughly 1.1 million he already holds through a family partnership.

Here is what nobody on your timeline has done: opened the customer. Not the power company, the customer. What is a data center, physically, as a thing that buys electricity? How many hours a day does it draw? What fraction of its cost is the power bill? Why did Amazon pay about $95 per megawatt hour for nuclear when a brand new gas plant makes electricity for roughly $90, and why did Meta pay less than that for the same fuel from the same company a few weeks later?

In our Energized write-up, we argued the market was paying for announcements and refusing to pay for delivery. Vistra is the sharpest single expression of that. It has more contracted nuclear than any company in America and its stock is down about 35% from last September.

The thesis compresses to three sentences. A data center is not a bigger customer, it is a differently shaped one, drawing the same power at 3am in April as at 5pm in August, on a grid that does not. That shape, not the price of electricity, is how AI reaches this company’s income statement. And the market is pricing the 3% while ignoring the 97%, because the premium above market that the Amazon and Meta contracts add is roughly 3% of 2027 earnings, while everything else is capacity payments, retail, and running plants that already exist for more hours of the year.

What a data center actually is

Megawatts tell you how fast, not how much

Every note on this sector quotes megawatts. Announced megawatts, queued megawatts, contracted megawatts. Almost none of them stop to say what a megawatt is, and the distinction turns out to be the whole trade.

A megawatt is a rate. It is a speed, like miles per hour. It tells you how fast electricity is being delivered at a given instant, and nothing about how much gets delivered over a day or a year.

A megawatt hour is an amount. It is the distance actually traveled: one megawatt, sustained for one hour. A gigawatt is a thousand megawatts, roughly the delivery rate of one large nuclear reactor.

A power plant’s size is quoted in megawatts, which is its top speed. What it earns depends on megawatt hours, which is how far it actually drove. Those two numbers are related by one thing: how many hours a year the plant is allowed to run. That number is where this entire write-up lives.

The same distinction applies on the customer side. A data center described as “one gigawatt” is telling you its top speed, not its consumption. To get consumption, you need to know what fraction of that gigawatt it actually pulls, hour by hour, across a year.

Load factor, the one word that reorders the sector

That fraction has a name. Load factor is the share of its maximum draw a customer actually pulls, averaged over a year. A customer with a 50% load factor consumes half of what it would consume if it ran flat out constantly.

The numbers are not close:

• Data center, as operators plan it: 90%

• Data center, as utilities actually observe it: 80% to 85%

• E3’s modeling assumption: 86%

• Duke Energy’s planning assumption: 80%

• The US grid overall: 55% to 60%

Think about why. A house is a mountain range. Nothing much overnight, a bump in the morning, a spike when everyone comes home and turns on the air conditioning at 6pm. A steel mill is lumpy, running when there are orders and idle when there are not. An office building is dark half the day and every weekend.

A data center is a flat line. The servers do not know it is 3am. Training runs do not pause for the weekend. Inference traffic dips a little overnight, and the cooling load moves with the weather, but the shape barely bends. It is the most boring electrical customer ever invented, and boring is the entire point.

What that does to Texas

Now put a flat line on top of a mountain range.

ERCOT, the Texas grid, set a record peak of 91.3 gigawatts on July 22nd. Peak means the single highest instant of the year, the hottest afternoon in July when every air conditioner in the state is running at once. At a 57% load factor, the grid’s average draw across all 8,760 hours of the year is roughly 52 gigawatts.

Add 10 gigawatts of data center running at an 85% load factor and watch what happens to each number separately.

Peak goes from 91 to 101 gigawatts, up 11%. The data centers add to the hottest afternoon, like everyone else.

Average goes from 52 to 60.5 gigawatts, up 16%. Because the data centers are also drawing at 3am in April, when almost nothing else is.

The system load factor moves from 57% to 60%.

Energy grows faster than peak. That is not a forecast, it is division. And it matters because peak and average are paid for by completely different things. Peak capacity is what you build new plants for, and it is expensive and slow. Average energy is what you serve by running existing plants for more hours.

Vistra’s chief executive has been saying exactly this on every call for two years, in language so flat that most people read it as spin. His formulation: overall load growth outpacing peak demand, resulting in higher utilization across the system rather than short-duration peaks alone, and therefore the economics of existing generation assets improving on a sustained basis.

He is not spinning. He is describing load factor.

The trade is utilization, not price

Capacity factor is the same number, seen from the plant

Flip to the generator’s side of the meter, and the same fraction gets a different name.

Capacity factor is the share of its maximum output a power plant actually produces over a year. A 1,000 megawatt plant running flat out for all 8,760 hours would produce 8,760,000 megawatt hours. If it actually produces 5,256,000, that is a 60% capacity factor. It spent the equivalent of 40% of the year switched off.

Vistra’s combined cycle gas fleet, which will be about 26 gigawatts once the Cogentrix acquisition closes, runs at roughly 60%. Burke put that on the record himself.

So the question that decides this stock is not what a megawatt hour sells for. It is how many hours those plants get to run.

Why a gas plant sits idle 40% of the year

This explains why the AI load lands where it lands.

Every hour, the grid operator stacks up every generator that wants to run and starts with the cheapest ones. This is called the merit order. Solar and wind bid near zero, because sunshine and wind cost nothing once the panels and turbines are paid for, so they clear first whenever they are producing. Nuclear runs constantly because it is expensive to turn off and its fuel is cheap. Coal and gas fill whatever demand is left over.

Which means a gas plant’s capacity factor is set by subtraction. It runs in the hours when demand exceeds what the free and near-free resources can cover. Midday in spring in Texas, when solar is flooding the grid and demand is mild, a gas plant is unnecessary and sits idle. On a July afternoon, or at 9pm when the sun is gone, it is essential.

Now add a customer that draws the same amount at 3am in April as at 5pm in August. Solar contributes nothing at 3am. Wind contributes whatever it feels like. The extra flat hours have to come from something that can run continuously, and in ERCOT and PJM that means gas, coal, or nuclear.

The plants that pick up those hours are the ones already built and already connected to the wires. That is the mechanical link between “AI needs power” and Vistra’s income statement, and it runs through hours, not price.

What management has actually said, and the sentence that matters

Burke’s stated target for that fleet is the mid-80% range, up from 60%. That is a twenty-five-point move.

The sentence that turns this from a hope into a thesis is different: strong market fundamentals can drive higher volumes and revenue without significant incremental investment.

Read that literally, because it is literal. Running an existing gas plant more hours each year requires no new turbine, no new permit, no new transformer, no new interconnection agreement, and no place in a seven-year queue. The plants exist. The staff is already employed. The maintenance is already in the budget. The only incremental cost is fuel, and the spark spread math below already nets that out.

That is a very unusual growth engine. Almost every other way to sell more electricity in America right now requires steel in the ground and a decade of permitting.

What twenty-five points is worth

Here is the calculation explained simply.

One percentage point of capacity factor across 26 gigawatts is 26,000 megawatts times 8,760 hours times 1%, which is 2,278 gigawatt hours a year of extra electricity.

Now, what is a gigawatt hour worth to Vistra? Not the full price of electricity, because they have to buy gas to make it. The relevant number is the spark spread: what a megawatt hour sells for, minus the cost of the gas burned to produce it. It is the gross margin per unit. Vistra discloses its own for 2027: $22.40 per megawatt hour in Texas and $29.10 in the East. Call it $25 blended.

But you cannot apply $25 to the new hours, and this is where most bullish versions of this argument go wrong. Vistra already runs the profitable hours. It is not leaving money on the table at 60%. The hours between 60% and 85% are precisely the hours that were previously uneconomic, the ones where power was cheap enough that burning gas did not pay. Those hours will clear at a lower spread than the average.

So halve it. Use $12.50.

Twenty-five points, 56,940 gigawatt hours, at $12.50 per megawatt hour, is roughly $712 million of additional gross margin a year, from assets that are already built and already paid for.

For scale: the Cogentrix acquisition cost $4 billion and adds about $552 million.

The utilization gain is worth more than the largest acquisition in the company’s history, and it requires no capital.

It is already in the numbers

None of the above is a projection. It is visible in the last quarterly filing.

Texas combined-cycle capacity factor rose from 51.5% in the first half of 2025 to 56.2% in the first half of 2026. In the second quarter alone, it rose from 54.8% to 62.5%, a 7.7-point jump, meaning the same physical plants produced roughly 14% more electricity than a year earlier. Texas coal went from 49.6% to 60.0% in the quarter, though part of that is the Martin Lake Unit 1 restart after a 2025 outage, so read the gas number as the clean one. Total generation across the fleet was 50.2 terawatt hours in the quarter against 45.6 a year ago, up 10%.

So far this is a Texas story. East combined cycle was flat at 57.6% against 57.2% for the half, because PJM plants were already running harder and the East fleet grew with Lotus. The lift is showing up first where the grid is tightest.

Now the part that matters. All of that happened while the ERCOT 2027 forward curve fell 19%.

Prices down. Volumes up. Margin up. That is the physical signature of new load being absorbed by existing plants rather than new entrants, and it explains the segment results better than any price story does. Management’s own phrasing for the same thing on the second-quarter call was optimizing the run profile of their flexible gas generation assets to capture margin opportunities.

Nobody argues about the power price

What it costs to run a gigawatt of AI

To understand why hyperscalers pay what they pay, you have to see the bill they are actually paying.

Epoch AI published a full cost model of a one-gigawatt AI data center in May. Up-front capital of $38 billion. Annual operating cost of $0.9 billion. Annualize the capital over the equipment’s useful life, and the total cost of ownership is about $8.5 billion a year.

Servers and GPUs account for $5.0 billion of that, or 60%. Buildings, cooling, networking, and staff are another $2.9 billion. Electricity is $0.6 billion, about 7%.

Other estimates put electricity at 10% to 20% of a GPU cluster’s cost, around 15% of a facility’s, and about 11% of the cost of a rented GPU hour. Call the range 7% to 15%, and use the low end for frontier AI, because the chips have become so expensive that everything else shrank as a share of the total.

