This model is not a capex forecast. It is a demand model expressed in MW-equivalent. The dollar output is the implied capex if 100% of demand were satisfied. Compared to top-down consensus, the difference is the SUPPLY GAP — the size of the physical execution constraint (power, permits, chips, labor) that the AI economy is now hitting. Hyperscalers themselves describe markets as supply-constrained, not demand-constrained (Moody's Apr 2026, Futurum Feb 2026).
Five knobs drive the demand build. All headline numbers in this page reflect the knob settings shown below. Default settings are bolded.
Bottom-up token demand and the demand-implied capex it would require at current chip-mix throughput, workload realization, and $/MW. Cumulative capex through 2030 lands at $7.5T.
| Year | Demand (T tokens) | BU capex ($B) | Cumulative ($B) | YoY Δ |
|---|---|---|---|---|
| 2026 | 149,293 | $794 | $794 | — |
| 2027 | 326,118 | $950 | $1,744 | +20% |
| 2028 | 790,167 | $1,340 | $3,084 | +41% |
| 2029 | 2,042,787 | $1,855 | $4,940 | +38% |
| 2030 | 5,282,855 | $2,605 | $7,545 | +40% |
Bottom-up demand-implied capex compared to the mean of eight authoritative top-down forecasts (Goldman, Dell'Oro, BCG, Bain, Moody's, Futurum, McKinsey, Deloitte). A positive gap means BU demand exceeds what consensus expects will actually be built.
| Year | BU capex ($B) | TD consensus mean ($B) | Supply gap ($B) | Gap as % of supply |
|---|---|---|---|---|
| 2026 | $794 | $573 | $221 | 39% |
| 2027 | $950 | $692 | $258 | 37% |
| 2028 | $1,340 | $836 | $504 | 60% |
| 2029 | $1,855 | $1,004 | $851 | 85% |
| 2030 | $2,605 | $1,208 | $1,398 | 116% |
The gap widens dramatically in 2030 (116% of planned supply) as the BU demand-implied capex steps up while TD consensus follows a smoother trajectory.
The agentic chain is the model's load-bearing demand source. It decomposes across 11 occupational buckets, each with its own workforce size, penetration curve, and per-worker annual token volume. Sorted by 2030 agentic contribution.
| Function | Workforce 2030 (M) | Penetration 2030 | Tokens / worker / yr | 2030 demand (T) |
|---|---|---|---|---|
| F1 Software engineering | 27.3 | 72% | 44.8B | 1,231,803 |
| F5 Healthcare | 37.7 | 58% | 26.4B | 981,952 |
| F4 Knowledge ops | 46.5 | 42% | 18.6B | 581,234 |
| F8 Arts & media | 8.1 | 78% | 52.5B | 428,714 |
| F11 Pharma R&D | 1.0 | 82% | 284.4B | 336,969 |
| F9 Sciences | 6.0 | 60% | 45.0B | 241,552 |
| F3 Sales & marketing | 18.0 | 55% | 16.2B | 209,399 |
| F7 Engineering (non-SWE) | 12.7 | 50% | 21.0B | 173,562 |
| F2 Customer support | 11.7 | 62% | 7.5B | 108,518 |
| F6 Education | 30.7 | 40% | 4.2B | 77,262 |
| F10 Government | 12.3 | 30% | 4.3B | 22,232 |
| Total agentic 2030 | 4,393,197 |
Tokens-per-worker varies ~65× across functions, from F10 Government (~4.3B/yr) to F11 Pharma R&D (~284B/yr — reflecting literature-search and trial-design workloads that scale very differently from chat). SWE, Healthcare, Knowledge Ops, and Arts/Media together contribute over 70% of 2030 agentic demand.
Demand: 5 workload chains — consumer chat (user-based), embedded chat (seat-based), agentic enterprise (labor-based across 11 function buckets via a 6-factor formula), ambient hyperscaler (top-down), and consumer agentic — summed and × induced-demand multiplier (Knob 3).
CapEx: token demand → MW_Bridge (chip-mix throughput × workload realization × derating) → required inference MW → refresh treadmill on installed base × $/MW = inference capex. Training capex layered via Deloitte inference-share trajectory × same $/MW. Total = inference + training.
Knob 5 (workload realization) drives the chip-throughput derate factor: Bear 75% (chat-like workloads), Base 59% (research-derived 60% reasoning + 25% summarization + 15% chat mix), Bull 45% (heavy agentic). Top-down consensus inputs sourced from Z_Sources S190-S200.
22 of 24 integrity checks pass. The two exceptions are a stale 2025-anchor control (being corrected) and an intended high-intensity stress-corner — Pharma R&D × High intensity exceeds the 5T/yr per-worker ceiling at 5.69T — a documented calibration finding whose check rationale anticipates the breach, not a model bug.