Evidence-led market review
AI compute and semiconductor market weekly
Accelerator demand meets a tougher efficiency benchmark
Constructive on delivered AI silicon demand; cautious on assuming the same revenue for each unit of useful output. Efficiency is a competitive variable, not proof of falling total demand.
Edition evidence
Supporting, opposing, and catalyst signals
Broadcom reports AI semiconductor revenue growth of 221% year over year.
OpenAI reports 1.5–1.9x peak tokens per rated chip watt in selected benchmarks; this challenges a fixed hardware-intensity assumption.
The next comparable issuer disclosure and independently measured customer outcomes.
Analyst interpretation: compare useful workload delivery, not peak claims alone. The benchmark range is dated and company-reported; its lower endpoint is the numeric observation.
01 · Executive synthesis
The decision layer
Accelerator demand meets a tougher efficiency benchmark
Constructive on delivered AI silicon demand; cautious on assuming the same revenue for each unit of useful output. Efficiency is a competitive variable, not proof of falling total demand.
04 · Leadership watchlist
Questions for the room
Recheck the cited metric on the same reporting basis.
The next comparable company disclosure falls below this edition's cited level.
Reassess the directional thesis; do not carry it forward automatically. No new disclosure means this test remains unresolved.
Observe by · The next comparable company disclosure falls below this edition's cited level.Compare: compute.broadcom_ai_revenue_growth < 221%09 · Sources & methodology
Public inputs, traceable assumptions
We prioritize official statistics, regulatory filings, investor materials, and provider pricing. AI Atlas estimates are explicitly labeled rather than presented as published facts.
Broadcom via PR Newswire
Third Quarter Fiscal Year 2026 Financial Results
Issuer-distributed mirror of the same quarterly release; not independent corroboration.
OpenAI
The Work Now Within Reach
Company-reported adoption, serving software cost reduction and chip benchmark. Rated chip power is not facility power; benchmark endpoints are not fleet averages.
Model note. Market heat, composite indices, tracking baskets, and selected forecasts are editorial models for directional comparison—not live prices, official statistics, or investment advice.
Publication review. Every Wednesday and Saturday candidate passes automated checks for source availability, value ranges, bilingual completeness, and traceability. Publication proceeds only after an auto-publish decision bound to the exact candidate; source conflicts and anomalous changes are escalated for human review.