SemiAnalysis in Detail

[ikaijua]

SemiAnalysis in Detail

An independent intelligence shop for the chip and AI-infrastructure industry, and among the most-cited first-hand sources for tech giants and investors.

What it is. Founded by Dylan Patel in 2020 (started as a one-person newsletter); ~60-85 people in 2026; covers the full semiconductor and AI supply chain.
Signature capability. Proprietary supply-chain data plus self-built models (Accelerator/HBM, datacenter models, ClusterMAX, InferenceX); called out by Jensen Huang at GTC 2026.
Readership. #1 technology publication on Substack; 300K+ subscribers (mid-2026).
Pricing. Paid subscription ~$500/yr (20% group discount); institutional data products quoted separately.
Caveats. Figures are mostly model estimates, not official data; the founder is also an investor - judge conflicts yourself.

Below, floors are organized by: what it is -> content & products -> how to use it -> FAQ & resources.

Table of Contents

What it is

1. What is SemiAnalysis?
2. Why people listen (method & influence)

Content & products

3. Content & pricing
4. Data models & benchmark products (institutional, custom-priced)
5. Signature reports (famous and contested)

How to use it

6. Who reads it & how to use it
7. Value and caveats (honest assessment)

FAQ & resources

8. FAQ
9. Sources

Click a TOC link to jump to its floor; if the in-page jump does not work, open the link in a new tab.

[ikaijua]

What is SemiAnalysis?

SemiAnalysis is an independent research firm covering semiconductors and AI (a newsletter plus consulting), tracking the full supply chain from chipmaking equipment and materials, foundry and packaging, to servers, datacenters, AI models, and inference infrastructure. Its tagline: "bridging the gap between business and the world's most important industry."

• Founder Dylan Patel: started as a one-person newsletter in 2020; no semiconductor degree (studied data analytics / risk management); wrote anonymously in the "silicon Twitter" scene for years before going public; broke through with reports on GPT-4's architecture and DeepSeek
• Team: ~12 people in 2024 → ~60-85 people across ~10 countries in 2026 (San Francisco, NYC, Singapore, ...), with its own hardware teardown lab (STEEL Team)
• Readership: 编程工具可以添加aws提供的codewhisper #1 technology publication on Substack, 300K+ subscribers as of mid-2026
• In one line: an independent intelligence shop for the chip and AI-infrastructure industry, read by engineers, investors, and executives

[ikaijua]

Why people listen (method & influence)

Method (ground-up, starting from the technology):

• No extrapolating from earnings calls: works from the chips, systems, and models themselves; attends 40+ industry conferences a year and talks to people at every layer of the stack
• Builds its own data: satellite imagery, permit and property records, a teardown lab, and self-run benchmarks - independent of company disclosure
• Covers every layer from equipment/materials to models and connects the causal chain between layers

Influence (2025-2026 track record):

• Jensen Huang called out SemiAnalysis' open-source inference benchmark InferenceX on stage at GTC 2026 and walked through it for several minutes
• A critical report on AMD's MI300X led AMD CEO Lisa Su to arrange a 90-minute one-on-one within a day
• Microsoft CEO Nadella is a subscriber; CNBC's Jim Cramer called SemiAnalysis "the arbiter, they're like God"
• Widely cited by CSIS, The Information, and the tech press (usually with the caveat that its numbers are estimates)

[ikaijua]

Content & pricing

Free tier: many long-form articles (supply chain, architecture, market analysis) are public on the website and Substack.

Paid subscription (~$500/yr): unlocks every article and the deeper conclusions in the deep dives; groups get a 20% discount, corporate subscriptions a larger one; roughly one deep-dive every two weeks; payments via Stripe, no trial period.

The real revenue is institutional: per The Information, SemiAnalysis did ~$20M in 2025 and expects $100M+ in 2026 - subscriptions are a fraction of it. The money comes from selling deep research and data to startups, investment firms, and corporate teams (including Nvidia).

Where to start: read a few free articles to judge the style and quality, then consider subscribing; institutional data products are quoted separately.

