中文

AI Overexpansion's Hard Grid Barriers and Energy Arbitrage · Chase Lock Miller

2026-06-11 · A faithful, transcript-grounded reading by PodLens

Original episode:https://youtu.be/4zk-hJ50vmU?si=9d_9Y7wbYavqltYO　·　Timestamps are clickable — they seek the player in place

compute energydata centersdigital laborvertical integrationgrid infrastructure

What This Episode Is About

In Stanford's MS&E435 (Economics of the AI Supercycle) classroom, host Apoorv Agrawal and Crusoe founder and CEO Chase Lock Miller explore the economics and energy supply architecture underlying the data center boom. Opening with the exponentially growing AI capital expenditures of the five major Hyperscalers, the conversation introduces a Cobb-Douglas model of AI as "Digital Labor" and argues that the hard physical constraint has shifted from chip supply to "Energized Data Centers." Chase Lock Miller details the logic behind Crusoe's "Energy-First" strategy in building gigawatt-scale compute campuses in Abilene and Claude, Texas — including on-site capture of cheap wind and solar power, bidirectional across-the-meter grid interconnection, and self-owned natural gas power plants. He publicly discloses for the first time the per-megawatt CapEx/OpEx and revenue return models across physical infrastructure, IT equipment, power systems, and construction labor. The conversation closes with first-principles analysis of older compute commoditization, long-cycle uranium enrichment fuel supply bottlenecks, orbital data center physical barriers, and the disruptive opportunity in electrical equipment supply chains (Eaton, Schneider).

Timeline Theme Map

Core Viewpoints List

  1. AI CapEx's explosive growth is fundamentally the first time in human history that physical infrastructure investment directly creates "Digital Labor." Traditional macroeconomic GDP growth is heavily constrained by national population reproduction and growth cycles (typically a 20-year incubation period), while AI's "Digital Labor" can achieve instant, batch, and infinitely scalable supply expansion by purchasing GPUs and building data center shells. [05:58-06:54] | Type: Viewpoint | Note: Chase argues this is the digital reshaping of Delta L in the Cobb-Douglas production function equation.
  2. The current binding bottleneck of AI compute expansion has shifted from chip supply shortages of previous years to a shortage of "Energized Data Centers / Powered Shells." In the current environment, owning expensive chips does not equal owning compute — finding physical sites that can safely connect to sufficient grid load and power on the chips is the real competitive hard boundary. [09:50-10:27] | Type: Fact | Note: Chase also emphasizes that the bottleneck is dynamically migrating in the supply chain and may shift back to switchgear or cooling units at any time.
  3. Seeking remote, excess "Stranded Energy" outside traditional network hubs like Northern Virginia and adopting "compute in-place, move data" is the most efficient breakthrough approach. Facing grid interconnection delays and red-ocean competition, directly co-locating data centers where renewable energy (wind/solar) is abundant but transmission corridors are blocked can obtain power at extremely low cost and avoid traditional infrastructure congestion. [12:15-12:45] | Type: Viewpoint
  4. Through "Across the Meter" bidirectional grid interconnection arbitrage, hyperscale compute campuses can simultaneously achieve green power local consumption and grid peak-shaving. When local generation is surplus, data centers feed power back to the grid reducing costs for local ratepayers; when local wind/solar output is insufficient or during maintenance, they draw from the grid to firm up load — forming a benign distributed symbiotic system. [28:20-29:08] | Type: Viewpoint
  5. The primary cost inflation in physical data center infrastructure at $20M/MW CapEx comes from skilled trade shortages and oligopolistic control of natural gas power generation equipment manufacturing. Due to industry-wide shortages of electricians, welders, and pipefitters, construction labor costs have been pushed to $4.7M/MW; simultaneously, GE Vernova, Siemens, and a few other giants controlling gas turbine capacity have driven equipment prices to $3M/MW. [23:30-25:53] | Type: Fact
  6. Due to the explosion of intelligent agentic tasks and complex cloud scheduling, the current AI compute foundation faces a severe CPU supply gap. The industry typically focuses on GPUs, but CPUs — the coordination and scheduling hub for compute tasks — are currently experiencing serious supply chain bottlenecks, directly blocking some new cluster deployments. [31:11-31:25] | Type: Fact
  7. Data center evolution toward ultra-fast power distribution will force transformation of the electrical power distribution supply chain, with a long-term bearish view on traditional vendors like Eaton and Schneider lacking innovation. Traditional equipment will lose dominance as 900V DC direct supply and solid-state power electronics emerge; data centers' specific high-current demands will ultimately reset century-old electrical bus designs. [42:19-43:37] | Type: Prediction | Note: Chase adds that due to these traditional vendors' large backlog of pending orders, their short-term financial performance will still be very strong.
  8. Due to the inability to physically swap failed cards in space and extremely challenging thermal management, orbital data centers cannot achieve commercial scale within 5–10 years. Although space computing avoids land costs and grid interconnection regulatory delays, the naturally high physical GPU failure rate without human maintenance capability will directly erode the depreciation economics of orbital modules. [45:41-47:22] | Type: Prediction | Note: The technology's viability entirely depends on whether SpaceX Starship can compress launch payload costs by two orders of magnitude.