Where Vistra's contracts land on that bill

Now price the actual contracts against that stack. A gigawatt at a 90% load factor consumes 7,884 gigawatt hours a year, so every $1 per megawatt-hour of contract price is about $7.9 million a year.

At $69 per megawatt hour, the cheapest nuclear deal anyone has signed, the annual power bill is $544 million, or 6.4% of the $8.5 billion.

At $95, which is Comanche Peak, it is $749 million, or 8.8%.

At $150, a price nobody has paid, it is $1.18 billion, or 13.9%.

So the entire distance between the cheapest nuclear contract in existence and a number well above anything signed is about seven percentage points of the customer’s total bill.

Now the other side of the ledger. At $8.5 billion a year, a finished but unenergized gigawatt burns roughly $708 million a month in idle capital. The GPUs are bought, the building is built, the staff is hired, and nothing is earning.

One month of delay costs more than a full year of paying an $80 per megawatt hour premium over the cheapest deal in the market.

That single comparison explains every strange thing about how these contracts price. The buyer is not price sensitive. The buyer is date sensitive.

The four nuclear contracts, decoded

Four nuclear contracts have been signed with hyperscalers. Laid out side by side, they look like a random scatter of prices for physically identical electricity. They are not random at all.

Two words do all the work here.

Behind the meter means the customer sits on the same site as the plant, connected directly, with the power never entering the public grid. Front of the meter means the plant sells into the grid and the customer buys from the grid, even if a contract links the two on paper. Behind-the-meter is faster because you don’t have to wait for grid upgrades, and it is scarcer because it requires the plant and the customer to physically share land.

Additionality means whether the deal adds new electricity to the system or just reassigns electricity that already existed. Restarting a shut reactor is additional. Signing a contract on a reactor that has been running for thirty years is not.

Now read the ladder. Talen and AWS at Susquehanna, front of meter after being restructured out of a behind-the-meter arrangement, at about $69. Vistra and Meta at Perry and Davis-Besse, front of meter, at about $84 to $89. Vistra and AWS at Comanche Peak, behind the meter with the campus on Vistra’s own land, at about $95. Constellation and Microsoft at Crane, a restarted reactor, at about $100 to $102.

The step from Meta to Comanche Peak is roughly $10 per megawatt hour. That is the price of a shared fence line. The step from Comanche Peak to Crane is roughly another $6. That is the price of new supply.

Price depends on adjacency and additionality. It was never a function of electrons.

Which means the widely repeated take that Vistra left money on the table at Comanche Peak, because institutions wanted $120 and got $95, misreads the product. Vistra sold the middle rung. Constellation sells the top rung because it is bringing a dead reactor back. Talen sells the bottom because it converted its deal to ordinary grid supply.

Why Meta paid less than Amazon

There is a specific, physical reason the Meta deal priced below Comanche Peak, and it is a map.

PJM, the grid covering thirteen eastern states, is divided into transmission zones. Power moves freely inside a zone. Moving it between zones means using the transmission system, paying for it, and facing congestion.

Vistra’s three PJM nuclear plants, Perry, Davis-Besse and Beaver Valley, sit in the ATSI zone, which is FirstEnergy territory. Meta’s Prometheus supercluster is in New Albany, Ohio, which sits in the AEP zone.

Different zones. Power can’t flow behind a meter from one zone to the other. It goes into the grid, and Meta draws from the grid, which is why Meta’s own announcement says the projects deliver power to the grids that support its operations, rather than to the buildings.

Note which of Meta’s other nuclear partners is in the right zone. Oklo, whose 1.2-gigawatt Aurora campus in Pike County sits in AEP, same as the data center.

So Meta bought 2,176 megawatts of existing output from Perry and Davis-Besse, 433 megawatts of future uprates across those two plants and Beaver Valley, clean energy attributes, twenty years of price certainty, and a great deal of political goodwill in Ohio and Pennsylvania. It did not buy speed, because at that distance speed is not for sale. It paid $84 to $89 accordingly.

2300 Coates Road, Granbury

The Comanche Peak deal, by contrast, has a street address.

Amazon’s campus, filed with Hood County as Project Spectrum, sits at 2300 Coates Road in Granbury, Texas. Roughly 1,265 acres in Hood County, with a separate 435-acre agreement in Somervell County next door. Up to 18 two-story buildings. About $5 billion of investment. Construction starting in 2027. On and adjacent to land Vistra owns.

The county deal is a Chapter 381 agreement, Texas’s mechanism for a county to share tax revenue with a developer: $2.5 million at groundbreaking, then $300,000 per building each year after.

The plant is across the county line from the campus. Both are cooled by the same body of water, Comanche Peak Reservoir, 3,275 acres and 151,418 acre-feet, built for one purpose, which is keeping a nuclear reactor from overheating. Cooling water is one of the hardest things to secure for a large data center in Texas, and this one sits next to a lake dug for exactly that job.

Somervell County has about 10,000 residents. Hood County narrowly voted down a data center moratorium, and local reporting suggests another resolution may come. That is the political risk on this specific site.

One caution we hold throughout. Vistra’s SEC filings describe the Comanche Peak counterparty as undisclosed. Hood County News, DCD, and WFAA have reported that Amazon is the counterparty, but the company has never confirmed it. Treat it as very likely and not certain.

What Vistra actually owns

Seventy sites, and the wire attached to them

Read more

Every Sunday, you get the week laid out before it starts. What is happening, what each thing actually measures, and what would have to show up for it to matter. The stuff that will move prices Monday through Friday, explained properly.

The reason we are starting this is simple. The calendar itself is free. Any brokerage app will tell you the jobs report is Friday. What almost nobody does is explain what is inside the jobs report that matters, why one line in it is worth more than the headline, or why an earnings report from a company you do not own tells you something about one you do.

That is the job. Let us get into it.

Where things stand

Read more

Every machine meant to define the next decade has the same problem, and it is not intelligence.

A humanoid robot walking through a warehouse. A car that drives itself. A delivery drone. A robot arm working next to a person instead of behind a cage. The models that run them have improved faster than almost anyone predicted. What has not kept pace is the part that comes before thinking: perceiving the physical world in real time, on a battery, in the rain, at forty degrees below zero, without ever failing in a way that hurts somebody.

That is a silicon problem. And it is a solved silicon problem, in exactly one industry, which spent the last fifteen years and enormous amounts of money solving it because regulators forced it to. The automotive industry already builds cameras that work at tunnel exits, radar that sees through fog, and processors that turn both into decisions in under thirty milliseconds, all qualified to survive twenty years of vibration and never quit.

The companies that did that work are sitting on the sensing stack for physical AI, already built and already paid for.

One of them is a $900 million company in Aliso Viejo, California, called indie Semiconductor.

Except that is not quite what it is either

Go looking, and the company gets strange fast.

There is a business outside Munich whose radar stops cranes from crushing people in an Austrian steel mill, tracks containers at the Port of Hamburg, and runs telemetry on roughly ten thousand locomotives. Last year it certified the first industrial radar in its category to the European machine safety standard, which is the exact certification a robot needs before it is allowed to work next to a human.

There is a business outside Zurich making the light source inside fiber optic gyroscopes. A fiber optic gyroscope is how an aircraft or a missile knows where it is and which way it is pointing when GPS is jammed or spoofed. Its stated markets are avionics, aerospace, navigation, and eye surgery.

There is a laser tuned to 399 nanometers, which is the precise color of light that cools ytterbium atoms almost to absolute zero. That sentence only means something if you are building a quantum computer, which is exactly who buys it.

indie owns all of it. None of it has ever been named on an earnings call. All of it is reported as a single line.

And then there is how they got into robots

They didn’t.

Robot companies went looking for a chip that could handle several camera feeds at once, in real time, without draining a battery, and that had already been proven reliable at scale. They found automotive parts sitting in a distributor’s catalog and bought them. indie found out afterward and turned the orders into relationships.

Two of the largest names in humanoid robotics are now customers of a company that has never built a robotics product, never had a robotics division, and never spent a dollar acquiring a robotics customer. The chief executive says it works because the car parts are “basically 100% compatible” with what robot makers need.

That is not a claim you have to take from a small-cap chief executive. The largest robotics program on earth is built the same way. Tesla’s Optimus runs eight autopilot-grade cameras and an adaptation of the Tesla FSD computer, the same silicon platform that processes vision data from Tesla’s vehicle fleet. Andrej Karpathy (who ran Tesla’s AI team and built much of Autopilot) put it more memorably: “The early version of Optimus once thought it was a car, because it used the same computer, the same cameras, and the same algorithms that ran on cars.”

That is the whole thesis in one sentence, said by somebody with no reason to flatter it. The robot and the car are the same perception problem. Whoever already solved it for the car is holding the parts.

Nobody paid for that. It arrived because somebody else’s qualification bill had already been settled.

What we found

We went through all of it. Every filing, every transcript, every investor deck, the subsidiary websites, the trade show listings, and one Form 8-K that the company points to in every press release and that nobody appears to have opened.

Six things came out.

The company is five businesses wearing one costume, and it says so itself. Buried in the accounting policy is an admission that indie has “multiple business activities” run by “individual segment managers” held accountable for their own results. It reports all of them as one segment anyway.

The most important number in the business stopped being published in February. We rebuilt it from what is left, two different ways, and the real figure is far better than the one everyone is quoting.

The Chinese subsidiary everybody treats as a hole in the accounts is the opposite. It is 42% of revenue and only 10% of costs. Selling it makes the path to profitability dramatically shorter, not longer, and the math is in a filing nobody reads.

The chip driving growth right now works because the world ran out of memory. It was designed to save a few dollars. Then the memory market broke, and it became the only way some carmakers could keep a production line running.

The fastest-growing product is illegal in the United States. That explains why American revenue has halved since 2023. It also means an entire market opens on the day a dormant federal rule finally moves.

And the customer behind the biggest number in every model is named exactly once, in the risk factors, and never again.

None of this is secret. It is spread across four company names, several countries, and one line item.

indie keeps solving problems it was not aiming at. That is either the luckiest small-cap in semiconductors or a pattern. We think it is a pattern, and the rest of this is why.

The sentence that started this

On the August earnings call, an analyst asked Donald McClymont where gross margins land now. You said 55%. Not long ago you hoped for 50% by year-end. Where are we?