[ikaijua]

Data models & benchmark products (institutional, custom-priced)

Beyond the newsletter, SemiAnalysis sells institutional data products (quoted to enterprise clients, prices not public):

• Accelerator & HBM Model: historical and forecast AI accelerator production, ASPs, and installed base by company and SKU - Nvidia (A100 through B200/R100), AMD, Google TPU, Huawei Ascend 910B/910C, ByteDance custom silicon, etc.; includes HBM capacity/layers/manufacturers and foundry/packaging demand
• Datacenter Industry Model: 5,000+ datacenter facilities worldwide, per-region capacity, delivery curves, and tenant identification
• China Datacenter Model (new Sep 2026): 1,000+ facilities across 60+ operators in China at building-level granularity - the model behind repo discussion #1172
• AI Cloud TCO Model: full ownership economics of buying accelerators and reselling GPU cloud (GB200 NVL72 all-in cost ~1.6-1.7x H100, etc.)
• ClusterMAX: tiered rating system for GPU clouds (Platinum/Gold/Silver/Bronze); v3.0 Sep 2026; CoreWeave the only Platinum
• InferenceX (formerly InferenceMAX): open-source inference benchmark run nightly on GB200, B200, H100, MI355X, etc., with a free public dashboard

[ikaijua]

Signature reports (famous and contested)

• GPT-4 architecture reveal (Jul 2023): judged GPT-4 a sparse MoE of ~1.8T parameters - later corroborated by multiple parties
• Google "We Have No Moat" (May 2023): republished the leaked internal Google memo, igniting the open-source AI debate
• 100,000 H100 Clusters (Jun 2024): a 100K-GPU training cluster needs >150MW and >$4B in server capex
• AI Datacenter Energy Dilemma (Mar 2024): global datacenter IT power 49GW in 2023 → ~96GW by 2026
• DeepSeek R1 deep dive (Jan 2025): among the first to systematically break down its training cost and architecture
• The Chinese AI Infrastructure Boom (Sep 2026): China 24GW+ datacenters (2nd globally), the East-Data-West-Compute mechanism, 100MW delivered in 12 months (full breakdown in repo discussion 中国 AI 基础设施热潮：SemiAnalysis 中国数据中心模型解读 #1172)

Caveat: these rest on anonymous sourcing and self-built models - numbers are estimates, not official figures; companies involved often deny them.

[ikaijua]

Who reads it & how to use it

• ML teams / engineers: architecture analysis, inference cost, and benchmarks (InferenceX and ClusterMAX ratings are free and public); TCO models for buy-vs-rent decisions
• Investors: subscriptions plus institutional models for GPU supply/demand, datacenter buildout, and semiconductor equity fundamentals
• Cloud providers & chip companies: buy institutional data and consulting (clients include Nvidia and major hyperscalers); also attend events SemiAnalysis co-hosts
• Casual readers: the free articles plus public benchmarks are plenty; institutional products are enterprise-priced

Suggested path: start with free content → decide whether the information density justifies paying → subscribe; never base a major decision on a single source.

[ikaijua]

Value and caveats (honest assessment)

Value:

• Information lead: supply-chain exclusives often beat mainstream media by weeks
• Data becomes the industry reference: ClusterMAX / InferenceX are cited by Nvidia's own developer blog; Jensen Huang called InferenceX out at GTC
• Real depth: the full chain from equipment/materials to models - something most outlets can't do

Caveats:

1. Numbers are estimates, not facts - built from anonymous sourcing, satellite imagery, and inference; companies often dispute them. CSIS, citing its DeepSeek report, explicitly noted "GPU counts are estimates"
2. Conflicts of interest - founder Dylan Patel is also an investor in ~20 AI startups (Thinking Machines Lab, Enfabrica, ...), and the firms it covers are often also paying clients; analyst/investor/publisher roles coexist, with the line drawn by Patel himself
3. Reputation risk signal - 2026 litigation with former employee Wei Zhou (mutual lawsuits involving Fluidstack coverage and Patel's investment ties)

Bottom line: treat it as one of the industry's most informed sources - but cross-check any key figure.

[ikaijua]

FAQ

Is $500/yr worth it? If the free articles click for you and you track chips/AI infrastructure long-term - yes. If you only want conclusions, free content plus secondary sources is usually enough.

Is there free content? Yes. The website and Substack host many free long-form pieces, close in quality to the paywalled tier.

How is it different from The Information or Tom's Hardware? More supply-chain depth plus proprietary data models; not a news chaser. Granularity (per-SKU shipments, per-facility capacity) is unmatched elsewhere.

How do I buy institutional products? Contact sales@semianalysis.com for a quote; prices are not public. The Accelerator model is an annual subscription with quarterly updates.

Are the numbers always accurate? No. It claims to "know more than anyone", but estimates lack independent replication - cross-verify before key decisions.

[ikaijua]

Sources

• Official site: https://semianalysis.com
• X (Twitter): @SemiAnalysis_
• Newsletter (Substack): https://newsletter.semianalysis.com
• About / team: https://semianalysis.com/about , https://newsletter.semianalysis.com/about
• FAQ (subscription & pricing): https://semianalysis.com/faq
• Related: The Chinese AI Infrastructure Boom - deep dive (based on its China Datacenter Model)