Who depends on whom: reading customer concentration in SEC filings

Almost everything in this catalog is a lead. A placement tagged reported came from trade press; one tagged inferred came from a taxonomy seed and has no citation behind it at all. We say so on every row, because the alternative — presenting a plausible supply relationship as a confirmed one — is the failure mode that makes most supply-chain databases useless.

There is one source that does not have this problem. US-listed companies have to disclose in their 10-K, 10-Q or 20-F when a single customer accounts for a material share of revenue or accounts receivable — usually 10% or more. It is a compelled statement, made under liability, about who the company depends on. That is as close to ground truth as this industry offers, and the disclosed relationships table now holds 197 of them, from 101 companies.

It is also far less useful than it sounds, in three specific ways worth understanding before you lean on it.

One: almost nobody names the customer

The disclosure requirement is about the concentration, not the counterparty. A filer satisfies it completely by writing “one customer accounted for 17% of total revenue” and stopping there. Of the 197 disclosures in the table:

• 2 name the customer outright.
• 16 use letters — Customer A, Customer B — anonymised in the filing itself, not by us.
• 57 say only “one customer” or “our largest customer”.
• 122 are a combined group: “our five largest customers accounted for 28.8% of revenue”.

The two that name names are worth reading in full. STMicroelectronics’ 2025 Form 20-F states that its largest customer, Apple, accounted for 17.7% of total revenues. CoreWeave’s 10-K for the year ended December 2025 states that approximately 67% of revenue came from its top customer, Microsoft. Those are disclosed facts with a filing URL attached, and they are the only two in this dataset that are. Everything else tells you a supplier is concentrated without telling you on whom.

Two: “five largest customers” is not concentration

The single most common way to read this data wrong is to treat a group figure as a single-customer figure. “Our five largest customers comprised 28.8% of revenue” is a much weaker statement than “one customer accounted for 28.8% of revenue” — the first is consistent with a healthy book of business, the second is an existential dependency. They are the same sentence shape and the same percentage.

The table now separates them into single and combined scope and never mixes the two, because the group rows outnumber the single ones — 122 to 75. A summary that averages across both is measuring nothing.

Three: the extraction is the hard part, not the reading

A filing contains a great many sentences that look exactly like a concentration disclosure and are not one. Pulling candidate sentences by pattern produced 573 of them; 197 survived review. The other 376 failed in four distinct ways:

• 162 were negations. “No single customer accounted for more than 10% of net sales in fiscal 2025.” Same words, same threshold, opposite meaning. Published as a concentration disclosure, this says the reverse of what the company filed — the single worst error this dataset could make, and it was 28% of the first pass.
• 43 were measuring something else. “Revenue transferred to customers over time accounted for approximately 95% of revenues” is revenue-recognition timing under ASC 606. “Sales to customers in Asia were 92% of total revenue” is geography. “Approximately 80% of revenue comes from serving Virginia jurisdictional customers” is a class of ratepayer. None is a customer.
• 132 had no concentration construction at all — a percentage and the word “customer” in the same sentence, joined by nothing.
• 39 were table scrapes. “Customer B * 21 % Customer C 11 % *” — the numbers are real, but the pairing of label to value does not survive the scrape, so the row cannot be published truthfully even though every character in it came from a filing.

The extractor now requires positive evidence of a concentration construction and clears all four gates before a row is published. Precision over recall: a missing row costs a lead, a wrong row costs the thing the site is for.

What it is good for

Used carefully, the table answers a question the rest of the catalog cannot. Everywhere else, this site can tell you that a company makes a part. Here, a company tells you — on the record, with a filing URL — how much of its business rides on one buyer. Where the buyer is named, that is a disclosed edge in the supply graph. Where it is not, the concentration number is still a real measure of fragility, and a supplier at 60%-plus of revenue from one unnamed customer is a different kind of risk from one at 12%.

Read it alongside the bottlenecks page, which covers the other direction: parts where supply, not demand, is the binding constraint. And the caveats that apply everywhere on this site apply here too — see sources & method.