The whole answer:

“We don’t typically guide gross margin, through the divestiture of Wuxi, which I would say is tendentially a drag on gross margin. We’re in a good spot where we can get to our corporate goals.”

Gross margin is the simplest number in a business. Take what you sold, subtract what it cost to make, and the percentage left over tells you whether you have a real product or a commodity. It is the number that decides whether growth eventually turns into profit.

The first clause of that answer is false. indie guided gross margin in every quarterly release through Q3 2025. “In the range of 46% to 47%.” “49% to 50%.” Second bullet in the press release, quarter after quarter, for years.

Then it stopped. Not just the guidance. The metric disappeared from the company’s reported figures entirely.

So the single most important number for a semiconductor business at an inflection point is no longer disclosed. Every model built on this company runs mainly on assumptions.

We rebuilt it. A few minutes of math and one sentence from a Form 8-K.

Our thought process

Three sentences.

#1: The core business earns roughly 57% gross margin, not the 45% the blended print implies. The gap is a Chinese lighting subsidiary that is 42% of revenue, runs about 30% margins, and is under contract to leave.

#2: When it leaves, the math gets far easier than anyone models. That subsidiary is 42% of revenue and 10% of operating expenses. Selling it drops quarterly breakeven revenue from $84 million to $56 million, against $40 million of core revenue guided this quarter.

#3: This is five businesses reported as one, and three of them are free. Automotive vision, automotive radar, European industrial radar, defense and medical photonics, quantum. One segment. Four brand names. Zero analyst mention for anything except cars.

• Part One: how a machine actually sees, what indie makes, who runs it, and the one idea underneath all of it.

• Part Two: the reconstruction. The missing number, rebuilt two ways, and what selling China actually does.

• Part Three: the company nobody models.

• Part Four: memory, mandates, a regulatory unlock in America, and the robot supply chain being fenced off in Washington.

Part One: The Machine

How a machine sees

Numbers mean little until you understand what is being sold. So start with the physics.

Cameras, and the problem of light

A camera sensor is a grid of light buckets. Each pixel collects charge in proportion to the photons that land on it. Read out the grid ,and you get a spreadsheet of brightness values, with a color filter on top so each bucket sees only red, green, or blue.

That is not an image. It is raw data, and turning it into something a computer can act on takes four hard jobs.

Reconstructing color. Each pixel measures only one color, so the other two must be inferred from its neighbors. Do it badly, and you get false color fringing on every edge in the frame.

Handling dynamic range. This is the hard one. Picture a car coming out of a tunnel. Sunlit tarmac outside might be 100,000 lux. A pedestrian standing in the tunnel shadow might be 10. That is a ten-thousand-to-one range in a single frame, and an image sensor captures maybe a thousand-to-one in one exposure. So the sensor takes several exposures of different lengths, and something has to fuse them into one frame that keeps detail at both ends. Get it wrong, and the pedestrian is a black smear.

Undistorting the lens. A wide-angle automotive lens bends straight lines into curves. Every pixel has to be mapped back to where it should be.

Doing all of it in time. At seventy miles per hour, a car covers a meter every thirty-two milliseconds. The whole pipeline has to finish faster than that, thirty times a second, on every camera in the vehicle.

That job belongs to an image signal processor, and it is the largest business inside indie.

Here is the part that matters commercially. A conventional ISP handles those stages one at a time, and between each stage it writes the whole frame out to a memory chip sitting next to it and reads it back. Five round trips to external DRAM, thirty times a second, per camera. The memory chip is there because the processor needs somewhere to hold an entire image while it works on it.

indie’s iND880 never holds an entire image. It processes the picture in a rolling window of scan lines, using memory built onto the die itself, and passes each line downstream as it finishes. No full frame, no external memory chip. The part that removes the requirement costs less than the requirement it removes.

For years that was a modest saving. One fewer component on the board, a little less power, a little less space. Then the memory market broke, and we will come back to what that did.

Radar, and the problem of resolution

Read more

Two Dead Smelters

Hawesville, Kentucky, sits on the Ohio River, and for sixty years its aluminum smelter did exactly one thing: turn staggering amounts of electricity into metal. At full tilt, it pulled around 480 MW, day and night, one of the largest single power draws in the state. It shut down in 2022 because electricity got too expensive to justify the aluminum.

On 6 July 2026, Anthropic signed a 20-year, $19bn lease on that site.

Five weeks later, on 11 August, Anthropic committed another $9.1bn for a bitcoin mining campus in Rockdale, Texas, built on the carcass of an Alcoa smelter that poured its last metal in 2008.

$28bn. Five weeks. Two dead smelters.

Understand what was actually purchased, because it was not buildings. A shuttered smelter is a big shed with good bones. What Hawesville and Rockdale have that almost no site in America has is a live connection to the grid, sized for hundreds of megawatts, with an interconnection agreement somebody signed decades ago. The wire is worth more than everything attached to it.

One of the most sophisticated buyers of computing on the planet just told you twice in five weeks, with $28bn, what the scarcest input in artificial intelligence is. It is not chips. Chips ship in weeks. It is not capital. Capital is throwing itself at this sector. It is a connection to power that works now, and the going rate for skipping the line is being set in public.

So the trade is the line: the seven-year queue for power, the companies standing at the front of it, and the companies selling the way around it.

The rules changed this summer

Because it was not just the buyers who moved. In the six weeks before this was published, four governments acted on the same conclusion:

• June 30th: The Department of Energy issued an emergency order ahead of an eastern heat wave with a first buried inside it: utilities were told to curtail data centers with their own backup generation before cutting anyone else. The federal government now ranks who loses power first. Data centers came top of the list.

• July 14th: PJM, the largest grid operator in the country, published its capacity auction for 2028/29. It cleared at the legal maximum price for the third consecutive year, in every zone, and still fell 6,831 MW short of its own reliability target. $16.4bn for one year of capacity, and the auction still could not buy enough. When that happens three years running, the price cap is not the story. The shortage is.

• July 14th, the same day: New York froze permitting for every new data center of 50 MW or more, the first statewide moratorium in the country.

• August 3rd: The governor of Texas ordered ERCOT to halt new data center connection approvals and audit its entire large-load queue: 474 GW of requests, in a state whose record demand, set three weeks earlier, is 91.3 GW. Five times more capacity is asking to connect than the state has ever used at once.

Buyers paying billions to skip the line. Governments freezing the line. Same summer, same conclusion, opposite ends of the problem.

Our thought process

The whole thesis is three sentences long.

#1: Demand is real, and it moves faster than power. American electricity demand is growing again after twenty flat years, and the machines driving it are ordered on an eighteen-month cycle while the things that power them take two to eight years to build.

#2: Everything that would normally close the gap is already sold. Turbine slots are gone into the next decade. Eaton, the company that makes the electrical guts of these buildings, counts 307 GW of demand in its US pipeline and calls it fifteen years of backlog at Q2 2026.

#3: When the manufactured path is sold out, only two things are left worth owning: what already exists, and whoever delivers fastest. Existing gas plants, existing pipelines, existing nuclear, and the small set of companies that can put power on a slab in months instead of years.

The structure:

• Part One explains what is physically happening to American electricity.

• Part Two covers what happened so far in 2026, and what happened the last time capital tried this.

• Part Three is our top picks.

The Number Nobody Finishes

Every research note this year leads with the same statistic: US data center announcements have grown from about 18 GW in 2023 to about 295 GW now, a sixteen-fold rise in three years.

Almost none of them print the second half of the sentence:

Through 2030, the American grid is forecast to add about 93 GW of accredited supply.

One word carries the whole sector.

That word is accredited, and once you understand it, you will read every energy headline differently.

A power plant has two sizes. The size on the label, called the nameplate, is what it produces under ideal conditions. The size that matters is what the grid operator is willing to count on during the worst hour of the year, a stagnant August evening when every air conditioner on the seaboard is running, and the sun is dropping. That second, smaller number is the accredited capacity, and it is the only one the system actually plans around.

The gap between the two can be enormous. Take a 100 MW solar farm in the eastern market. On the label: 100 MW. At 7 pm on the evening that matters: almost nothing. Accredited value: 14 MW.

The other 86 MW exists. It generates power. It earns money. It is simply not something the system can lean on when it matters.

This is why the United States can be adding an enormous amount of generation and still be short. In 2026, it will add roughly 86 GW of new capacity, the largest single year since 2002:

But when the grid operator works out how much of that it can count on for a peak evening, the number collapses.

295 announced. 93 accredited. That gap does not close by building more announcements.

The Five Layers

The best way to understand this sector is to follow an electron from the ground to the chip. There are five stages. Each has a lead time and a price.

Layer one: the molecule

Almost every fast answer to the power problem burns natural gas. On-site turbines burn it. Reciprocating engines burn it. Fuel cells run on it. New utility plants burn it.

None of that works if the gas is not physically there.

And gas does not move through a pipeline on its own. Pressure drops as it travels, so roughly every hundred miles you need a compressor: a very large engine that squeezes the gas back up to pressure and pushes it along. Without compression, a pipeline is an extremely expensive empty tube.

Gas sitting in the US power interconnection queue grew 86% in the year to mid-2026, to 253 GW. But pipelines are permitted over years, litigated for longer, and cannot be ordered from a factory.

In New Mexico in July 2026, a state regulator blocked the route for a 17-kilometer lateral pipe to a single data center. Seventeen kilometers!

Layer two: the generator

Large gas turbines, meaning machines above 100 MW, are the industrial answer. They are effectively sold out. Smaller industrial turbines from 1 to 39 MW and reciprocating engines are the fast answer. A reciprocating gas engine starts in seconds and reaches full output in about two minutes, which matters more than efficiency when the alternative is nothing.

Build times and capital costs, from the July 2026 levelised cost tables:

New nuclear costs roughly six to ten times what a gas plant costs and takes three and a half times as long. That is why nobody is building it for this cycle, whatever the press releases say.

And the 24 months for a gas plant is only the concrete. The turbine itself is now the long pole.

