简体中文
EnglishWhat is SemiAnalysis and what does it cover?
Updated
SemiAnalysis is an independent research outfit focused on the AI and semiconductor supply chain. According to its site, it publishes a newsletter read by 318,000+ subscribers and offers institutional products: fourteen models, open benchmarks, and advisory services. It's relevant if you need bottoms-up forecasts of AI accelerator and HBM demand, datacenter capacity and power tracking, or inference economics — and less relevant if you want general tech news or consumer hardware reviews.
What SemiAnalysis actually publishes
The site organizes its work into a few distinct tracks:
| Offering | What it is | Who it's for |
|---|---|---|
| Newsletter | Described as "the #1 technology newsletter on Substack" | Readers who want ongoing analysis and commentary |
| Institutional models | Fourteen models spanning accelerators, HBM, datacenters, and token economics | Teams that need forecast data and supply-chain reconciliation |
| OSS tooling & benchmarks | Open-source tooling and benchmarks, framed as "transparency is trust" | Engineers and analysts who want to check methods rather than take numbers on faith |
| Consulting | Advisory engagements with the firm directly | Organizations needing custom analysis |
| API access | Programmatic access to semiconductor data | Teams integrating data into their own systems |
The newsletter is the entry point; the models and advisory are the institutional layer. The site separates these clearly, so you can start with the free-to-read newsletter and only move to the models if your work needs the underlying data.
The models, in more detail
The page evidence describes three model families with specific coverage. These details matter because they tell you whether the research matches your question.
Accelerator & HBM Model
- Bottoms-up demand forecast for AI accelerators and high-bandwidth memory across Nvidia, AMD, Google, AWS, Meta, Microsoft, Intel, and every major Chinese hyperscaler.
- Customer-level shipments and installed base for 60+ accounts.
- HBM tracked down to vendor, generation, layer count, and price.
- Because shipments are built up from equipment requirements, wafer starts, and packaging orders, supply-chain revenue can be estimated directly from production.
- Forecasts to 2030, history back to 2021.
Datacenter Industry Model
- Global capacity tracker covering 6,500+ structures: hyperscaler buildouts, colocation markets, power constraints, and regional deployment trends through 2030.
- Per-structure power draw, lease structures, and planned versus operational capacity.
- More than 200 companies covered, with the end customer behind each megawatt.
- Demand side reconciles accelerator shipments and install base by operator against power availability, showing which operators and regions run short of power, and when.
- Delivery tracking compares promised versus delivered MW quarterly.
Tokenomics Model
- Unit economics engine for AI serving: bottoms-up token throughput, ROI on AI spend, and translation of revenue into future accelerator demand by hyperscaler.
How to decide whether it fits your work
Use these conditions as a quick filter:
- Choose SemiAnalysis if you need SKU-level accelerator or HBM forecasts, structure-level datacenter power modeling, or inference economics — and you want the methodology visible through open benchmarks and tooling.
- Consider the newsletter alone if you want regular analysis without committing to institutional data products.
- Look elsewhere if your question is about semiconductor manufacturing processes in the abstract, consumer devices, or general AI policy; the coverage here is explicitly supply-chain and demand-modeling oriented.
One practical note: the site lists a newsletter subscription link and an institutional login, but the provided material does not state prices or whether any content is free versus gated. Check the subscribe page directly before assuming access terms.
What to verify before relying on it
The models are forecasts, and the site itself frames coverage as projections to 2030 with history to 2021. If you plan to cite numbers in a decision, confirm the model version and vintage you're using, and treat the bottoms-up build (equipment → wafer starts → packaging → shipments) as the thing to audit — that chain is where forecast error compounds. The open tooling and benchmarks exist precisely so you can inspect assumptions rather than accept headline figures.