ATLAS

Where the AI dollars flow

Six scenarios trace different revenue and capital flows through the stack: training, inference, enterprise software, consumer AI, sovereign AI, and the physical data-center build-out. Widths are directional — not GAAP — and meant to show which layers actually capture incremental dollars under each story.

Training value chain

Frontier-model training dollars start at hyperscalers and frontier labs, flow into GPU clusters, then through advanced packaging, HBM, leading-edge wafers, and ultimately into equipment, raw inputs, and grid infrastructure.

Model training spendCloud / GPU cloudsFrontier labsData centersAI serversAccelerators / GPUsHBM memoryNetworking & opticsAdvanced packagingFoundriesEquipmentRaw inputsPower & cooling

Layer breakdown

Where value lands in this scenario — share of total flow captured at each layer.

Total flow weight: 595
LayerCapturedShareDistributionPassed downstream
Cloud / GPU clouds9516.0%
85
AI servers7512.6%
90
Foundries7512.6%
50
Frontier labs6010.1%
55
Data centers6010.1%
80
Accelerators / GPUs6010.1%
60
Equipment508.4%
15
Advanced packaging406.7%
15
Power & cooling305.0%
5
HBM memory203.4%
30
Raw inputs203.4%
0
Networking & optics101.7%
10

Flow widths are illustrative weights, not reported revenue. They convey directional capture so you can compare scenarios side-by-side. Quantitative values for any layer should be reconciled with primary filings before use in any decision.