Topic research

AI 产业的信号与噪音

Quantified, source-graded tracking of the AI industry across seven boards. Claims carry the A1–A5 evidence-maturity ladder: committed (A2), physically auditable (A3), and priced daily (A4) — before end-demand economics (A5) are proven.

Open canonical wiki

Boards

Accelerator siliconGPU / TPU / ASIC / wafer-scale — public tracker
Compute capexHyperscaler capex, guidance language, Stargate-class commitments
Supply bottlenecksTSMC monthly revenue, HBM, CoWoS, data-center power
Tokenomics$/M tokens, OpenRouter mix, GPU rental spot pricing
Models & frontier labsRelease cadence, benchmark-per-dollar, lab revenue
Agents & memoryCaching economics, context pricing, memory infra
Applications & revenueAI-attributed revenue — the A5 falsification layer
China AI stackAscend/Cambricon/SMIC, domestic token price war, policy

Companies

NVIDIAentity page
AMDentity page
Google TPUentity page
Huawei Ascendentity page
Cambriconentity page
Cerebrasentity page
Custom ASIC (Trainium / Maia / MTIA)entity page

Core data series

Hyperscaler capex (quarterly), TSMC monthly revenue (~10th each month), GPU rental spot pricing, and token pricing — the four hard sequences everything else must reconcile against.

Evidence posture

Every artifact updates a series table or a canonical claim; no narrative-only research. Signal is what moves a series or a claim stage — everything else is noise and is not logged.