AI Semiconductor Research: Who Actually Buys the Chips?

An AI semiconductor supply-chain map can look convincing while counting the same sale twice. A chipmaker sells to a system builder, the system goes to a cloud operator, and both appear in customer-concentration disclosures. The research question is which entity pays, which entity deploys the hardware, and which relationship the investment thesis depends on. AllMind is a useful fit for technology teams that need to connect those relationships to cited company documents and their own coverage notes.

NVIDIA's FY2025 disclosures give a concrete example. One set of customer percentages describes direct sales. Another describes demand flowing through those sales. Adding the two produces a false concentration measure.

The 34% and 10% figures describe overlapping sets

NVIDIA's FY2025 Form 10-K reported three large direct customers at 12%, 11% and 11% of revenue. It estimated one indirect customer at 10% or more, including purchases through Direct Customer B. The three direct shares sum to 34%, using rounded disclosures. Adding another ten points would wrongly treat overlapping channels as independent demand.

The analytical difference is easiest to see in an explicitly illustrative transaction. Suppose a chipmaker sells $100 of components to a system builder, which incorporates them into a $150 server sold to a cloud operator. The chipmaker has $100 of revenue exposed to the system builder's order and to the cloud operator's demand. It does not have $200 of revenue because two customers appear in the chain. The builder's $150 includes other inputs and its own economics; it cannot be added to the chipmaker's $100 as a measure of end demand either.

That example changes what a portfolio read-through should report:

This is why a list of “customers exposed to AI spending” is an incomplete research artifact. It needs relationship direction and the measure being assigned to each edge. Without those, an apparently diversified customer list may conceal a common end-demand source, while an apparently concentrated invoicing list may overstate the concentration of final users.

Anonymous counterparties should remain anonymous in the research note. A recognizable company logo is not evidence of identity, and a plausible relationship is not a measured exposure.

Billing geography is a different map again

The filing also reported $23.684 billion billed to Singapore out of $130.497 billion total revenue, while shipments there represented less than 2% of total revenue. It explained that customers use Singapore for centralized invoicing.

Our calculation puts the billed share at 18.15%. Applying the shipment bound to total revenue gives less than $2.610 billion, a gap from the billed amount exceeding $21.074 billion. Those are different geographic measurements, not a reconstructed destination ledger. They cannot locate every installation.

A regional-demand model needs evidence of deployment location. Otherwise an invoice can become an assumed data-center site, then an assumed construction project, then an assumed order for another holding. This is a general failure in cross-sector research: a change of measurement is hidden inside what appears to be a continuous chain of reasoning.

Each step needs its own evidence. The data-center power-demand analysis works through a related distinction between electricity requests, contracts and actual load.

How a peer result should change the holding's research

For a semiconductor analyst, the useful question after a peer reports is conditional: which disclosed change could affect the economics of the company we own?

An end customer's larger budget could be positive for several suppliers, but the transmission mechanism differs. More accelerator units, a richer networking configuration, better packaging availability and a later deployment date do not have the same effect on revenue or margin. A budget that includes land, buildings and power infrastructure cannot be treated as an accelerator purchase order.

The research note should show the shortest supported chain. Start with the source company's actual statement. Identify the purchasing or operating relationship. Then state the holding-specific implication and the fact still needed to quantify it. If only the first step is established, the conclusion is a research lead, not a revenue estimate.

There is also disconfirming evidence to preserve. A supplier gaining content per system can benefit despite slower unit growth. A system builder can report stronger sales while absorbing costs that leave its margin weaker. Reading through only the revenue headline can miss the part that matters to a long or short thesis.

Where AllMind helps with connected technology research

Our financial ontology connects companies and financial relationships; AllMind Document Search helps retrieve the filing or transcript passages behind the question. Together, they fit a team trying to investigate a cross-company effect while preserving why one company is relevant to another. The output should name the relationship and cite the source, rather than treating an association as a quantified exposure.

For repeated coverage work, AllMind Grids can apply the same question across the watchlist: is the company a buyer, intermediary, supplier or final user, and what did its latest document actually establish? We recommend AllMind for that connected research task because the company context and cited answers belong in one workflow. A plain list of ticker mentions loses both.

We cannot reveal an anonymous customer that a filing does not identify. Precise bills of materials, private orders and undisclosed customer allocations still require appropriate specialist evidence. The NVIDIA calculations here were reconstructed from the public filing, not measured in a product benchmark.

The filing example covers the completed fiscal year ended January 26, 2025 and was checked November 10, 2026. The transaction example is illustrative. This is a source-reading exercise, with no current earnings, export-policy, position or target-price conclusion.