Class #3 | MS&E435: Economics of the AI Supercycle Stanford University Spring '26 Apoorv Agrawal

ELI5/TLDR

Chase Lockmiller, founder of Crusoe, walks a Stanford class through what it actually costs to build an AI data center, line by line. The hyperscalers are spending $650B on AI infrastructure — more than the Manhattan Project, second only to the US defense budget. A single gigawatt of AI capacity costs roughly $60 million per megawatt to build ($60B total), earns $15-30M per megawatt per year, and pays back in two to four years if the chips hold their value. The bottleneck keeps moving — first chips, now powered land and blue-collar labor.

The Full Story

Why this money is being spent

The framing Lockmiller offers is the Cobb-Douglas production function — GDP grows through changes in labor, capital, and technology. Labor is normally the hardest to move. A baby takes roughly twenty years to become a useful worker. For the first time in history, money spent on data centers and GPUs is directly buying new units of labor, just digital ones. Every capex dollar is being poured into the delta-L term of the equation.

“For the first time in history, what we’re able to do is actually change this delta L digitally through sort of the investment in buying data centers and buying GPUs and really accelerating the growth in the economy by actually accelerating the growth in digital labor force.”

That framing — capex as a lever on the labor supply — is what makes the scale of spending rational. The pot at the end of the rainbow is not chatbots, it is GDP growth.

What a data center actually is

Strip away the mystique: a data center is a building with power, cooling, and a slab you can plug computers into. The engineering complexity comes from doing this at the scale of a small city’s electricity use, concentrated in one site, with millisecond-level reliability. The campus Crusoe is building in Abilene, Texas — the original Project Stargate site for Oracle and OpenAI — is 2.1 gigawatts. For context, one gigawatt is roughly the power draw of the entire city of Denver. So the campus is two Denvers’ worth of electricity, running GPUs.

Why Abilene

Crusoe’s insight from day one was that the scarce input in Western-world computing is energy, not compute. The old data center map — Northern Virginia, Silicon Valley — was built for Web 2.0 traffic. AI training doesn’t care about latency to end users, so you can put the building wherever the electrons are cheap.

West Texas, it turns out, is structurally oversupplied with renewables. Federal production tax credits pushed wind and solar developers to overbuild, and transmission couldn’t move the power out of the region. Prices sometimes went negative — producers paying the grid to take electricity. Crusoe walks in and says: we are the buyer you were missing.

The Abilene campus has the largest privately owned substation in the US (one gigawatt), a 350-megawatt on-site natural gas power plant, and 9,000 construction workers on site in a town of 120,000 people. Another Crusoe site is in Claude, Texas — population 1,500, site headcount 3,500. The site is literally outnumbering the town.

Where the $60 million per megawatt goes

Lockmiller splits it into two chunks.

The building + power plant: ~$20M/MW

Electrical equipment (transformers, switchgear, power distribution centers) stepping 345kV down to 480V
Cooling (chillers, a million gallons of recirculating water per building, pumps, pipe fitters)
Structural — steel, concrete poured 24/7 from an on-site batch plant
The gas turbine itself — used to cost $1M/MW, now $3M/MW because GE Vernova, Siemens, Mitsubishi and Pratt & Whitney can’t make them fast enough
Labor: $4.7M/MW. For a gigawatt, that is $4.7 billion in wages during construction alone. This is capitalized labor — it goes on the balance sheet, not the income statement.

The IT fit-out: ~$40M/MW

GPUs: $30M/MW — three-quarters of the compute spend, which is why Jensen Huang is always smiling
Networking: $4M/MW — NVLink within a rack, InfiniBand or RoCE between racks, stitching 72-GPU trays into one coherent cluster
CPUs and storage: $3M/MW — currently short supply, because agentic workloads need CPUs to orchestrate the GPU jobs
Tenant fit-out and labor: ~$4M/MW

The bottleneck keeps moving

Four years ago it was GPUs. Two years ago it was power. Today, Lockmiller says, it’s powered shells — actual buildings with energized substations that can accept a chip shipment. Chips have loosened. Trades people — electricians, welders, plumbers — have not. A project like Abilene competes for the same labor pool as every other data center going up, and the labor pool was not built for this.

“We don’t have enough of these trades people. We don’t have enough electricians. We don’t have enough welders. We don’t have enough plumbers.”

The payback math

Spend $60M/MW. Rent the chips out at roughly $15M/MW/year. That’s a four-year payback on revenue — shorter once you subtract $1-2M/MW of ongoing opex (power, maintenance, cable replacement). Stack a managed inference service on top — hosting models, serving tokens, charging per API call — and you add another $5-15M/MW, getting to roughly $30M/MW/year and a two-year payback. This is why every infrastructure company is trying to climb the stack toward model-serving.

The wall-street-critical question is depreciation. Most hyperscalers depreciate GPUs over six years. The bear case says next-gen silicon makes last-gen worthless. Lockmiller points to a Bloomberg chart of H100 spot prices: after initially falling, they have recovered and now trade above their launch price, because agentic demand picked up everything the training labs dropped. Blackwell is tracking the same curve. If the useful life is longer than six years, every hyperscaler’s P&L gets prettier.

