Who Gets Paid When NVIDIA Ships a GPU? We Mapped It for the 100 Biggest Stocks

The question a stock screener cannot answer

Search "AI" in a company database and you get hundreds of names across a dozen sectors. What you will not get is the specific chain of companies that bills someone when a data center actually goes up: the foundry that etches the die, the firm that licenses the CPU instruction set, the company that builds the rack, the one that keeps it from overheating, and the neocloud that rents it out by the hour.

That is a different question from keyword search, and it needs a different kind of data. So we built it.

The new ecosystem pages map the suppliers, customers, partners, competitors and channel relationships around the 99 largest US companies by market cap. That is 1,420 mapped relationships. You can browse them all at pro.stockalarm.io/ecosystem, or jump straight to NVIDIA's map.

Every relationship carries a receipt

The easy version of this product is an LLM listing plausible suppliers. That version is worthless, and worse than worthless when it is confidently wrong.

So the rule is: no relationship renders on the page without a quote behind it.

A relationship counts as confirmed only when it can be tied to a specific sentence in a primary document. Not a summary, not a paraphrase, not a model's recollection. The actual sentence, with the speaker and the date, stored alongside the relationship and shown to you on the page.

Right now that is 1,107 of 1,420 relationships confirmed, or 78%. Seventeen companies are at 100% coverage, including Apple at 28 of 28 and Cisco at 18 of 18. The remaining 313 are counted on the page as unconfirmed, but the names are not shown. If we cannot prove it, you do not see it.

Here is where that evidence comes from:

Why you have to read both sides

The single most important thing we learned building this: a big company rarely names its own suppliers.

NVIDIA does not spend its earnings calls listing the firms that keep its GPUs cool. Vertiv does. Vertiv brought up NVIDIA 8 times across its recent calls, because that relationship is material to Vertiv in a way it simply is not to NVIDIA.

Mine only the hub company's disclosures and you find none of the power, cooling and optics half of its ecosystem. You end up with a map of the famous relationships and none of the load-bearing ones.

So the map reads three layers, and it genuinely needs all three:

1. The counterparty's transcript, which catches everything a smaller supplier says about its largest customer. This is the single richest layer at 407 confirmations.
2. The hub company's transcript, which catches what the large company chooses to acknowledge, 144 relationships.
3. SEC filings in both directions, 556 relationships and the layer that catches the certain ones.

That third layer matters more than it sounds. TSMC never names a customer. But AMD names Taiwan Semiconductor outright in its own 10-Q, because concentration in a single foundry is a risk factor AMD is obligated to disclose. Risk-factor language is where supply chains get admitted.

The foreign filing trap

One detail worth calling out, because it silently breaks this kind of dataset.

Foreign private issuers do not file a 10-K. Taiwan Semiconductor, ASML, Arm, STMicroelectronics and Nokia file a 20-F. Canadian issuers such as Magna file a 40-F.

Search only 10-K and 10-Q filings and every one of those relationships becomes structurally unconfirmable. Not wrong, not missing, just permanently unprovable, which looks identical to "no relationship exists." Including those forms confirmed 23 relationships that were otherwise invisible, including Taiwan Semiconductor's supply into Microsoft, Alphabet, Amazon, Cisco, Qualcomm and Analog Devices.

Mentions is a dependency ranking

Each relationship carries a mentions count: how many times that counterparty discussed the hub company across its last four earnings calls. It is not a revenue estimate and we do not present it as one. It is a rough measure of how much the relationship is on that company's mind.

For NVIDIA, sorted by mentions:

A supplier that names its largest customer 59 times in a year is telling you something about concentration risk that no ratio on its balance sheet will.

What the shape of the data says

Across all 99 companies, the 1,420 relationships break down as 484 supplier links, 396 competitor links, 233 customer links, 211 partner links and 96 channel links.

Competitors outnumbering customers is not an accident of collection. Companies talk about rivals constantly and about customers carefully, because naming a customer invites questions about concentration and pricing. The disclosure incentives are asymmetric, and the data reflects that.

Two files, on purpose

For anyone curious about the mechanics: the map is two separate files.

One holds hand-authored intent, which company supplies whom, in what role, with a plain-language note. The other holds that same map plus the machine-gathered evidence. Re-running verification can never overwrite a human judgment call, and a diff shows editorial changes separately from new proof.

That separation exists because the failure mode we most wanted to avoid is a dataset that quietly rewrites itself and leaves nobody able to say which parts a person actually decided.

Where to start

Open NVIDIA's ecosystem map and sort by mentions. Then try a company whose supply chain you think you already know, and look at the competitor column, which is usually the one that surprises people.

If a relationship on the page looks wrong, the quote that justifies it is right there. That is the entire point.