• The largest maker has sold the vast majority of its 2027 through 2030 production slots

• Order-to-delivery runs beyond five years

• Commissioning lags shipment by roughly another two years

• Turbine prices are up 195% since 2019, heading toward roughly $600 per kW by the end of 2027

The order data says the frenzy is still accelerating. The second quarter saw roughly 38 GW of global gas turbine awards, the largest quarter ever recorded, up 71% on a year earlier. The first half alone booked 67 GW against about 100 GW in all of 2025. Advanced-class machines (the big H and J frames) were 20 GW of it, up 93% on the year. The awards now carry expected operation dates out to 2032.

The factories are responding, and the response has a number: the three large makers plus the rest of the field are on track to expand from about 55 GW of annual production capacity in 2025 to about 85 GW by 2028, and roughly 95 GW beyond 2030, in simple-cycle terms.

Meanwhile, the price of the finished plant keeps climbing. From a tracker of 78 announced North American projects: the median quoted cost of a combined-cycle plant delivered in 2031 is up about 200% against one delivered in 2025, with early-2030s quotes now running $2,800 to $3,000 per kW and above. A simple-cycle plant for 2028 delivery is up about 50% in the same window.

And the fallback fuel is in its own storm. The default bridge while you wait for gas equipment is a diesel generator: 90 to 110 week lead times, and now the fuel itself has broken. Diesel cracks are up 86% since June 2026, US distillate inventories sit at multi-year lows, Russia has banned diesel exports through January 2027, and three of the world’s four refining hubs are impaired. So the stopgap costs more, arrives later, and, since the 30 June federal order, is the first thing a utility is told to shut off in an emergency. Everything about 2026 keeps pushing the same direction: permanent, permitted, gas-fired, on your own land.

Mitsubishi Power, the third of the three large makers, said in July 2026 that lead times have gone from two years to more than five, that it is signing contracts for delivery in 2031 to 2034, and that industry demand is running above 100 GW a year against a market built to make about 30.

One more thing happened to this layer in the last two weeks, and it matters for every project modeled on rented power. The best-known shortcut was mobile turbines on temporary permits. On July 30th, Mississippi’s environmental regulator put a retirement schedule on the 69 unpermitted temporary turbines at one AI campus: removals begin 18 August 2026, most are gone by April 2027, the last run to mid-2027. The temporary turbine now comes with an eviction date. The workaround era of this buildout is ending, permit by permit, and the replacement is permanent, permitted, owned generation.

Layer three: the connection

This is where the sector actually breaks.

The US interconnection queue held 2,061 GW of proposed projects at the 2026 national survey. The median time from entering that queue to producing electricity:

• Five years nationally

• Nearly seven years in the largest eastern market, and rising

On an average rather than a median basis, that eastern figure is over eight years, up from under two years in 2008.

Seven years. A generation of AI chips is roughly eighteen months.

There is no version of this where the grid catches up on its own schedule. That is not a criticism of the grid operators. It is what happens when a system designed to add 1% a year is asked to add 5%.

The queue is shrinking

The 2,061 GW figure is down 10% over 2025. More than 750 GW of projects withdrew during the year against roughly 600 GW of new requests. That is the second consecutive year in which withdrawals exceeded additions.

The composition of what is left tells you what developers now believe:

And the historical conversion rate is the number that should govern every forecast in this sector:

Of all the capacity that requested interconnection between 2000 and 2020, only 13% had reached commercial operation by the end of 2025. Roughly 75% withdrew.

A further 549 GW currently holds a draft or executed interconnection agreement and still has no operating date.

Put those two facts together. The queue is both the reason everyone believes in enormous demand and the least reliable forecasting instrument in the industry. It has never converted above about one in five, and it is currently shrinking while the gas share of it nearly doubles.

Everyone is in line. Almost nobody makes it to the counter. And the people quoting the length of the line as evidence of demand have never once checked how many were served.

Layer four: the step-down

Electricity arrives at a site at very high voltage, because that is how you move it efficiently. A building cannot use it in that form. A transformer steps it down. Switchgear routes and protects it.

These are not exotic machines. They are, however, custom-built, and the factories that make them were sized for a market growing at 1% a year.

Lead times as of mid-2026:

Demand for generator step-up transformers is up 274% since 2019. Roughly $1.8bn of new North American transformer capacity has been announced. Most of it lands in 2027 and 2028.

And the company that sits across most of this layer put a single number on it at its Q2 2026 results, 31 July: Eaton counts 307 GW of demand in its US data-center pipeline and calls it “15 years of backlog.” Its data center revenue grew 65% in the quarter. When the shovel seller tells you the line outside the store is fifteen years long, believe the line before you believe any individual customer’s schedule.

Layer five: the room

A traditional server rack draws about 10 kW. An AI rack draws about 100 kW.

Ten times the power in the same physical footprint means ten times the heat in the same physical footprint. Air conditioning stopped working as a solution. That is why liquid is now pumped directly onto the chips through cold plates, and why the cooling business quietly became a chemistry business rather than an air business.

The cost changes shape too:

Inside the facility portion of that cost, electrical equipment is 48% and mechanical is 30%. Nearly four-fifths of the building cost, excluding the computers themselves, is moving power and removing heat.

The One Sentence That Summarizes the Stack

Put all five layers on one axis, and the picture is uncomfortable.

Every single way of making or connecting electricity in America takes longer than the technology cycle it is being asked to serve.

• Solar is fastest at 12 to 18 months, and it counts for 14% of its nameplate

• A gas plant is 24 months, plus a turbine you cannot get until 2030

• A transformer is up to four years

• A grid connection in the largest market is eight years

• New nuclear is seven years of construction after roughly a decade of everything else

Against a GPU generation of eighteen months.

So the buyers stopped waiting.

The Price of Not Waiting

The most important number in the sector is also the least quoted.

“On-site” here means generation that sits on the data center’s own land and never touches a utility wire.

Two to three times the cost of the electrons, paid willingly, by the most numerate buyers on earth. Which tells you the electrons were never the product.

The product is the delivery date. Power is just how it ships.

The math

One operator’s management sizes a 1 GW data center at $1bn to $2bn in monthly revenue once it is running.

• Assume a 20% to 25% net margin

• That is $200mn to $500mn of profit per month

• Every month of delay costs the operator that amount

Now weigh that against paying a few dollars more per megawatt-hour for electricity. It is a rounding error.

Priced properly: a single month of schedule is worth roughly $5 per MWh over the life of a contract. Pull the start date forward by a year, and the entire premium is paid for several times over.

This is the sentence that reorganizes the sector:

Levelised cost of energy does not price time, and time is now the only thing being bought.

That is why a fuel cell deployable in 90 days competes against a grid connection that is nearly free at the margin but arrives in eight years. It is not competing on cost. It is not close on cost. It is competing on a calendar, and on a calendar it wins by seven years and nine months.

The clearest proof the premium is real

Go back to the two smelters from the opening. Between them, the two leases cover about 592 MW of contracted capacity on sites whose grid connections were built for smelters, and both had been picked up by bitcoin miners years ago precisely because miners were the only buyers who valued a live wire in the middle of nowhere. Now divide Anthropic’s $28bn by anything physical on those sites and the deals look insane, until you divide by the megawatts. Then they look cheap. The fastest way to build an AI data center in 2026 is to buy a bitcoin miner and evict the bitcoin.

Powered land is the asset class. The rest is drywall.

Where Fifty Million Dollars Per Megawatt Actually Goes

Building one megawatt of AI data center costs roughly $50 million. Building one megawatt of a traditional data center costs about $20 million. Here is where the money goes.

The top-level split

Note what happened. The computers got roughly 3.5 times more expensive per megawatt. The building got only marginally more expensive. So the facility fell from 46% of the total to 22%, which is why people stopped paying attention to it.

That is a mistake, because the 22% is what determines whether the 78% can be switched on at all.

Inside the facility

Of that roughly $11 million per megawatt:

Nearly four-fifths of the building, excluding the computers, is moving power and removing heat.

And now the generation, which is a separate bill entirely

None of the above includes actually making the electricity. If you are connecting to the grid, you do not pay this. If you are building on-site because the grid is seven years away, you do:

Those are not strictly comparable. The engine number is capital cost; the fuel cell number is revenue per megawatt to the supplier, which includes their margin. But the range is the right order of magnitude, and it tells you what “bring your own power” actually costs.

The number that follows

Add the electrical infrastructure to the on-site generation, and you get roughly $7 million to $9 million per megawatt.

The entire power layer, from the generator to the busway, is about 15% to 18% of what an AI data center costs. And it is the only part that determines whether the other 82% ever produces a token.

That single ratio is why this sector behaves the way it does. The power layer is small enough that the buyer will overpay for it without blinking, and critical enough that they cannot proceed without it. A hyperscaler weighing a few million dollars a megawatt against $39 million of idle accelerators is not running a procurement process. It is running a countdown.

It also explains the content figures the equipment makers now quote. One infrastructure supplier sizes its own content at $3.25 to $3.75 million per megawatt, having raised that range by 10% to 15% in a single year. On our numbers, that is roughly 40% of everything spent inside the building. One turbine maker sizes its electrification content at $0.2 to $0.3 million per megawatt today, moving to $0.4 to $0.9 million as new products land, which is a tripling of its take from the same megawatt.

Everyone in this chain is quietly increasing their share of a bill the customer has stopped scrutinizing.

Behind the Meter

“Behind the meter” is jargon for something simple: a power plant on the customer’s own property, wired directly into the building, that never uses the public grid.

It solves the queue problem completely, because there is no queue.

It is also, currently, the most over-announced category in American energy.

The census

From the most careful bottom-up count available: 158 GW of announced on-site gas generation across 109 discrete projects.

Now the reality ladder on that number:

Between 30% and 50% of the announced number is real by 2030. The remainder is a press release stapled to an architectural rendering.

Narrow it to the front of the funnel, and it is starker:

Of roughly 90 GW of announced behind-the-meter data center power, about 2 GW is operating today. That is 2.2%.

The opportunity is still real

Take the unconstrained demand forecast of 177 GW of new capacity needed. Subtract the 93 GW of accredited grid supply. You get an opening of 80 to 84 GW that the grid cannot serve. Texas alone is 61% of all accredited net additions in the country.

