Supply Chain

2 corrections · latest Jul 15, 202610d

Who supplies whom across the AI compute chain, where the structure concentrates, whether that concentration is tightening, and what is moving through it now. Named-entity bilateral dependencies rather than market-share proxies; the hourglass of single-supplier and single-jurisdiction chokepoints; a four-dimension stress reading with adjustable weights; the semiconductor cycle that sets the near-term ceiling; and the forward supply model of what is buildable.

Supply Chain

2 corrections · latest Jul 15, 202610d

Compute Supply Model

How much AI compute is buildable in the coming years, and which physical limit binds at each horizon? Three independent supply curves — power availability, chip production, advanced packaging — are projected forward from current baselines; the lowest curve at each quarter is the binding constraint, and it shifts over time as different bottlenecks resolve. Assumptions (PJM approval rate, construction-delay factor, chip-yield improvement, CoWoS capacity growth, GPU density per MW) are adjustable; confidence bounds widen from 30% to 60% over the five-year horizon, with no projection beyond the cited papers’ horizon. Controls aimed at a non-binding constraint won’t reduce supply; controls at the binding one will.

No date

Sources

Epoch AI
FRED
WSTS

Processing

Three independent supply curves (power, chip production, packaging) projected forward from current baselines. The binding constraint at each quarter is whichever supply curve is lowest. Power: RAND RRA3572-1 baseline (21 GW) + PJM queue additions × approval rate. Chips: Epoch AI chip sales baseline (13.7M H100e/yr) × yield improvement rate, cumulative. Packaging: TrendForce CoWoS wafer capacity (780K/yr) × growth rate × estimated chips-per-wafer (6, not disclosed by TSMC). Installed compute baseline: 20.3 ZettaFLOPS dense BF16 (~20.5M cumulative H100e × 989.5 TFLOPS ≈ 2.03×10²² FLOP/s), the same per-chip convention used on /flop-engine and /cost-index. Epoch AI expresses the same fleet against H100 FP8-dense (1,979 TFLOPS/chip), which reads as ~40.5 ZettaFLOPS; the figures describe identical hardware in different unit systems. Confidence bounds widen from ±30% (near-term) to ±60% (5-year). All projections are model outputs, not forecasts — they depend on the adjustable assumptions shown.

Stress Index

Current measurement of the constraints projected here

Semiconductor Cycle

Historical context for the projections

EconomicsCompute & Capability

Pricing implications of supply constraints

Sovereign AICompute & Capability

Sovereign program demand layered onto the projection