When NVIDIA sells through a cloud provider, the visible buyer may not be the final AI company using the hardware. That makes customer concentration difficult to read. A cloud company can look like one customer in a filing while its own demand is connected to several model companies, startups, and enterprises.

NVIDIA’s 2026 annual report discusses indirect customers and notes that one AI research and deployment company contributed a meaningful amount of revenue through customers purchasing cloud services. That disclosure is a reminder to look beyond the first name on the invoice.

Direct and indirect customers

A direct customer buys from NVIDIA. An indirect customer reaches NVIDIA through a distributor, cloud provider, system builder, or other channel. The commercial chain can be perfectly normal, especially when a cloud provider buys hardware and rents compute to many users.

The risk is analytical. If several cloud providers ultimately depend on the same model company, counting each cloud provider as independent demand may overstate diversification. The reverse can also happen: one cloud provider may serve thousands of smaller customers, making the visible customer look more concentrated than the end usage really is.

Why the disclosure matters

Concentration changes how a growth story behaves under stress. If one important customer delays a deployment, the supplier may see lower orders, slower collections, or more inventory. If end demand is broad and contracts are strong, the impact may be smaller.

The filing also discusses growth in compute and infrastructure spending. Those facts do not establish wrongdoing. They show that demand and capital intensity are linked and that investors should test whether the infrastructure chain can keep converting spending into usage.

Spend 2 hours drawing the chain from hardware maker to cloud provider to end workload. Mark where cash, credit support, and a purchase commitment appear. That map is usually clearer than a headline about “AI demand.”

What to compare in filings

Read revenue concentration, indirect-customer language, accounts receivable, inventory, capital commitments, and the timing of major product transitions. Then compare those disclosures with cloud-provider filings and large customer announcements.

Do not add every announced contract together. Ask whether the values overlap, whether the contract is delivered over several years, and whether the same end customer appears in multiple arrangements.

Also look at timing. A customer can announce a purchase before the supplier recognizes revenue, and a cloud provider can announce capacity before an end customer uses it. Write the dates beside each claim. That simple timeline often explains why two public statements seem inconsistent without requiring a dramatic theory.

Spend 2 hours checking whether the risk is measured by revenue, backlog, inventory, or receivables. Each denominator answers a different question, and mixing them can make concentration appear lower or higher than it really is.

Keep two ideas separate: customer concentration and customer quality. A large customer can be reliable, profitable, and diversified across its own users, or it can be highly dependent on one project. The filing rarely answers every part, so use the disclosure as a starting point for questions rather than a complete picture.

This is especially important during a hardware transition. Customers may buy early to reserve supply, then change the mix of products as a new generation arrives. A strong analysis records the product, delivery date, and end workload instead of treating all chip revenue as the same kind of demand.

Use the same timeline when comparing quarterly reports, because a single quarter can reflect shipments rather than lasting consumption.

A fair conclusion

Indirect customer risk is not an accusation. It is a reason to avoid simplistic conclusions such as “NVIDIA has thousands of customers” or “one AI lab explains everything.” The truth may be a network of customers with different levels of independence and different ability to pay.

Sources and limits

Use NVIDIA’s 2026 annual-report filing, its latest 10-Q, and CoreWeave’s 2025 filing. These are reported disclosures, not predictions or proof of misconduct. Recheck later filings for updated concentration.