And on August 4th, the market finally got a fully priced public example of what serving that opening is worth. NRG disclosed a 1.2 GW gas plant in Texas, built for a single unnamed hyperscaler under a bring-your-own-power contract: 15-year minimum term, $3.2bn of capital, about $500mn a year of EBITDA, running from late 2029. That is $2,700 per kW to build and roughly a 15% cash yield on cost, contracted, before a single share of upside. And $2,700 no longer even carries a premium: the median quoted cost of a grid combined-cycle plant for early-2030s delivery has reached $2,800 to $3,000 per kW. The dedicated plant now costs about the same as the shared one and arrives years sooner.

But the gap between 90 announced and 2 operating is the whole investment problem in this sector, and it is where the money was actually made and lost in 2026.

Who Is Allowed to Plug In

One more layer, and it changed fastest this year.

For twenty years, the constraint on building a data center was land, fiber, and money. In 2026, it became permission.

The price signal that turned the politics

Retail electricity prices in the year to May 2026:

Seven of the ten fastest-rising states are in the eastern grid, which is where the data centers are.

New England, which has the highest absolute electricity prices in the country and no data center boom, is flat to down.

That contrast is the strongest available evidence that these increases are load-driven rather than generic energy inflation.

What cuts against it

Of the $9.4bn of utility rate increase requests filed in the first quarter of 2026:

• $4.4bn was Western wildfire recovery

• $2.7bn was Southern storm recovery

• Roughly 76% had nothing to do with data centers at all

Anyone using the headline rate-request number as proof of data center backlash is overstating their case. The state-by-state contrast is the better evidence.

The polling

In March 2026:

• 71% of Americans said they oppose a data center being built near them

• For a nuclear power plant, the figure is 53%

People now object to data centers more than they object to reactors. It took the nuclear industry four decades and two accidents to earn that number. Data centers got there in about three years. National support for data center construction fell from 65% to 36% in thirteen months.

And the legislatures moved

• More than 300 data center bills were filed across more than 30 states in the first six weeks of 2026, against 267 in all of 2025

• New York’s governor issued an executive order on 14 July 2026 creating the first statewide moratorium on hyperscale facilities, running up to a year pending an environmental review. The legislature had passed a stiffer bill in June; it remains unsigned

• Texas froze its queue on 3 August

• A bipartisan federal ratepayer protection bill cleared its House committee 52 to 0 on 21 July

A Democratic governor’s moratorium and a Republican governor’s queue freeze arrived twenty days apart, in the two largest data center markets in the country, in a midterm year.

What Part One Establishes

Five things, all checkable.

#1: American electricity demand was flat for two decades and then turned. Consumption in 2017 was lower than in 2007. From 2005 to 2020 it grew at 0.10% a year. From 2020 to 2025 it grew at 1.77%.

#2: The forecasts moved much faster than the demand. The five-year-ahead peak load growth forecast went from 23 GW in the 2022 vintage to 166 GW in the 2025 vintage, a seven-fold increase in three years.

#3: Physical supply cannot respond on the timescale being asked. Every layer of the stack has a lead time longer than a GPU generation, and the eastern capacity auction has cleared at its own price ceiling three years running while producing 317 MW of genuinely new generation in its most recent round.

#4: So buyers pay two to three times the cost of electricity to skip the queue, and what they buy is a date rather than a commodity.

#5: The political system noticed, in both parties, in the same month.

PART TWO: The Year, and The Last Time

2026 in Four Acts

Act one, December to February: the rules change

• December 18, 2025: The federal energy regulator found the largest grid operator’s tariff unjust and unreasonable for lacking clear rules on power plants co-located with data centers, and ordered three new transmission services within 60 days.

• December 17, 2025: The capacity auction for delivery year 2027/28 cleared at its price cap of $333.44 per MW-day, 6,623 MW short of the reliability requirement.

• January 6, 2026: The Texas grid operator disclosed a large-load queue of 233 GW, up roughly 300% in a year, noting its process “was designed to consider 40 to 50 large loads at a time.”

• January 14, 2026: The largest grid operator cut its own near-term load forecast for the first time in the cycle, by 2,564 MW for 2026, citing stricter vetting of planned data centers, while raising its ten-year path.

Act two, March to May: the narrative cracks, the capital arrives anyway

• March 8, 2026: The planned expansion of a flagship Texas AI campus from 1.2 GW to 2.0 GW was reported canceled over financing terms and winter damage to cooling infrastructure. One of the parties publicly disputed the reporting the same week.

• March 16-30, 2026: The Department of Energy issued four emergency orders deferring roughly 4.4 GW of coal retirements, part of more than 40 such orders since May 2025. The federal government is now physically preventing power plants from closing.

• March 2, 2026: One of the largest independent power companies agreed to be taken private.

• May 18, 2026: The largest regulated utility merger ever proposed was announced at roughly $420bn of enterprise value.

• May 12, 2026: A geothermal developer completed the largest clean energy IPO on record.

Act three, June to July: the regulator and the voter arrive together

• June 18, 2026: The federal energy regulator issued show-cause orders to all six regional grid operators simultaneously, giving them 60 days to justify or reform their large-load rules. That is an unusual instrument. It means the regulator has concluded the existing rules are probably unjust and has put the burden of proof on the operators.

• June 30, 2026: The Department of Energy issued an emergency order ahead of an eastern heat wave that, for the first time, directed utilities to curtail data centers with on-site backup before touching other load. Two days later, the eastern grid peaked at 161,910 MW, about 2% under its all-time record.

• July 22, 2026: Texas set its all-time demand record, 91.3 GW. Real-time prices peaked at $378 per MWh, in a market that touched its $5,000 cap in 2024, because record solar and a record 12 GW of battery discharge absorbed the peak. Storage is quietly flattening scarcity pricing in Texas, which is precisely why our merchant power exposure sits in the eastern market and not there.

• July 14, 2026: New York’s governor issued an executive order creating the first statewide data center moratorium, bypassing a stiffer bill the legislature had passed in June and left on her desk. On the same day, the capacity auction for 2028/29 cleared at its cap again, and the shortfall got worse, to 6,831 MW.

• Late July 2026: Every large cloud company raised its capital spending guidance.

• July 31, 2026: The grid operator filed a one-time backstop auction with a price cap of $555 per MW-day, seventy per cent above the regular cap. That filing is an admission that the capacity market as designed does not build things. When the fix for an auction that keeps hitting its ceiling is a second auction with a higher ceiling, you are allowed to notice that the plan is to turn the amplifier up to eleven.

Act four, July to August: the forecasts split, then the freeze

• July 29, 2026: A research house raised its 2030 US data center capacity outlook by 52%.

• July 30, 2026: The following day, another reported the third consecutive quarterly decline in announced capacity, with the first quarter of 2026 down 19%.

Both are defensible. One models demand, the other counts announcements. They cannot both be describing the same 2030.

The two most-cited forecasters in the sector published opposite directions on consecutive days. That is the best single description of where knowledge actually stands.

Then on August 3, Texas.

And then one more filing, four days after Texas and largely unread: on August 7th, the eastern grid operator asked the federal regulator to approve an interim rule under which data centers coming online after June 1, 2027 must bring their own power, register in a large-load registry, and accept curtailment in emergencies if they do not. The largest grid operator in the country just proposed making bring-your-own-generation the law of its market.

What the Market Paid For

The complex split in half, and the two halves went in opposite directions by more than 2,000 basis points.

• Equipment and electrification returned between 45% and 99% this year

• Merchant power and nuclear returned between negative 2% and negative 37%

• The utilities index returned 4.1% against 13.8% for the broad market

But sector labels are the wrong tool. Here is the right one.

The test

For each large name, take three things:

1. What management guided for 2026 back in January or February

2. What they guide today

3. The stock price move

The difference between the guidance change and the price change is what the market decided to pay, separate from anything the business actually did.

Guidance is from Q4 2025 and Q2 2026 company disclosures, and prices are year-to-date.

Read the top and bottom together

• Eaton raised its 2026 guidance by 1.9%, and the market paid 43% more for it.

• Constellation raised guidance by 4.3%, and the market paid 24% less.

• Vistra did not change its guidance at all and lost 9% of its multiple.

That is not a market pricing fundamentals. It is a market pricing visibility, and all year it has paid for one specific thing: revenue that has already shipped.

The exception

Bloom is different. Its guidance did rise 92%, from $1.33 to $1.48 of non-GAAP earnings at the start of the year to $2.55 to $2.85 now. It earned most of its move. It is the only large name in the sector where the fundamentals did more work than the sentiment.

And what is missing from that table

There are five more names on it.

In four of them, the gap between what the business delivered and what the market paid is wider than anything shown above. In one of them, it runs the opposite way to everything here: guidance up double digits, multiple down double digits, in the same year.

Those five are in Part Three.

Announced Against Delivered

The market’s behavior makes more sense once you measure the sector by what is actually running.

Roughly 90 GW of behind-the-meter data center power has been announced. About 2 GW is operating.

Now hold the famous names against that ruler.

The fuel cell company

One large fuel cell maker markets a backlog of $20bn.

The remaining performance obligations in its audited quarterly filing, dated 28 July 2026, are $494.1mn.

That is a ratio of roughly 40 times. Across a set of ten comparable companies, the widest equivalent gap is about two times.

And the detail hardest to explain:

Product performance obligations grew by $48mn in the first half of 2026, from $394.4mn to $442.4mn, in a half where the company recognized $1.589bn of product revenue and announced a 2.8 GW customer expansion.

The audited forward book barely moved while the marketed one grew.

The counter-evidence is real and belongs here too: customer deposits went from $78.2mn to $360.6mn in six months.Large customers are writing cheques. Both things are true at once.

The turbine maker

One large turbine maker reports a backlog of $176.3bn. Its audited remaining performance obligations are $176.3bn. Identical.

It is also the only company in the sector that separately discloses firm orders from slot reservations, at 53 GW and 63 GW respectively, and it disclosed 10 GW converting from one to the other in a single quarter.

That is the best disclosure practice in the industry.

The merchant power names

The largest merchant nuclear and independent power companies are delivering essentially zero data center megawatts in 2026.

