Gross margin by layer of the AI stack. The scarce layers price like monopolies; the labs buying from them run below 50 cents on the dollar.
0%25%50%75%100%
FoundryTSMC
59.5%
Memory · HBMMicron
74.4%
AcceleratorsNVIDIA
75.0%
Model labsOpenAI
30–50%
Model labsAnthropic
30–50%
Reported in filingsOur estimate, shown as a range
Jun 24, 2026 · Micron guided the next quarter to ~81% gross margin, with HBM sold out for CY2026 and booked into 2027. The squeeze is still tightening. (Micron earnings call)
Jun 30, 2026 · Model-vendor margins are expanding, not compressing: capacity-constrained labs price to destroy demand, so falling per-token cost accrues to the labs through 2027. (Dylan Patel, SemiAnalysis)
Sources: TSMC quarterly results, 3Q 2025 · Micron earnings release and call, Jun 24 2026 · NVIDIA Form 10-K, fiscal 2025 · Blended estimate; free-tier ChatGPT subsidy weighs on it · Dylan Patel via The Information: ~$13-14B compute on $20B ARRAs of July 2026Peregrinations
Download the data (CSV)Cite this page: peregrinations.xyz/markets/charts/ai-stack-gross-margins
How to read this
Reported bars are consolidated gross margin from company filings; the pure-plays (TSMC, Micron, NVIDIA) stand in for their layer because none of these companies discloses segment-level gross margin in structured filings. Lab bars are estimates assembled from analyst reporting, shown as a range, not a point. Layers are ordered upstream to downstream.