Internal Tension and Self-Correction

Plain English Retelling

Let's talk about what Chase Lock Miller shared this session. While everyone was debating how much larger language models had gotten and when we'd reach AGI, Chase and the MS&E435 class looked at the hardest, most physical layer of the world — the power grid, concrete, and cooling pipes.

Chase presented an interesting economic lens: why are the tokens AI generates so expensive and so pursued by capital? Traditional macroeconomics requires labor (L) and capital (K) inputs for GDP growth. But the human brain as labor takes at least 20 years from conception to the point of being able to do meaningful work, and birth rates can't change on a dime. But AI represents the first time in human history we can batch-produce "Digital Labor" directly. Buy GPUs, build powered data center shells, and you can inject labor directly into the economy. This is a fundamental disruption to the production function.

But to get this "Digital Labor" factory running, you're now bottlenecked by the grid. A few years ago everyone was competing for GPUs; now that chip scarcity has eased slightly, the biggest headache has become the "Energized Data Center" — a powered data center shell. Today you can have the money to buy the cards, but without grid access, they're just useless silicon.

Crusoe's core commercial intuition is contrarian thinking: instead of competing on expensive power and congested grid connections in traditional hubs like Northern Virginia, compute in-place and move data. They went to places like Abilene in West Texas — lots of wind, abundant sunshine, renewable developers frantically generating green power to claim government clean energy subsidies, but with no local population to consume it, and transmission networks built too slowly to move it out, so power prices frequently go negative. Crusoe built a 1GW-scale mega-substation right next to the wind and solar generators. The Abilene campus's Stargate project is the joint supercompute cluster for Oracle and OpenAI. Beyond that, Crusoe built a 350MW natural gas plant on-site as stable support for grid fluctuations.

Many people ask: with hundreds of billions invested, can this actually make money? Chase just walked through the math:

  1. Physical infrastructure + power plant (physical CapEx): roughly $20M per megawatt. Within this, construction labor (electricians, welders, and other skilled tradespeople) and natural gas turbines take the largest share, with equipment prices having skyrocketed from $1M/MW to $3M/MW due to GE and Siemens oligopoly control.
  2. GPU and networking (IT CapEx): roughly $40M per megawatt, with $30M going to Jensen Huang (Nvidia), and the rest buying high-speed interconnect networks (NVLink, InfiniBand) and increasingly scarce CPUs.
  3. Total upfront CapEx: roughly $60M per megawatt.
  4. Revenue and payback period: bare chip rental to customers generates roughly $15M/MW annually, about 4 years to payback. But with vertical integration providing managed API services — serving tokens directly to customers — annual revenue can reach $30M/MW, meaning 2 years to payback.

This explains exactly why Crusoe won't settle for being a "landlord" and must build "Crusoe Cloud" on top: only by integrating vertically from "Electrons" all the way to "Tokens" can you capture the most substantial profits as hardware rapidly commoditizes through its depreciation cycle.

Finally, Chase gave very clear-eyed intuitions about future technology evolution. Space data centers (no permitting approvals, direct solar power, optical interconnection) sound beautiful, but the physical constraints are too obvious: failed GPUs in space can't be swapped by human hands, and as hardware naturally dies, depreciation costs become unmanageable — this won't scale commercially in 10 years. The real innovation will actually happen in century-old electrical chains — how to deliver high-voltage power step-down directly into racks at 900V DC. That's the massive blue ocean for the next generation of engineers.

Recommended Segments for Close Listening

Resonances with past episodes

A faithful reconstruction and plain-language retelling of the episode, generated by PodLens.

This is one source-grounded reading, not a replacement for the original. Every point is anchored to its source, so you can check it yourself — and corrections are welcome.