• One restart is scheduled for 2027

• One uprate begins in the fourth quarter of 2027 and reaches full volume in 2032

• One advanced reactor developer booked $1.21mn of revenue last quarter, the first in its history

• Another booked about $75,000, down 99% year over year, with zero backlog

The scoreboard, as of 12 August 2026

Here is the same question asked of everyone at once. Not what is announced. What is running.

And the universe itself is inflating faster than it converts. The same survey counted 46 projects and 56 GW in February 2026. By mid-year it counted 59 projects and 90 GW. 92% of all of it was announced after the start of 2025.

So a reasonable person could conclude the market is right

The names with the biggest announced numbers have the least audited substance behind them, and the market has been discounting them accordingly.

The counter-argument, which is our argument, requires showing which companies are actually delivering electrons today and demonstrating that nobody is paying them for it.

That is Part Three.

A Vocabulary for Reading Announcements

First, a tool you can use on any press release in this sector.

Power agreements exist on a ladder from meaningless to binding. The language tells you which rung you are on.

The tells to watch for: “up to,” “collaborates with,” “toward a definitive agreement,” “subject to.” Each one drops you a rung.

The trap

Some binding contracts produce no backlog disclosure at all.

If an arrangement is accounted for as an operating lease, it sits outside the regime that generates a remaining-performance-obligation number. There is at least one company in this sector with more than 2 GW under long-term contract and no disclosed backlog whatsoever.

Absence of a backlog number is not absence of a contract.

Anyone screening this sector on backlog alone will throw away the careful companies and keep the promotional ones. Which is exactly backwards.

The Last Time

Almost nobody runs the comparison that follows.

The closest historical analogue to 2026 is the American merchant power buildout of 1999 to 2002.

That buildout was itself partly justified by a data center boom. Same argument, same decade as the fiber glut, and it ended in seven Chapter 11 filings.

The number that frames it

The 2001 National Energy Policy called for 1,300 to 1,900 new power plants and 393,000 MW of new capacity, premised on electricity demand growing 1.8% a year for twenty years.

The federal energy outlook that year forecast 34% demand growth by 2020.

Actual growth was 14%.

The buildout was not killed by a demand collapse. It was killed by demand arriving at roughly 40% of the forecast rate. That gap alone destroyed more than $100bn of equity in two years.

A capital cycle does not need a bust to end badly. It only needs a forecast to be wrong by half.

Six lessons

#1: The supply response is violent and fast.

Additions peaked in 2002 at 66.3 GW of nameplate gas capacity, or 56.9 GW measured the way the government reports it today, and fell to 11 GW by 2006, a decline of about 80%. Roughly 220 GW went in across 1999 to 2006, expanding the national fleet by about 30%.

#2: Equipment makers survive. Producers do not.

The largest turbine maker went from $22.9bn of revenue and $6.3bn of segment profit in 2002 to $18.5bn and $4.1bn in 2003. Down 19% and 35%, and profitable throughout.

It had guided in July 2002 that US gas turbine orders would fall 80% or more over two years. Its main plant shipped 86 heavy-duty turbines in the second quarter of 2002 against 24 a year later.

What carried it was not the order book. It was $29.3bn of outstanding service commitments on the installed base, against which it signed $1.3bn of new service work in a single quarter of 2003.

Meanwhile:

• Four large merchant power producers filed for bankruptcy

• One large international power company fell 97%

• One large trading and generation company fell about 95%

• One European turbine maker effectively failed and required a state rescue

And it took 22 years to reclaim the 2002 turbine order peak. It was first matched again in 2024.

#3: The killer is the capital structure, not the asset.

One of those bankruptcies was not caused by weak margins at all. It was caused by a $1.85bn collateral posting requirement triggered by a credit rating downgrade, which the company could not fund.

At the top of the cycle, the three most levered producers carried 8.0x, 8.0x and 8.8x debt to EBITDA, the last with 1.7x interest coverage.

The assets did not disappear. The equity did, in weeks.

#4: The asset survives, the equity does not, and the second owner makes the money.

Distressed merchant power plants transacted at $90 to $225 per kW against a build cost near $500 to $600 per kW.One 5,280 MW portfolio of eight gas plants changed hands at roughly $90 per kW, which is about fifteen cents on the dollar.

In the fiber glut next door, a network that cost $15bn to build sold whole, nine years later, for $3bn.

In all three of the great capex cycles of the last thirty years, the technology thesis was correct, and the first-round equity was still a near-total loss.

The value went to whoever owned the asset in round two.

And the detail that explains the price dispersion: the high prices went to plants bought by entities with ratepayer backing. The contract backstop, not the asset, set the price.

#5: The equity clock and the physical clock are different clocks.

Equities bottomed in the fourth quarter of 2002 and made new highs by 2007. Combined cycle utilization did not normalize until roughly 2015.

The trade bottoms years before the asset does, and it tops years before the asset does too.

#6: Long-dated contracts are credit instruments, not certainties.

This is the lesson that threatens the entire contracted-revenue argument in this sector, so it gets the most space.

• In the 1970s, a public power system in Washington State wrote an explicit dry-hole clause into its agreements: the participant pays “whether or not any of the Projects are completed, operable or operating.” That is stronger language than exists in any data center power contract today. In 1983 the Washington Supreme Court voided those agreements as beyond the participating utilities’ statutory authority, producing a $2.25bn default, the largest in municipal bond history.

• In 2001 California signed 52 long-term power contracts with a face value of $42.9bn. Twenty-three were renegotiated. Present-value savings to consumers came to $580mn, and the litigation ran until a federal ruling in December 2023.

• In 2018, a bankruptcy court approved rejection of multiple power purchase agreements under the ordinary business judgment standard, holding that “the economic disappointment a power contract counterparty experiences cannot be avoided by invoking the Federal Power Act.” The counterparty gets an unsecured damages claim and nothing else.

Three stress tests. Three failures to deliver the certainty the drafters intended.

The Analogue Nobody Draws

There is one historical case where the buyer of power was the creditworthy party driving the buildout, which is exactly the structure of today’s market.

It is not encouraging.

A federal power agency in the Pacific Northwest built enormous generating capacity anchored by long-term contracts with large, creditworthy, single-purpose industrial customers. The first aluminum smelter contract was signed in December 1939 for 32.5 average megawatts. By the 1960s, the industry ran ten smelters consuming 3,150 average megawatts, 40% of all US aluminum smelting capacity.

It ended in three moves:

1. The agency raised rates nearly 250% between 1979 and 1983, largely to pay for a nuclear program it had underwritten.

2. In 1996 and 1997 the smelters relinquished their contracts to buy cheaper power on the open market, cutting the agency’s supply to them by 40%.

3. In the 2000 to 2001 price spike, they shut down and resold their contracted power at a profit rather than smelt aluminum.

By the summer of 2001, all ten were shut or running minimally. The last two were idled permanently in 2016. One company paid $7.3mn in 2017 merely to preserve the option to restart, and never used it.

An investment-grade anchor buyer does not guarantee an investment-grade load.

The question this sector has to answer is not whether the largest cloud companies will pay. They are among the most creditworthy institutions on earth.

It is whether inference is still worth doing at that location, on that hardware vintage, at that price, in year twelve of a nineteen-year contract.

Scoring This Cycle Against the Last One

The literature on capital cycles converges on seven signals that a cycle is topping.

Five of seven are fully present. We are not going to pretend otherwise, and anyone selling you this theme without that table is selling you the theme.

On signal three specifically

• Hyperscaler debt as a share of capital spending went from 9% in fiscal 2024 to 32% by mid-2026

• Off-balance-sheet lease obligations across the five largest run near $820bn

• One large software company was downgraded to the lowest investment-grade rung in July 2026 on capital spending, negative free cash flow and customer concentration, with lease exposure near three times its own next-year capital spending guidance

• In June 2026, one of the five raised $84.75bn of equity, the largest corporate equity transaction on record

The One Real Difference

And now the fact that makes this cycle different from 2002, which is why we are long anything at all.

In 2003, the Texas grid operator’s reserve margin was 24% on a conservative basis and 33% including all resources. The system was long, with excess committed capacity averaging about 9,000 MW.

In 2026, that same reserve margin is 18.3%:

And those figures already haircut every new data center request to 49.8% of what was asked for.

In the eastern market, the most recent auction cleared 6,831 MW short of the reliability target at the price cap, and the supply response was almost nothing.

Read the auction reports carefully, because the headline number is not what most people quote:

The figure usually cited, 774 MW falling to 525 MW, includes uprates to plants that already exist and reactivations of plants that were switched off. Genuinely new generation was 351 MW, then 317 MW.

Three consecutive auctions at the price ceiling have produced roughly 300 MW a year of new power plant. A single large gas turbine is bigger than that.

The last cycle began from glut. This one begins from scarcity, which is why it can run harder and longer before it breaks.

The counter travels with that sentence

Those reserve margins are computed against a load forecast whose input is an interconnection queue that historically converts at 10% to 22%.

• The largest grid operator expects roughly 22% of its 65 GW large-load pipeline to materialize

• A survey of 25 large utilities found that among the ten whose requests exceeded half their peak load, none expected the realized five-year share to exceed 35%

• The definitive academic study of utility forecasting accuracy found that the utilities projecting the highest growth rates experienced the lowest actual growth, with concentrated industrial load the largest single source of error

Data centers are the most concentrated industrial load ever connected to the American grid.

That is the same category of error as 393,000 MW at 1.8% a year, arriving from the opposite direction.

Where That Leaves It

The 2001 comparison does not map onto the largest cloud companies. It maps onto whoever sits between them and the turbine.

In 2001, the seller of power was the levered speculative party, and the buyer was diffuse. In 2026 the buyer is mostly investment grade with real cash flow, but a new levered layer has been inserted between the buyer and the asset.

The risk did not disappear. It moved one rung down the capital structure and became harder to see.

Which is the practical question this sector actually poses, and the one Part Three answers:

Which layer of this stack gets paid whether or not the forecast is right?

PART THREE: Our Top Picks

The roster. Full theses, falsifiers and valuations below.

