The AI Cloud Boom Has a Few Very Big Customers

The AI infrastructure boom looks like the creation of an entirely new industry. Chip manufacturers sell processors. Cloud operators rent GPU clusters. Data-center developers build facilities. Electricity providers connect enormous new loads. Banks and investors help finance the expansion.

There are many companies involved—and from a distance it can look like a broad, diversified market.

But what if several layers of that market depend on the same handful of buyers?

That is the hidden concentration problem disclosed in the public filings of companies building the infrastructure for artificial intelligence. The demand can be real, the contracts enormous, and the customer base surprisingly narrow, all at the same time.

One customer can account for most of the sales

Nscale's September 18, 2026 U.S. IPO filing provides a revealing example.

In the six months ended June 30, the company's largest customer represented 52% of reported revenue. In 2025, the same disclosure said its largest customer accounted for 73%. These are concentration metrics from Nscale's own prospectus, not estimates inferred from a press headline.

The filing also described statements of work with Microsoft providing for payments of up to approximately $43.8 billion through 2033 and separate Anthropic services agreements with potential payments of up to approximately $44.6 billion.

Those prospective commitments are large, but they are not the same as the amount already earned. Delivery and service-availability requirements matter. Optional extensions, future capacity and the precise terms of each contract affect how much activity ultimately becomes recognized revenue.

The underlying business has demand. What it does not necessarily have is a wide distribution of independent customers.

For an infrastructure operator, both facts can be true.

CoreWeave shows the pattern is bigger than one IPO

CoreWeave's filings disclose a similarly concentrated business.

Its second-quarter 2026 report said that its three largest customers accounted for approximately 36%, 26% and 10% of quarterly revenue. That means the top three together represented around 72% of sales in that period.

The 2025 annual report said Microsoft had represented about 67% of CoreWeave revenue for that year. These percentages are tied to different reporting periods and should not be compared as if nothing changed between them.

What matters is the general pattern: a relatively small number of companies have the ability, demand and resources to rent enormous amounts of frontier AI computing capacity.

For the cloud provider, a big customer contract can make expansion possible. It can also give that customer immense importance to the business's future.

The risk travels down the stack

The next step is where the story becomes less obvious.

An AI laboratory or hyperscaler buys GPU-cloud capacity from a specialist provider. The specialist provider, in turn, leases or contracts space and power from a data-center landlord. That landlord invests in facilities, electrical infrastructure and cooling equipment to deliver on the arrangement.

At each layer, a different company may record revenue, debt, or a long-term commitment.

But the chain can still depend on demand that began with one large enterprise decision.

Core Scientific's June 2026 filing makes the dependency concrete. It disclosed that CoreWeave accounted for approximately 77% of Core Scientific's total first-half revenue, while its colocation business was entirely derived from CoreWeave at that point.

Core Scientific also said that colocation represented 77% of its total first-half revenue. These are two related disclosures, not two additional independent customer categories.

Imagine demand for a particular generation of AI servers changes. A cloud buyer may slow its planned expansion or adjust where it purchases new computing capacity. The GPU-cloud operator could then face a slower capacity ramp. Its landlord might still have equipment, power, financing and construction commitments tied to the original plan.

That chain reaction is a business risk, not a prediction that the companies will default or that the AI boom is imaginary.

Long-term contracts are real protection

The concentration argument has an important counterweight.

Many AI infrastructure contracts are designed to provide greater revenue visibility than selling compute capacity one hour at a time. They can include committed minimums, prepayments, capacity reservations, long tenors and negotiated consequences if conditions are not met.

CoreWeave's annual filing, for example, discusses committed arrangements and customer prepayments that help finance its infrastructure needs.

These terms matter. They can reduce uncertainty around whether there will be a buyer for an expensive cluster once it has been installed.

But they do not eliminate all risks. Delivery can be delayed. A customer can change its architecture or mix of suppliers at renewal. Customer credit quality may change. The cost of equipment and debt can shift. And newly purchased accelerators can become less economically attractive if software or hardware efficiency advances much faster than the contract's expected lifespan.

A contract makes the economics more predictable within specified terms. It does not suspend the technology cycle.

Why the same names keep appearing

The number of potential users for AI products is very large.

The number of organizations capable of committing tens of billions of dollars to advanced computing infrastructure is much smaller.

Large hyperscalers and frontier AI companies have an unusual combination of financing access, engineering demand, product distribution and the ability to absorb massive up-front costs.

That is why their names appear throughout different cloud providers' sales presentations, contracts and filings.

If one cloud company wins a huge deal with a hyperscaler, and another provider wins a huge deal with a frontier model company, the market can look like it has gained two wholly separate demand engines. In reality, the final demand may still depend on overlapping end-user adoption, a small set of technology platforms, and a relatively concentrated group of companies choosing where to place workloads.

The supply chain becomes diversified faster than the buyer base.

The questions a backlog headline cannot answer

A gigantic contract announcement can be an important indicator of future demand. It is less informative when stripped of its surrounding financial and operational details.

Before treating a large AI-cloud contract as a guaranteed source of future sales, ask:

Who is the ultimate buyer? Does that customer also appear behind other infrastructure companies' headline contracts?

What has actually been delivered? A multi-year commitment for proposed capacity is not the same as revenue already recognized.

Who finances the build-out? Which party owns or leases the servers, who has made construction commitments, and who pays for unused capacity?

What protects both sides? What do minimum commitments, prepayments, service requirements and termination clauses actually cover?

How quickly can the technology change? An expensive facility and an accelerator generation may have very different economic lifespans.

The AI boom does not need to be fraudulent to carry concentration risk. It can be one of the largest authentic investment cycles in history and still have a customer structure that makes individual decisions unusually consequential.

The demand is real. The buyer base is narrow. That is the distinction worth following.