How SemiAnalysis Approaches AI and Semiconductor Supply Chain Research

SemiAnalysis builds its research from the bottom up: it starts with physical production inputs—equipment requirements, wafer starts, and packaging orders—and works forward to SKU-level shipments, installed base, and revenue. That approach is what distinguishes its models from top-down market sizing, and it is the lens to keep in mind when deciding whether its research fits your needs. The method is best suited to readers who need to connect upstream capacity to downstream deployment—supply chain analysts, infrastructure planners, and investors tracing how fab and packaging constraints translate into available compute.

The Core Mechanism: Build Shipments from Production Inputs

The central design choice is that demand forecasts are not estimated from spending intentions or survey data. Instead, shipments are constructed from what has to be produced to make them possible:

• Equipment requirements feed into capacity estimates.
• Wafer starts translate capacity into die output.
• Packaging orders capture the advanced packaging step that constrains many AI accelerators.

Because each layer is modeled explicitly, revenue for companies across the supply chain can be estimated directly from production rather than inferred from reported sales. The stated coverage runs forecasts to 2030 with history back to 2021.

A concrete example of why this matters: if packaging capacity for high-bandwidth memory is the binding constraint in a given quarter, a top-down demand model will overstate accelerator shipments. A bottoms-up model that tracks packaging orders separately will show the bottleneck before it appears in shipment data—and will also show which upstream suppliers benefit.

Linking Upstream Capacity to Downstream Deployment

The research connects production to where compute actually lands. The Accelerator & HBM Model covers Nvidia, AMD, Google, AWS, Meta, Microsoft, Intel, and every major Chinese hyperscaler, with customer-level shipments and installed base for 60+ accounts. HBM is tracked at the same granularity as compute—down to vendor, generation, layer count, and price.

On the deployment side, the Datacenter Industry Model tracks 6,500+ structures, including hyperscaler buildouts and colocation markets, with per-structure power draw, lease structures, and planned versus operational capacity. More than two hundred companies are covered, with the end customer identified behind each megawatt. A supply-demand reconciliation shows which operators and regions run short of power, and when.

The Tokenomics Model adds a third layer: unit economics of AI serving, including bottoms-up token throughput, ROI on AI spend, and translation of revenue into future accelerator demand.

Independence and Transparency

SemiAnalysis describes itself as independent research on the AI and semiconductor supply chain, read by 318,000+ subscribers, with fourteen institutional models, open benchmarks, and advisory. The transparency claim is backed by open-source tooling and benchmarks published alongside the models—useful if you want to inspect methodology rather than accept conclusions. Note that the site presents these as its own positioning; treat the independence claim as a stated principle and evaluate the models on their published assumptions.

How to Access the Research

Access splits into three tracks, and the right one depends on your use case:

• Newsletter — the Substack publication, positioned as the entry point for individual readers. A subscribe link is available at newsletter.semianalysis.com/subscribe. Pricing and login terms are not specified in the available materials, so check the subscribe page directly before assuming cost or access level.
• Institutional models — the fourteen models above, accessed through an institutional login on the site. Each model page offers "View model" and "Talk to Sales," which indicates a sales-qualified process rather than self-serve signup.
• API access and consulting — programmatic access to semiconductor data, and direct engagement with the research team.

Choosing Between Tracks

• If you need to follow industry developments and occasional analysis, start with the newsletter.
• If you need SKU-level forecasts, installed base, or datacenter capacity data for planning or investment work, the institutional models are the relevant product—plan for a sales conversation.
• If you need to pipe data into your own systems, ask about API access during that conversation.

Common Pitfalls When Using This Research

• Don't read bottoms-up forecasts as precise point estimates. They are conditional on equipment, wafer start, and packaging assumptions; when a constraint shifts, the forecast shifts. Check which input is binding before citing a number.
• Match model granularity to your question. The datacenter model answers power and capacity questions; the accelerator model answers shipment questions. Using one for the other's purpose will mislead.
• Verify access terms before budgeting. The available site information confirms a subscribe link and an institutional login but does not state prices or whether any content is free.

For a quick orientation, the site's own framing is the clearest summary: the definitive source on the world's most important industry, built on models that start from production and end at deployment.