From Electrons to Tokens: The Physical Economics of AI Factories

In the third lecture of Stanford MS&E 435, guest speaker Chase Lochmiller (Co‑founder & CEO of Crusoe) explains the end‑to‑end economics of AI factories, tracing the chain from electricity to token revenue.

Key Insights

GPU accounts for only 50% of full‑stack cost. The other half consists of transformers, cooling systems, concrete, steel, natural‑gas turbines, and construction labor, meaning market focus on NVIDIA is disproportionate to its cost share.

The bottleneck has shifted from chips to the "Power Shell". Chip supply improves, but sufficient power, cooling, and physical space to host chips remain scarce.

Labor is the largest hidden bottleneck. Blue‑collar workers cost $4.7 M per MW, competing with semiconductor fabs and infrastructure projects for electricians, welders, and pipefitters.

Agent demand inverts the GPU price curve. Sustained demand for AI agents underpins the economic model; if agent product‑market fit fails, price support collapses.

Control of electricity equals control of tokens. Owning power infrastructure brings high returns but also concentration risk.

CapEx Explosion

Five hyperscalers’ AI CapEx is projected to rise from ~$150 B in 2023 to ~$650 B by 2026 – a three‑fold increase, surpassing most historic U.S. investments outside defense.

Why Tokens Are Valuable

Using a Cobb‑Douglas production function, GDP growth = ΔLabor + αΔCapital + (1‑α)ΔTechnology. Tokens act as digital labor, directly boosting ΔLabor, and indirectly accelerate ΔTechnology through large‑scale training, making tokens the first input that simultaneously accelerates two GDP drivers.

Cost Breakdown: $60 M per MW

The full‑stack cost of an AI factory is $60 M per MW, split into:

IT layer ($40 M per MW): GPU $30 M, networking $4 M, CPU & storage $3 M, rack equipment $3 M, deployment & logistics $1 M.

Data‑center + power layer ($20 M per MW): Labor $4.7 M, soft costs $1.5 M, natural‑gas turbine $3 M, tenant fit‑out $2.6 M, electrical equipment $1.8 M, mechanical equipment $1.7 M, materials $3.9 M.

Labor alone dominates the non‑GPU cost, and a 1 GW cluster would require $47 B in labor alone (≈9 000 workers on site).

Rising Costs

Costs are increasing: natural‑gas turbine prices rose from $1 M to $3 M per MW due to limited capacity expansion, and electrician wages are climbing as AI data centers, chip fabs, and infrastructure projects compete for the same workforce.

Power Shell Bottleneck

"The current bottleneck is a powered shell—an enclosure with sufficient electricity, cooling, and networking to host chips. Chip supply has improved, but finding space that can be powered and cooled remains the limiting factor."

Vertical integration (as pursued by Crusoe) mitigates this by controlling the entire power‑to‑compute stack.

GPU Price U‑Curve

H100 spot/lease prices fell in 2024‑25 with increased supply, then rebounded sharply as agent demand surged, surpassing launch levels. Blackwell one‑year contracts dropped to a low in mid‑2025 and rose >25 % by early 2026, driven by agent demand.

CPU Shortage

Agent workloads also demand large CPU resources for orchestration, causing a hidden CPU shortage. IT CapEx allocates $3 M per MW to CPU/storage; if CPUs become scarce, they could limit agent scaling.

Strategic Bypasses

Crusoe: Vertical Integration + Energy‑First

Crusoe builds data centers in regions with excess renewable power (e.g., West Texas) and pairs a 1 GW private substation with a 350 MW gas turbine for firming. Their "Energy‑First" approach avoids congested grids and leverages negative‑price wind.

Crusoe’s modular "Crusoe Spark" prefabricated data‑center units reduce construction labor by 30‑50 % and cut weather‑related delays.

IREN: Time‑Window + BTC Balancer

IREN secures power pipelines (≈6 GW total) and builds data‑center parks with direct renewable feed‑in, achieving $0.028‑0.035 /kWh electricity cost. Their BTC mining arm balances power usage, providing revenue when AI demand is low.

CoreWeave: Light‑Asset Model

CoreWeave leases colocation space and focuses on GPU procurement. When bottlenecks shift to physical infrastructure, they face delivery delays and lack control over upgrades, exposing a vulnerability.

Common Risks

All three players share concentration risk: Crusoe binds to Oracle/OpenAI & Microsoft, IREN to Microsoft & NVIDIA, CoreWeave to a few large cloud customers. Heavy, specialized assets limit short‑term repurposing, effectively guaranteeing demand for hyperscalers.

Economic Model

Assumptions:

GPU assets depreciate slowly (effective life >6 years).

AI agent demand continues to grow, sustaining high GPU utilization.

Data‑center and power‑plant assets have 20‑30 year depreciation.

OPEX (electricity, insurance, maintenance, GPU replacement) stays within $1‑2 M per MW per year.

Financials per MW:

CapEx: $60 M (IT $40 M + infrastructure $20 M).

Revenue: $15 M‑$30 M per MW per year (GPU rental alone $15 M; adding managed services up to $30 M).

OPEX: $1‑2 M per MW per year.

Payback: 4 years (rental‑only) or 2 years (managed services).

The model shows a heavy‑asset, long‑duration cash‑flow profile more akin to utilities than venture‑backed startups.

Conclusion

While the physical infrastructure for token production is now in place, converting that capacity into ten‑fold productivity gains remains uncertain, prompting the next lecture to explore why enterprise output has not matched the $700 B AI investment.

References

Chase Lochmiller, “From Electrons to Tokens,” Stanford MS&E 435: Economics of the AI Supercycle, Spring 2026.

SemiAnalysis, “GPU Rental Pricing Trends (1‑year contract),” 2026.

Agrippa Investments, “IREN: A Hyperscaler in the Making,” 2025.