SemiAnalysis (Dylan Patel)
Dylan Patel — founder, CEO & chief analyst of SemiAnalysis; went from anonymous Silicon-Twitter chip poster to one of the most-cited AI-infrastructure analysts, cited on stage by Jensen Huang at GTC.
The single most-followed semiconductor/compute KOL of this cycle — SemiAnalysis is the most-subscribed technology newsletter on Substack (250k+ subscribers) and advises labs, hyperscalers and hedge funds. Patel's influence on the AI trade runs through his calls and benchmark reports (InferenceX, CoWoS/HBM/power bottleneck tracking), which move sentiment on specific names. Listed here for his signature theses, NOT a portfolio.
- Not a fund — A research, media or advisory business with no managed book to score.
Top holdings
bullish call — His signature "silent winner" thesis — builds Google's TPU and Meta's MTIA, capturing custom-silicon volume without brand risk. A published CALL, not a disclosed holding.
bullish on TPU — Argues TPUv7 ("900lb gorilla") beat Nvidia to N3 and is scaling fast on internal + external (Anthropic) demand. Thesis/call, not a position.
structurally cautious — Argues the GPU "monoculture" is increasingly vulnerable to Trainium 3, AMD MI450X and Google TPU unless Nvidia sustains extreme co-design. A directional view, not a short position.
MI450X cited as a credible rack-scale challenger; a competitive-dynamics call, not a disclosed holding.
AI-cluster connectivity (active electrical cables / SerDes)
Recent moves
2026: published the InferenceX benchmark (cited on stage by Jensen Huang at GTC); pushed the Broadcom "silent winner" and Google TPUv7 theses; argued the AI bottleneck has rotated back to semiconductor fabs in 2026 and that hyperscalers face a "Pascal's Wager" forcing capex over profitability through 2027 (newsletter.semianalysis.com / latent.space, 2026).
Our take
By far the highest-reach, most market-moving voice in this cluster — when Patel publishes a benchmark or a custom-silicon thesis, it reprices sentiment on AVGO/GOOGL/NVDA, and Jensen Huang citing him on stage is the credential. The essential honesty caveat: he runs NO public book, so unlike the other three entries there are no positions or returns to track — his entry surfaces calls, not holdings, and we deliberately did not invent a portfolio. He's a thesis/flow signal of the highest quality (custom silicon and power over GPU monoculture), but cannot be scored on realized P&L.
The interesting thing marketing won't tell you is where the money is.
The revenue mix isn't publicly broken out — institutional models, dashboards and consulting are reported alongside subscriptions (The Information, Apr 2026) — so whether the free post is the product or the distribution isn't knowable from outside. That opacity is the point: the commercially valuable work is the part you don't see. Second, the reflexivity — InferenceX numbers were used on NVIDIA's own GTC 2026 keynote stage. A benchmark adopted into a vendor's marketing is a genuine credential and simultaneously a constraint, because the benchmarker now has a commercial relationship with the party the benchmark flatters. Third, and most under-priced: there is no scoreboard. Reach is measurable; accuracy is not. The unresolved litigation below bears on disclosure and conflict-management practice, not on the published analysis.
Thesis
SemiAnalysis is not a fund and has no book — treating it as one is the first analytical error. It is a research and data business whose public newsletter sits at the top of a funnel for enterprise subscriptions, models and consulting sold to hyperscalers, labs and hedge funds. The right critique is therefore of a research provider: is the edge (supply-chain access) durable, is the incentive structure disclosed, and can any of the calls actually be scored? On the third question the honest answer is no.
Sell primary-source AI-infrastructure intelligence — fab/ODM/supply-chain channel checks, 14+ industry models (accelerators, HBM, AI-cloud TCO, foundry, memory, tokenomics), and open benchmarks (InferenceX, ClusterMAX) — free and viral at the top, priced at the institutional tier underneath. Reported ~$20M revenue in 2025 against a >$100M 2026 projection with ~60 staff (The Information, Apr 2026).
Assessment
- Primary-source access — fab, ODM, memory and power-chain channel checks — rather than re-reading sell-side notes. That is a real, hard-to-copy input.
- Open-sourced benchmarking (InferenceX, formerly InferenceMAX; Apache-2.0 on GitHub) is falsifiable in a way most paid research isn't — methodology and code can be inspected.
- Early on the structurally correct rotation: custom silicon, networking, HBM and power as bottlenecks rather than a GPU-only frame.
- Speed and specificity — named parts, named fabs, quantified capacity — at a granularity generalist tech media does not attempt.
- No scoreable record. Calls are rarely stamped with an entry date, a price, a horizon or an exit, so a hit rate cannot be computed. Reach is being mistaken for accuracy.
- Three roles in one entity — publisher, paid consultant to covered companies, and reported personal investor. We found no editorial-independence or conflict-of-interest policy on semianalysis.com (fetched Jul 2026).
- The edge is access, and access is revocable. A vendor that dislikes a report can withdraw the channel that produced it; the incentive runs toward not finding out.
- Key-person concentration is near-total. The brand, the sourcing relationships and the market-moving authority are Patel's, not the ~60-person firm's.
- Benchmark reflexivity: once a vendor cites your benchmark in its own keynote, publishing an unflattering result for that vendor carries a commercial cost it did not previously carry.
Record
There is no performance to report and we did not invent any. SemiAnalysis runs no fund, discloses no positions, files no 13F, and publishes no return series — the tickers associated with this entry (AVGO, GOOGL, NVDA, AMD, CRDO) are subjects of published theses, not holdings. That distinction is the whole point: every other entry of this type can be scored against a lagged but real book, and this one cannot be scored at all. What is measurable is the business, not the calls — ~$20M revenue in 2025 against a >$100M 2026 projection with ~60 staff (The Information, Apr 2026), and a subscriber list the company's own site states at 180,000+ (fetched Jul 2026). Treat commercial growth as evidence of demand for the research, not as evidence the research was right. Anyone using these theses as input needs their own attribution: date the call, mark it, and score it yourself.
Risks & fit
- SemiAnalysis filed a trade-secret and contract action in SF Superior Court in March 2026; a former employee's wrongful-termination counter-action followed in April. Both unresolved.
- Vendor dependence: a single major supplier cutting access could degrade the differentiated input overnight.
- Scaling from ~$20M to a >$100M target with ~60 staff strains the quality control that made the name; more output, thinner sourcing.
- Crowding — when 180k+ readers get the same channel check, the informational edge in any single call compresses fast.
- Competitive encroachment from sell-side semis desks and paid expert networks selling the same supply-chain primary sourcing.
Two things would change this read. Positive: a published conflict-of-interest and client-disclosure policy, plus a date-stamped auditable call log with entry dates and outcomes — that converts an unscoreable reputation into a measurable record and largely retires the concerns above. Negative: a report materially unfavourable to a large paying client, or the reverse — evidence that coverage systematically softens on clients — would settle the independence question either way. Absent both, research quality looks high and the incentive structure simply isn't disclosed enough to verify.
Readers wanting granular AI-infrastructure supply-chain detail — capacity, packaging, memory, power, interconnect — at a specificity generalist media doesn't reach, and who do their own attribution rather than treat a thesis as a signal. Institutional buyers should price in that the public newsletter and the paid product differ. Anyone seeking a track record won't find one here. This is a critique of a research firm, not a view on any security it covers.
Retail Substack access is reported at roughly $500/yr (The Information, Apr 2026), not confirmable on a public pricing page; most of the 180,000+ list is free-tier. Institutional models and consulting aren't publicly priced.