Operating rules
The product has two jobs: compare investable ETF expressions of the AI stack and keep the living thesis explicit, sourced, and falsifiable.
← Back to ETF monitor01 · Construction
SMH, SOXX, AIS, DRAM, and DISK are measured as exchange-traded funds. Adjusted close is preferred when the provider supplies it, with raw close as the explicit fallback. Fund fees remain embedded in observed performance.
fund return = ending adjusted close ÷ starting adjusted close − 1
The five funds span broad semiconductors, cross-stack AI exposure, and memory. Their mandates overlap and their inclusion is not an allocation recommendation. Each mandate, inception date, fund size, and expense ratio links to its sponsor.
Fund size is the sponsor-reported AUM or net assets on the labeled as-of date. Expense ratio is the sponsor’s current published total or gross ratio. These source-backed facts are stored separately from price history and are not inferred by the market-data provider.
7D and 30D are the trailing 7 and 30 calendar days. 1Y, 2Y, and 5Y use exact calendar anniversaries. The calculation starts on the first observed trading close on or after the cutoff. YTD requires a prior-year close; inception starts with the first valid fund close. A window disappears when the fund lacks the full history.
The ChatGPT window begins on the first trading session on or after November 30, 2022, the official launch date. OpenAI source ↗
VOO represents broad U.S. equities; QQQ represents the Nasdaq-100 growth cohort. Both compound over the fund’s exact comparison window. If a benchmark did not yet exist, that relative window is omitted rather than backfilled.
relative return = (1 + fund return) ÷ (1 + benchmark return) − 1
02 · Data
Production requests maximum available daily history through the existing server-side market-data provider interface. Responses are cached briefly, credentials are not required, and missing symbols remain visibly unavailable. All return arithmetic is deterministic server code.
Yahoo Finance chart data ↗03 · Research
Power
Capacity additions, standardization, or project delays remove the scarcity premium faster than demand grows.
Suppliers
Foundry and memory capital intensity falls structurally or tool reuse materially lengthens replacement cycles.
Packaging
Capacity catches demand, yields normalize rapidly, or architectures reduce packaging intensity.
Semis
Order concentration, inventory, or capacity additions turn the basket into a lower-quality proxy for AI demand.
Memory
HBM commoditizes quickly, conventional memory weakens, or aggressive capacity recreates a severe downcycle.
GPU
Utilization falls, alternatives erase platform advantages, or efficiency gains sharply reduce required accelerator units.
CPU
Accelerator integration or custom silicon compresses host CPU content and economics faster than cluster volumes grow.
Network
Standardization, integration, or slower cluster growth commoditizes switching and optics faster than volumes expand.
Neoclouds
Customer concentration, refinancing needs, or falling compute prices expose weak unit economics or stranded capacity.
Lab proxies
Governance, dilution, capital commitments, or a weak commercial link make the public security a poor proxy for frontier-model value.
Working five-layer map
04 · Governance