You can mirror our picks automatically

Energized runs on Autopilot. The nine positions below sit in a live account that mirrors straight into your own brokerage.

• Every entry and every exit is mirrored. When we add, trim, or sell, your account does the same thing at the same time.

• You hold your own money. The positions sit in your brokerage, in your name. We never take custody of anything.

This is built for people who do not have time to do any of this. You should not have to read 20,000 words, track a capacity auction calendar, or sit at a screen on the morning a position changes. Reading the work is optional. The account does the same thing either way.

And there is more coming. We will be starting weekly write-ups highlighting: every move we make across all our Autopilot funds, what triggered it, and the market news we thought actually mattered that week. We’ll continue to do basket write-ups and single-company write-ups over time.

The button below goes straight to Energized.

Energized Portfolio

Read more

Join our Autopilot here!

And for premium subscribers: a CV Research Discord is coming! Included with every paid Substack membership. Details soon.

Introduction

Every AI data center on Earth has the same problem, and it is not chips.

It is that electricity stops working as a way to move information. Every time the machines get faster, the distance a signal can travel over copper gets shorter. Ten years ago, a copper cable could carry data seven meters. At today’s lane speeds, it reaches about two. In the next generation, the copper that survives is not really a cable anymore. It is a connector fused into the chip package. Past that, there is no copper answer at all. The last two meters of the AI data center are the only part electricity still owns, and they shrink every generation on a published schedule. Everything beyond them has to become light.

That conversion is not optional, it is not gradual, and it is not fully priced. A single AI rack today contains 5,184 copper cables and zero optical connections between its GPUs. By 2028, on the current roadmap, that same class of machine converts to light almost entirely. The fibre count per GPU goes from zero to roughly 126, then to roughly 288. Nothing else in the AI supply chain, not memory, not power, not cooling, goes from zero to everything inside one product cycle.

The market knows optics matter. Optics stocks have run hard, they move violently, and everyone in AI has an opinion on them. What the market has not done is the harder thing: notice that “optics” is not one business. It is four businesses stacked on top of each other, with completely different economics, wearing one trade.

You can see it in the prices the parts themselves fetch. For example, in July, inside this one supply chain, in the same quarter, sold to the same customers: components for 1.6 terabit modules up 10% to 15%. Optical engines up 5% to 8%. Indium phosphide laser chips up 4%. VCSELs flat. And 800-gigabit optical components, the workhorse of the current buildout, down 3% to 4%. One supply chain. Prices moving 19 points apart at the two ends of it. That is not noise. That is a map of where the power sits, and this piece is that map.

You can see it again in the margins. One company in this chain took its highest-value product from immaterial to roughly two-thirds of revenue and watched gross margin go from 8.0% to 44.9% in five quarters. Another grew revenue 55% while its AI mix went from 32% to half of sales, and gross margin moved 1.3 points. Same box. Same buyers. One of them sets prices. One of them takes prices. Nothing about how the two stocks trade tells you which is which. There is a test that does, and we have run it across the entire stack.

And you can see it, most clearly of all, in what the buyers are doing with their cash. Inside four months this spring, five separate buyers paid up front to reserve capacity in this chain. Deposits of 30% to 50%. Contracts stretching to 2031. One deposit that turns nonrefundable if the buyer misses its own minimum. One shortfall liability that sits on the buyer, not the seller. Nvidia went further and put $6.5bn of equity directly into four of its own suppliers. When the most powerful buyer in the history of the industry starts leaving deposits with its component makers, pricing power has an address. Most of the market has not gone looking for it, because the visible price increase says 4% and 4% screens as nothing. The scarcity is real. It is just being paid in a currency that does not show up on a price screen, and we will show you exactly where it shows up instead.

Then, in early August, Washington accidentally proved the whole thesis in one afternoon. A reported restriction on Chinese optical transceivers sent the group vertical. China builds roughly two-thirds of the world’s transceivers, so a transceiver ban is possible. The laser chip layer underneath those transceivers is Western and Japanese, so the reverse ban is not available on the same terms. One layer of this supply chain is expendable, and one layer is leverage. The market bought both indiscriminately on the same afternoon, which tells you it still cannot tell them apart. The gap between those two layers is the whole piece.

One thing before we start, because the alternative is pretending. This is not an undiscovered corner of the market. The chief executive of one of the largest laser makers said from a public stage in early July 2026 that the indium phosphide shortage could get worse than the memory shortage. The names have many analysts already covering them. If you came here for a stock nobody has heard of, this is not that. What does not exist, anywhere, is the frame. We went through all the documents we could find and could not find the price list above read as one column, the margin test run across the full stack, or anything that separates the price setters from the price takers inside the same supply chain. The edge here is construction, not discovery. We are going to build it in front of you.

Four of our five picks report within the next three weeks of this write-up. Every pick carries a test with a number and a date.

Part One: The Map

Four layers, not one

Start at the bottom and work up. The physical chain that carries a bit between two accelerators has four distinct economic layers, and they are as different from each other as a copper mine is from a wire harness.

The substrate. Indium phosphide, grown as a crystal boule and sliced into wafers. Three companies control more than 90% of world supply. It is a materials business with the economics of a materials business, which means it is cyclical, capital hungry, and prone to gluts.

The epitaxy and the light source. The wafer goes into a reactor where atomic layers are grown on it, then diced into laser chips. This is where a $2,500 wafer becomes something worth many times that. It is a process know-how business, and process know-how does not transfer with a purchase order.

The signal path. The analog front end that converts electricity to light and back, and the digital signal processor that cleans the signal up. Merchant silicon, multiple suppliers, buyers who actively fund second sources.

The module and the assembly. Everything gets packaged into a transceiver and plugged into a switch. High volume, low margin, heavily concentrated in China.

Consensus treats these as one trade because they ship in one box. The gross margin data says otherwise, and the divergence is the entire piece.

Read more

Ask ten investors what Rocket Lab does and ten will say rockets. Rockets are now a third of revenue and the least valuable thing the company owns.

The bigger part, growing 57% a year, is a satellite factory anchored by the Pentagon, which just handed Rocket Lab the largest contract in its history to build missile-warning satellites. And a month ago the company signed the move that makes this write-up necessary: an $8 billion acquisition of Iridium, which brings 2.6 million paying subscribers, half a billion dollars a year of cash profit that arrives whether or not a rocket ever flies, and one of the last two deeds to globally coordinated L-band spectrum on Earth. Understand the club it just joined. SpaceX tried to buy the other deed and lost it to Amazon at $11.57 billion, then paid close to $20 billion for bare spectrum licenses as the consolation prize. No satellites, no subscribers, no revenue. Rocket Lab paid $8 billion for the most complete version, the one with the constellation, the certifications, and the Pentagon contract attached.

The market’s response was to cut the stock in half.

That is the setup, and the gap between the price and the machine is stuffed with things nobody has written down. The spectrum this entire deal is built around sits on Iridium’s books at fourteen million dollars, an accounting fossil from a bankruptcy, which means the market’s screens literally cannot see the asset being bought. The stock is not even trading on fundamentals right now: a merger collar has turned it into plumbing, and we show you the exact number it is chained to, why it snapped loose last week, and how to read every tick between here and closing. The most important contract inside the target, nearly a third of its revenue, comes up for renegotiation with the Pentagon in September, and not one sell-side note mentions it.

What follows is the full machine, layer by layer: the launch business, the factory, the landlord, the contract nobody is watching, the rocket the whole year leans on, the deal mechanics, the weighted scenarios, and the exact calendar on which all of it resolves inside eighteen months. The market is pricing the rocket. By the end, you will know what the machine is worth, and precisely what would prove us wrong.

Read more

Pull up any photo of an AI server rack and look past the chips everyone obsesses over. See the thick purple cables threading through the racks like veins? One company makes those, and over the past two years, it has quietly put up some of the most absurd numbers in the entire market: revenue up 126%, then up 206%, now guided up 80%+ again, all while keeping 68 cents of every sales dollar. Margins like a monopoly, growth like a startup, zero debt.

Now the strange part. This company sells to five of the six biggest AI spenders on earth, and the missing logo is the one everyone assumes matters most: Nvidia. Most investors hear that and file the stock under “risky.” We spent a long time inside the filings, the supply chain, and the models Wall Street actually uses to size this industry, and we came away convinced the market has it backward. The half of the AI buildout that gets modeled obsessively is Nvidia’s half. The half that barely gets priced at all is the one where Amazon, Microsoft, Meta, Oracle, and xAI design their own racks. That un-priced half is this company’s home country, and blind spots are where mispricings live.

So here is the full breakdown: what Credo actually sells, the five product engines, the origin story almost nobody knows, the year it nearly died, the financials, what the institutions think and where we think they are asleep, the bear case, and, because a thesis without a price is a slogan, our exact assumptions.

What Credo Actually Does

An AI data center is not one big brain. It is hundreds of thousands of brains that have to think together, as one. A brain is useless if the nerves connecting it to the others are slow or unreliable.

Credo makes the nervous system.

The flagship product is a smart copper cable with Credo’s chips built into both ends, cleaning and boosting the signal so GPUs can talk at full speed. Roughly $500 per cable, about 5 cables per 4 GPUs deployed. Every time you read “Meta ordered another 100,000 GPUs,” add: “...and ~125,000 cables from somebody. Probably Credo.”

Why anyone pays $500 for a cable

A training run is a race in which 100,000 cars must cross the finish line together. One flat tire and everyone waits. If a cheap cable “flaps,” meaning it drops its connection for even a moment, millions of dollars of hardware stalls while the system recovers. Credo’s brand is literally named ZeroFlap: the cable that never drops. When your Ferrari fleet costs billions, you do not shop for the cheapest tires. That is where the 68% gross margin comes from.

The second sale is power. Electricity, not money, is the binding constraint in AI data centers, and Credo’s founding obsession since 2008 has been moving data with dramatically less energy. Every watt saved on cables is a watt you can feed to another GPU. Every hyperscaler CFO understands that in one sentence.

Who buys, and how it locks in

Five of the six biggest AI spenders on earth: Amazon, Microsoft, Meta, Oracle, and xAI, plus the NeoClouds (startup AI clouds like CoreWeave), now account for about 20% of sales. These five each already spend $100M+ per year, and management said in July that no customer is stepping off the gas.

