ELI5/TLDR

Building the warehouses that run AI is one of the biggest construction projects in human history — bigger than the highway system or the Manhattan Project, second only to the US defense budget. Chase Lochmiller, who runs a company called Crusoe that builds these, walks Stanford students through where every dollar goes: roughly $20 million per megawatt for the building and power plant, another $40 million for the chips inside, and only about $15 million a year coming back from renting it all out. The whole game is a bet that the chips and buildings stay valuable long enough to pay it all back — and the trick to making the math work is to stop just renting computers and start selling the finished thing people actually want, which is AI answers.

The Full Story

Why anyone is spending this much

Start with the chart that opens the talk: the five biggest tech companies — the “hyperscalers” — pouring money into AI infrastructure, the line going, in the host’s words, “top into the right and it’s going top into the right fast.” The scale is hard to hold in your head:

This is bigger than space, bigger than our highway system, the Manhattan Project, second only to the US Defense budget.

The natural question is why. Lochmiller’s framing is economic. Imagine the whole economy’s growth comes from three ingredients: more workers, more capital (buildings, machines, money), and better technology that makes workers more productive. (Economists call this the Cobb-Douglas model; the name doesn’t matter, the intuition does.) For all of history, the “more workers” lever has been almost impossible to pull quickly — a new worker takes twenty years to grow, school, and feed. You change the labor supply via the birth rate, and that’s it.

AI changes that. For the first time, you can add labor by buying it.

For the first time in history, what we’re able to do is actually change this delta L digitally through the investment in buying data centers and buying GPUs.

When you hand an AI agent a task — “go build me a CRM for my new product” — that’s a unit of work done by something you bought, not someone you raised. So the trillion-dollar spend isn’t tech-bro exuberance in his telling; it’s an attempt to manufacture labor and goose the entire economy’s growth rate. Whether that pans out is a separate question, but that’s the bet on the table.

What a data center actually is

Strip away the mystique and a data center is almost embarrassingly simple:

A data center is a building that has power and cooling. And you can plug in computers. That’s pretty much what it is.

The complexity is entirely in the scale. Do this for a handful of servers and it’s a closet. Do it for a city’s worth of power and it pulls in every flavor of engineering at once — chemical (the cooling chemistry), mechanical (the chillers and pumps), electrical (taming high-voltage power), and computer science (how the chips talk to each other). Crusoe’s pitch is to be “vertically integrated,” meaning it does everything in this stack except two things: it doesn’t make chips, and it doesn’t make AI models. Everything in between — finding the energy, building the building, wiring the power, deploying the GPUs, serving the final answers — it tries to own.

The reason for owning the whole chain is that the bottleneck keeps moving. Four years ago the scarce thing was chips. Then it was power. Today, in his telling, it’s “energized data centers” — buildings that already have electricity flowing, so you can plug chips in and switch them on. Being vertically integrated means whatever the choke point is this quarter, you can attack it.

Why start with energy, and why Abilene

The founding insight of Crusoe (named, fittingly, after Robinson Crusoe and the idea of scarce resources) was that energy is the constraint nobody else was solving for. The instinct of a normal data-center developer is to build in an established hub — Northern Virginia runs a huge slice of the internet. Lochmiller didn’t want to be “the next data center developer in Northern Virginia.” So instead of moving energy to where the buildings are, he moves the buildings to where the cheap energy is stranded.

Abilene, West Texas, is the showcase. The place is consistently windy and sunny, so renewable developers had overbuilt there, chasing government subsidies that paid them to produce clean power whether or not anyone wanted it. The result was almost comic: so much electricity, and so few wires to carry it elsewhere, that power prices sometimes went negative. There was nobody to buy it.

So we said, great, have we got a power hungry application for you?

The Abilene campus is now one of the largest AI computing campuses in the world — eight buildings, fed by what he says is the largest privately owned substation in the United States, a one-gigawatt monster. For scale: a gigawatt is roughly what powers the entire city of Denver. The campus runs to 2.1 gigawatts in aggregate — two Denvers of electricity, all of it feeding computers. The first eight buildings are for Oracle and OpenAI (the project once known publicly as “Stargate”); an expansion to the south is for Microsoft. To smooth out the renewables, Crusoe also built a 350-megawatt natural-gas plant on site.

One detail worth sitting with: the chips across all eight buildings are wired together as a single coherent machine, so one training job can run across the entire campus at once. The buildings are physically separate; the computer is one.

The human scale, and the labor bottleneck

The campus needs a 5,000-car parking lot, and it’s full, because roughly 9,000 people are on site every day during construction. Abilene’s population is 120,000. A second site he mentions — in a town called Quad, Texas, population 1,500 — has 3,500 people working on it, more than twice the town’s population.