Commodity or not

Asked whether compute is becoming a commodity, Lockmiller hedges. Older generations will commoditize. Scale and cutting-edge silicon will not — running a two-gigawatt coherent cluster is genuinely hard to replicate. But he expects gross margins on GPUs to drift from today’s ~80% toward a more normal ~60% as competition arrives.

“The invisible hand of capitalism is, you know, a very powerful force.”

The stock tip

Asked for a long and a short, he offers the short first. The electrical stack — Eaton, Schneider, the switchgear and transformer incumbents — are on the critical path right now and will print money in the near term. Long term, if they don’t innovate, newcomers building solid-state transformers and 900-volt DC architectures will eat them. On the long side, he’s cagey except for two throwaway bets: open source models will keep gaining share, and data centers in space (he’s partnered with StarCloud) are real, but not material for at least a decade. Thermal management and the impossibility of sending an astronaut to reseat a failed GPU are the blockers.

Advice to Stanford students

The specific subject matter doesn’t matter. The process of learning does. Tenacity compounds. In five years, everyone will have “the workforce of a million people at our fingertips,” so focus on the how, not the what.

Key Takeaways

AI capex framed as a Cobb-Douglas trick: for the first time, capital spending directly buys labor supply (digital labor), compressing the ~20-year lead time of growing humans to zero.
Full-stack cost of one gigawatt of AI compute: ~$60 billion. ~$20B for the building and power plant, ~$40B for the chips, networking, CPUs, and storage.
GPUs are roughly 75% of the IT capex ($30M of $40M per megawatt). Nvidia’s gross margins are ~80%, likely to compress to ~60% as competition arrives.
Labor during construction is $4.7M/MW — capitalized, not opex. At gigawatt scale that’s $4.7B in wages baked into the balance sheet.
Gas turbine prices have 3x’d ($1M to $3M per MW) because only four or five companies make them (GE Vernova, Siemens, Mitsubishi Heavy Industries, Pratt & Whitney, Caterpillar’s Solar). Explains GE Vernova’s stock run.
Crusoe’s core thesis: site data centers where energy is stranded. Abilene had negative power prices from overbuilt wind/solar with no transmission out. Colocation lets Crusoe move data instead of electrons.
“Across the meter” model: generate power on-site, sell surplus to the grid (lowering local rates), draw from the grid when wind/solar isn’t cooperating. Mutually beneficial rather than parasitic.
One million gallons of water per building, but recirculating — annual consumption equals a single-family home. Kills the “AI is drinking the rivers” narrative for liquid-cooled sites.
The bottleneck migrates: chips → power → powered shells → trades labor. Each phase of the cycle is a different arbitrage.
H100 spot prices went up, not down, three years after launch. Agentic demand absorbed the chips the training labs stopped wanting. Implications for depreciation schedules: useful life may be longer than the 6-year standard.
Revenue economics per MW: $15M/year renting raw compute → $30M/year if you host and serve models. The managed-services layer halves payback from ~4 years to ~2.
Only four workloads concentrate enough demand to justify custom data center economics: AI training, AI inference, crypto proof-of-work, and large-scale scientific computing. Web apps don’t need this.
Nvidia’s GB200/GB300 rack has 72 GPUs on one NVLink domain. Racks interconnect via InfiniBand or RoCE (RDMA over Converged Ethernet). Scale is vertical within a rack and horizontal across them.
Data centers in space: real partnership with StarCloud, first H100s already launched, but thermal management and field repair make it a >10-year bet.
The electrical stack (Eaton, Schneider, etc.) is the likeliest disruption target — hasn’t innovated in a century, sitting in the critical path, ripe for solid-state transformer and 900V DC displacement.
Crusoe Spark: modular, factory-built data centers in 500kW (air-cooled) and 2MW (liquid-cooled) units. Cuts on-site labor by 30-50%. Fleet-deployable, opens up smaller stranded-power sites that can’t justify a gigawatt campus.

Claude’s Take

This is the cleanest live dissection of AI infrastructure economics you’ll find in under an hour. The “electrons to tokens” framing isn’t original, but watching someone price each line item with real numbers from a site they’re currently building is genuinely useful. The $60M/MW full-stack cost and the ~4-year payback are the two numbers worth memorizing — everything else is commentary.

Two ideas stand out. First, the reframing of AI capex as a direct purchase of labor supply via Cobb-Douglas is a more serious argument than the usual “AI will boost productivity” hand-waving. If you accept it, the capex chart stops looking insane — it starts looking like the US is buying millions of workers in the most efficient way ever available. Second, the depreciation question is where the whole hyperscaler investment case lives or dies. Lockmiller’s H100 price chart is the single most valuable data point in the talk. If GPUs hold value for ten years instead of five, every hyperscaler’s return on invested capital doubles.

The BS filter: Lockmiller is obviously talking his book — he is the vertically integrated AI infrastructure guy, so of course he tells you vertical integration wins and the bottleneck conveniently lives where Crusoe operates. His depreciation optimism cuts conveniently in favor of his business model. His short call on Eaton and Schneider is notably hedged (“they’ll do great in the near term, my partners, don’t quote me”) — he knows he has to work with them on every site. The space data centers cameo is a founder being polite about a partnership rather than a considered thesis.

Score: 8/10. Loses a point for the disorganized second half and the founder’s obvious incentive alignment. Earns its keep on the cost breakdown alone — this is the kind of ground truth that doesn’t show up in analyst notes.