Notice who is missing from the list: Nvidia. Hold that thought later.

Data centers are built like hotel chains: design one blueprint, build it fifty times. Once Credo’s cable is qualified into a customer’s blueprint, every copy of that data center buys Credo cables for years, no rebidding. Nobody rips out a proven part mid-generation when a failure idles millions in hardware. Credo does not win orders. It wins blueprints. And of the five major customers, only one is fully penetrated. The other four still have designs that Credo has not rolled into yet. That is multi-year growth sitting inside relationships already won.

The Product Lines

Think of Credo as a data delivery company, with a vehicle for every distance.

Read more

Every AI data center going up in America right now has the same unglamorous problem. Electricity arrives from the grid at hundreds of thousands of volts, and a server chip drinks it at about one. Somebody has to build the machines that step it down: the room-sized transformers, the switchgear that kills a fault in milliseconds before it becomes a fire, the refrigerator-sized power units feeding the racks. Those machines are custom-engineered, have a quoted lead time of 1 to 3 years across the industry, and America imports more than 75% of the most important one.

Forgent Power Solutions builds all of them in America and delivers them faster than the giants can.

The scoreboard shows what that is worth right now. Revenue grew 103% last quarter, every dollar of it organic. For every dollar of gear Forgent shipped, customers ordered roughly two more, for the third quarter straight. Backlog hit a record just shy of $2 billion, up 157% in a year. The guidance was raised so high that the new floor sits above the old ceiling. A brand-new AI cloud signed a $100 million-plus order for the first building of a multi-gigawatt campus and took delivery six months after the PO, in an industry that typically quotes a year or more for the same equipment. And Washington, which has formally decided that importing three-quarters of the nation’s transformers is a national-security problem, already taxes imported competition with a 15% Section 232 tariff, hits foreign metal content by up to 50%, and has just opened a fresh Section 232 investigation aimed squarely at electrical transformers. Forgent, for its part, just finished a $205 million program that gives its American plants the capacity to generate $5 billion in revenue.

So why can you buy this at ~$44, below the price of a stock deal that closed on Monday?

Because of who is selling it. Forgent is four months old as a public company, and its private equity sponsor has already sold stock four times in five months, at $27, $29.50, $47, and $49. It still holds another $7-$8 billion, and the IPO lockup expires on August 3. The market has stopped treating Forgent as a company and started treating it as a distribution schedule: trade the overhang, wait for the flush, forget the ticker.

We think that reflex is the opportunity. The market is pricing the seller. Policy is pricing the asset.

It is also the rare AI trade that asks you to predict nothing. You do not need to pick the winning model, chip, or cloud. Every one of them ends at the same physical requirement, and Forgent builds the machines that meet it. The electrons do not care who wins.

It is not a clean story, and anyone who sells it to you as one is hiding the file. The same sponsors ran this exact playbook twice before, on Array and on Shoals, and both stocks eventually collapsed into the single digits. That history is part of this trade, and we lay it all out.

Follow the electron

The phrase “electrical equipment” hides how physical this business is, so walk the path an electron takes and watch Forgent collect a toll at every step.

Think of voltage like water pressure. Power leaves a plant at fire-hose pressure, hundreds of thousands of volts, because high pressure is how you move it hundreds of miles without losing it. A server chip drinks through a straw, at about one volt. Everything standing between the fire hose and the straw is a machine somebody has to design, wind, weld, test, and ship. Forgent makes all of them essentially.

First stop, the transformer, the gearbox of the grid. It steps the voltage down, substation by substation, building by building, until power is usable. A large one is a custom-engineered machine the size of a room: miles of copper wound around cores stacked from specialized electrical steel that only a handful of mills on earth produce, then tested for weeks, because the failure mode of a bad transformer is a fire. The industry quotes 24 to 36 months to deliver one. Nothing energizes without them. There is no substitute, no workaround, no way to add them later. They are the reason this entire sector has pricing power, and they are the product Washington is now writing tariffs about.

Next, switchgear, the breaker panel in your garage, scaled up ten thousand times. It routes power and, more importantly, cuts it off in milliseconds when a fault occurs, so a short circuit does not become a burned building. High-voltage switchgear has been quoted at 52 to 65 weeks across the industry.

Then the low-voltage layer, switchboards, panelboards, and PDUs, power strips the size of refrigerators, feeding the server rows. AI density is inflating them in real time: Forgent’s average unit ran 400 to 500 amps two years ago, runs 1,000 to 1,200 today, and the engineers are drawing 1,500 and 1,600 amp designs now. Same buildout, more copper per box, more dollars per megawatt.

Sometimes the whole path ships at once as an e-house, a complete electrical room, transformers plus switchgear plus controls, built in Forgent’s factory inside a steel enclosure and trucked to the site ready to plug in. In one deal, the company closed a 180-plus-megawatt e-house order in 45 days and delivered it in about five months. Hold that anecdote next to the industry’s two-year transformer quote and you have the entire value proposition in one sentence: speed, at scale, on the hardest products.

And when a customer buys the full path as one engineered system, one design team, one schedule, one throat to choke, Forgent calls it a Powertrain Solution. Remember the name. It is where the growth is hiding.

Now, the three facts that explain the economics. About 90% of revenue is engineer-to-order, meaning built to one customer’s spec rather than pulled from a catalog, and the average production batch is 15 units per design. Commodity peers run hundreds or thousands of identical units per design. This is tailoring, not fast fashion, and you do not haggle with the only tailor who can deliver the suit before the wedding. Second, the customer has no choice but to buy, and because the cost of failure is measured in burned buildings and dark data centers, reliability outranks price in every negotiation. It is one of the few corners of industrials where being the best matters more than being the cheapest. Third, Forgent is vertically integrated: it manufactures in-house the components its competitors order from third parties, across campuses in Minnesota, Texas, Maryland, and Tijuana. When the entire supply chain is the bottleneck, owning your supply chain is the product. That is where the lead-time edge comes from, and lead time is why the share gains are real. Management pegs its end markets at around 20% annual growth, and Forgent at three to five times that rate. You do not grow five times faster than your market by being average.

Those end markets, about 85% of revenue, are the three loudest booms in the economy at once: data centers, the grid itself, and energy-intensive industrials like chip fabs and battery plants. The grid deserves one extra sentence, because it is the quiet giant here: it is the largest machine humans have ever built, much of it is forty years old, and America is renovating it while it runs. Three ways to win, one product set.

Know your seller

Read more

Around six months ago, the consensus on Qualcomm was a shrug: a melting smartphone-chip cyclical, collect the licensing checks, trade it around the handset cycle, forget it.

Look at what it is quietly becoming.

A Saudi sovereign, HUMAIN, has a confirmed 200-megawatt deployment of Qualcomm’s AI racks going live this year, scaling toward 1.9 gigawatts by 2030, with a Qualcomm engineering center built in Riyadh and Adobe reportedly its first tenant. ByteDance is the customer behind the “millions of chips” of custom silicon, Bloomberg reported. META picked Qualcomm’s Arm server CPU. NVDA opened its own racks to that CPU through NVLink Fusion. Automotive just printed a record quarter, up 38%. Those are not logos a dying cyclical collects. They are the footprint of a company threading itself through the single biggest buildout of the decade.

And here is what turns a story into a stock: the market is handing you most of it for free.

Underneath the “phone chips” label sits a patent toll booth running ~70%+ margins that throws off close to half of all of Qualcomm’s profit. It collects rent on essentially every 4G and 5G device on earth, even from companies that buy zero Qualcomm chips. The Street treats it as a melting ice cube that dies the day AAPL ships its own modem. It is dead wrong, and we’ll show you from the contract language itself why Apple’s modem cannot kill it.

Now stack it. Put that monopoly on top of an auto-and-edge business compounding 20 to 38%, a $20B buyback program retiring roughly 9% of the shares, and a data-center ramp landing on R&D that is already paid for, and the earnings inflect hard: from this year’s trough near $10.80 to roughly $15.50-16 by FY29. At today’s multiple, that is a $315-340 stock. JPMorgan already sees the shape of it, lifting its target from $160 to $265. Tigress is at $280.

The cleanest way to see the setup: the dividend floors the stock near $114, and the cash-generative core alone, the monopoly plus the chip franchise, is worth $174-200. Everything you pay above that is a cheap option on a data-center business that six months ago was a slide and today is shipping to a sovereign and to ByteDance. The comparable AI-silicon pure-plays trade at $20B (Groq) and $65B (Cerebras). You are getting Qualcomm’s shot at that, plus robotics and 6G, for a fraction of the pure-play tape, all sitting on a core that already protects your downside.

Read more

Look at where the next decade is being built, and you keep bumping into the same company without realizing it.

The satellites stitching the internet across the planet run on its chips. The AI data centers ripping out copper and replacing it with light to feed their GPUs are buying its optical silicon. The devices around you are filled with on-board intelligence running on their processors. The electric car replacing the gas one is packed with its power chips and its sensors. Four of the largest build-outs of the decade, and one chipmaker is threaded through every one of them.

Almost nobody thinks of it this way. The market still files STMicroelectronics under “European auto and industrial chips,” a commodity name to trade around the cycle and forget. That filing is the opportunity because the gap between how it is priced and what it is becoming is enormous.

Underneath the dull label sit two businesses that behave like monopolies. One controls more than 90% of a market growing over 30% a year, which you have never once heard ST mentioned in. The other is a fortress nobody can dislodge, quietly throwing off strong margins while drawing almost no attention. The stock pays for neither of them today.

Then there are the partners. Amazon tied its own equity to a multi-year deal. Nvidia is co-designing with it. SpaceX has built alongside it for a decade. Those are not logos that a commodity chipmaker collects.

And the prize is large. Management has drawn a path from roughly $12 billion in revenue at a 2% operating margin today, to $20 billion at a 30% margin by the end of the decade. Reach even the middle of that, and the earnings are several times what the current price reflects.

This is the full breakdown: the four engines and which two truly matter, the markets underneath them, the one number that unlocks everything, and where we think the stock can go from here.

Read more