That labor is its own bottleneck, and a surprising one given all the talk of silicon:

We don’t have enough of these tradespeople. We don’t have enough electricians. We don’t have enough welders. We don’t have enough plumbers.

The construction labor alone runs about $4.7 million per megawatt — meaning a single gigawatt site burns $4.7 billion a year in wages. This is blue-collar money, welders and pipefitters and concrete crews (Crusoe pours its own concrete on site, 24/7), and there are now so many of these projects competing for the same tradespeople that wages are climbing.

Where the $100 goes

Lochmiller breaks the spend into two big buckets, normalized per megawatt so the numbers stay comparable.

The building plus power plant runs roughly $20 million per megawatt. Inside that: electrical equipment to step the voltage down from substation levels (34,500 volts) to what a chip can use (480 or 415 volts); a battery system (UPS) to smooth the power; chillers and a closed water loop for cooling; mountains of steel and concrete; the gas turbine; backup diesel generators for the critical storage and networking gear; and that enormous labor bill.

Two things on cooling and gas worth flagging. First, on the persistent worry that AI is drinking the world’s water — his claim is that they fill the cooling loop once and it recirculates, so annual water use is “about the same amount of water as a single family home.” Whether that holds across the industry is a real question, but his architecture recirculates rather than evaporates. Second, gas turbines have roughly tripled in price (from ~$1M to ~$3M per megawatt) because only a handful of companies make them — GE Vernova, Siemens, Mitsubishi, and a few others — and they haven’t expanded production fast enough to meet the AI rush. (He notes, dryly, that it’s “been good to be a shareholder of GE Vernova.”)

The IT side — the actual computers — runs about $40 million per megawatt, and this is where the money goes to Nvidia:

30 million of that is going to the GPUs. That’s why you always look so happy. It’s always smiling Jensen.

The rest: ~$4M for networking (the high-performance fabric that stitches GPUs into one machine — NVLink within a rack of 72 GPUs, then InfiniBand between racks), ~$3M for CPUs and storage (and CPUs are now in surprising shortage, because all the new “agentic” AI workflows need lots of CPUs to orchestrate the work), and ~$1M for labor and shipping.

Add it up: about $60 million per megawatt of upfront capital. Build a gigawatt cluster and you’ve spent $60 billion before a single token of revenue.

Does the equipment hold its value — and how the business actually works

The bear case everyone reaches for: the next chip generation ships and your expensive hardware becomes a paperweight. The data says the opposite. Lochmiller shows the spot price of the H100 GPU — three years old now — and instead of decaying, the rental price rose past its launch level, pushed up by demand from AI agents. Older compute does eventually commoditize, but at any given moment the cutting edge and sheer scale command a premium. On whether GPU margins stay sky-high, he’s candid: Nvidia is at maybe 80% gross margin today, and “the invisible hand of capitalism is a very powerful force” — over time he’d expect it to drift toward a more normal 60%.

Now the core economics. You spent ~$60 million per megawatt. Ongoing operating cost is surprisingly light — a little over $1 million per megawatt a year. If all you do is rent out the bare chips, revenue is about $15 million per megawatt per year — roughly a four-year payback on a $60 million investment. The whole business then hinges on a single accounting question Wall Street is chewing on: over how many years do you depreciate the building, the chips, the power plant? (Most public companies use five or six years for compute; if the H100 price chart is right, that may be too conservative.)

The way you make it a good business, rather than a merely-okay one, is to climb the value chain — the title of his deck is “from electrons to tokens.” Don’t just rent computers; host the model and sell the finished answers. Provide the API endpoint that serves the ChatGPT or Claude query coming off someone’s phone. That managed-services layer adds another $5–15 million per megawatt, pushing revenue toward $30 million and collapsing the payback from four years to two. Crusoe is also trying to attack the upfront cost with a prefab modular data center called “Crusoe Spark,” built in a central factory to cut the brutal on-site labor bill by 30–50%.

Long, short, and space

Asked for a stock pick, Lochmiller hedges (“I’ll call you in a year”), but offers a genuine bear case: the electrical-equipment giants — Eaton, Schneider, and others — that haven’t fundamentally innovated in a century. They’re on the critical path right now and will do well in the near term, but he thinks the data-center boom will force a leap to new architectures (solid-state transformers, 900-volt DC racks, high-voltage power electronics) that could gut their costs over the long run. On the long side, he thinks open-source AI models will take share from closed ones.

And data centers in space? He’s genuinely interested — Crusoe has partnered with a company called Starcloud, which has already put H100s into orbit. The appeal: no concrete foundations, no permitting, no power approvals, and everything interconnected with light instead of fiber. The catch: cooling is brutally hard in a vacuum, and you can’t send an astronaut up to reseat a failed GPU. His verdict: not material for five years, probably not for ten, but likely a major part of the future after that — gated on whether SpaceX gets the cost of launching payload down by a hundredfold.

His closing advice to students sidesteps the “what should I study” panic entirely: focus less on what you learn and more on how you learn, because nobody knows what work looks like in five years when everyone has “the workforce of a million people at our fingertips.” The durable asset is the compounding you get from improving a little every day.

Key Takeaways

A data center is, at its core, “a building that has power and cooling” where you plug in computers. All the complexity is a function of scale, not concept.
The economic logic of the AI buildout: AI lets you increase the labor supply by buying it (data centers + GPUs) instead of waiting 20 years for the birth rate. That’s why the capex is so aggressive and broad-based.
Cost structure per megawatt: ~$20M for building + power plant, ~$40M for IT (of which ~$30M is Nvidia GPUs), totaling ~$60M upfront. Ongoing opex is light, ~$1M+/MW/year.
Revenue from renting bare chips is ~$15M/MW/year → ~4-year payback. Climbing to selling tokens (hosting the model, serving the API) adds $5–15M/MW → ~2-year payback. Moving up the value chain roughly halves the payback.
The whole investment thesis rests on a depreciation question: how many years do the chips, building, and power plant stay valuable? Public companies assume 5–6 years for compute.
GPUs are not depreciating as feared — the 3-year-old H100’s rental price rose above its launch price, driven by agent-workload demand. Scale and the cutting edge always command a premium; old compute eventually commoditizes.
“Energy first” siting: Crusoe builds where stranded cheap power exists (Abilene had negative power prices from overbuilt, un-transmittable renewables) rather than moving energy to traditional hubs. Move the building to the electrons, not the electrons to the building.
The bottleneck migrates: chips → power → energized “powered shells” → now skilled trade labor (electricians, welders, plumbers). Vertical integration is a hedge against wherever the choke point lands.
Gas turbines tripled in price (~$1M → ~$3M/MW) because only ~5 manufacturers exist and none expanded capacity fast enough. Hence GE Vernova’s stock run.
CPUs are now in shortage — agentic AI workflows need lots of CPUs to orchestrate the GPU work.
Water fear is overstated in this architecture: a recirculating closed loop, filled once, uses roughly a single family home’s annual water after that.
Nvidia’s ~80% gross margin is unlikely to last; competition should pull silicon margins toward a more normal ~60% over time.
Bear case on the electrical stack: century-stale incumbents (Eaton, Schneider) face disruption from solid-state transformers and 900V DC architectures — an opening for power-electronics engineers.
Space data centers (via Starcloud) skip foundations, permitting, and power approvals and use optical interconnects — but cooling and un-repairable hardware are hard. Gated on a ~100x drop in launch cost. Not material for 5–10 years.

Claude’s Take

This is unusually honest for a founder talk. Lochmiller is selling Crusoe, obviously — he calls himself “turbo long Crusoe” — but he keeps undercutting his own hype with real numbers, and a couple of times admits he built the slides “this afternoon” and may have double-counted. The four-year-versus-two-year payback framing is the single most useful thing here: it makes concrete why everyone is racing to climb from renting GPUs to serving tokens. It’s the same margin logic as any commodity business moving toward finished goods, just with eye-watering zeros.

The part to hold at arm’s length is the depreciation argument, because the entire thesis leans on it. The H100-price-went-up chart is real and genuinely surprising, but it’s also exactly the data point a man with $60 billion of depreciating assets on the books wants to be true. If chips hold value, his balance sheet is golden; if the next generation actually does crater used-chip prices, the four-year payback stretches and the math gets ugly. He’s honest that this is the open question Wall Street is debating — he just happens to land on the optimistic side. The water claim is similarly his architecture’s best case, not necessarily the industry’s.

The “AI is digital labor, so capex equals economic growth” framing is elegant and probably the cleanest one-paragraph defense of the whole AI buildout I’ve seen — but notice it assumes the labor actually gets used productively, which is the trillion-dollar bet nobody has settled. Score is an 8: dense, specific, refreshingly numerate, and a genuinely good map of where the money flows in AI infrastructure. It loses points only because it’s a sales pitch wearing a lecture’s clothes, and the most load-bearing claim (depreciation) is the one he’s least able to be neutral about.

Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Building AI Factories | Stanford Online