CoreWeave Customer Concentration: Two Buyers Paid 62% of Q2 Revenue

Two customers paid CoreWeave 62% of its revenue in the second quarter of 2026. That is the most diversified quarter the company has ever reported, and it is still a business where two signatures carry most of the income. CoreWeave customer concentration is easy to quote and hard to read, because the 10-Q names nobody. This post rebuilds who those customers are from every filing since the IPO, then sets them against the $35.6 billion of debt they ultimately service.

The short answer. CoreWeave's customer concentration is falling but still extreme. Its top two customers paid 62% of Q2 2026 revenue, down from a single customer paying 71% a year earlier. Microsoft was 67% of 2025 revenue. OpenAI has committed up to $22.4 billion and Meta about $35.2 billion, so three buyers anchor most of the backlog behind the debt.

CoreWeave customer concentration, quarter by quarter

CoreWeave reports concentration in one place each quarter: a small table in Note 2 of the 10-Q listing every customer above 10% of revenue. The customers are labelled A, B, C and D. Each filing warns that the labels "may represent different customers than those reported in a previous period."

That caveat matters, and I return to it below. Read the table as a sequence of snapshots, not as a tracked panel of the same companies. Even so, the direction is unambiguous across six filings.

FY2023, FY2025 and the quarterly rows come from CoreWeave's 10-K and 10-Q filings. The FY2024 split of 62% and about 15% is from S-1 coverage by The Register and Futuriom; the 10-K confirms only the combined 77%. Receivables shares are for Customer A as labelled in each filing.

Customer A is very probably Microsoft, but the filing will not say so

The 10-K names Microsoft as the top customer for 2025 at about 67%. Customer A carried 68% of receivables at 31 December 2025, which is consistent with the same company. Nothing in the 2026 quarterly filings names it.

My working assumption is that Customer A remains Microsoft through June 2026. The continuity of its dollar revenue, shown in the next section, supports that. It is still an inference, and you should treat every sentence below that relies on it as one.

Receivables tell the story a quarter faster than revenue

Receivables are bills sent but not yet paid. They lead revenue concentration because new contracts start billing before they dominate a full quarter. At 30 June 2026, Customers A and B each carried 32% of receivables, a dead heat.

At 31 March 2026, a third customer, labelled C, held 22% of receivables while staying below 10% of revenue. That is the fingerprint of a large contract ramping. By June a Customer C had crossed 10% of revenue, which fits the same pattern.

The largest CoreWeave customer stopped growing in dollars

Percentages hide the more useful fact. Multiply each filing's share by that period's revenue and you get the implied dollars paid by the largest customer. This is my arithmetic, not a CoreWeave disclosure.

Customer A paid about $707 million in Q1 2025 and about $861 million in Q2 2025. It then paid about $935 million in Q1 2026 and about $927 million in Q2 2026. Year on year, that is growth of roughly 8% in a quarter when CoreWeave's total revenue grew 112%.

For the full first half of 2026, Customer A's 40% share of $4,653 million implies about $1.86 billion. For 2025 as a whole, Microsoft's 67% of $5,131 million implies about $3.44 billion.

Diversification by addition, not by subtraction

The consensus reading of a falling top-customer share is that CoreWeave is becoming less dependent on Microsoft. I think that framing is half right. The dependence on Microsoft's dollars has not fallen at all. What has changed is that other buyers now stack on top.

Customer B is where the growth sits. Its 20% of Q1 2026 revenue implies about $416 million. Its 26% of Q2 implies about $670 million, a rise of roughly 61% in one quarter. Customers below 10% contributed about $727 million in Q1 and about $721 million in Q2, effectively flat.

So the Q2 revenue increase of about $497 million came mostly from Customer B and the newly visible Customer C. That is not a broad customer base filling in. It is two or three more very large contracts switching on, which is concentration of a different shape rather than its absence.

Microsoft, OpenAI and Meta: what each CoreWeave contract actually commits

CoreWeave names customers in two ways. It names Microsoft by revenue share in the annual report. It names OpenAI and Meta by contract ceilings in 8-K filings and press releases. Those are different measurements, and mixing them is the most common error in coverage of this company.

"Up to" figures are ceilings, not minimums, and some of each has already been recognised as revenue. The $4B OpenAI expansion is stated in CoreWeave's September 2025 press release; it was not located in the 10-K text reviewed for this post.

OpenAI: up to $22.4 billion from a counterparty the filing calls private

CoreWeave's own September 2025 release put the OpenAI relationship at "approximately $22.4 billion" across three tranches. The 10-K describes the first tranche as "up to approximately $11.9 billion through October 2030." It describes the third as "up to approximately $6.5 billion through May 31, 2031."

The same 10-K uses OpenAI as its example in the counterparty risk factor, introducing it as "a private company." That phrase is doing work. Microsoft and Meta publish audited cash flows every quarter. OpenAI does not, and how its own compute commitments compare with its revenue is examined in the breakdown of OpenAI's compute commitments against its revenue.

CoreWeave's filings also say they expect OpenAI "to be a significant customer in future periods." Customer B's jump to 26% in Q2 2026 is consistent with OpenAI capacity coming online. It is equally consistent with Meta. The filing does not let you choose, and I will not pretend it does.

Meta: two order forms worth about $35.2 billion

Meta's first order, signed 25 September 2025, committed "up to approximately $14.2 billion through December 14, 2031," with an option to expand through 2032. The second, signed 31 March 2026, committed "approximately $21 billion." That figure includes new capacity through 20 December 2032 and the exercise of an existing option running through 10 April 2032.

Both sit under a master services agreement dated 10 December 2023, so Meta has been a CoreWeave customer for nearly three years. On my reading the two orders are additive, because the 2026 filing describes new capacity plus an option that the 2025 order held open. The filing does not say so in words. If they overlap, the Meta figure is lower than $35.2 billion.

Meta's appetite for transferring data centre risk to other balance sheets is not unique to CoreWeave. The same instinct shows up in the Hyperion bond structure and who carries its risk.

Revenue concentration and backlog concentration point at different customers

Revenue tells you who pays today. Remaining performance obligations, or RPO, tell you who has signed for the future. RPO is contracted revenue not yet recognised, billed or unbilled. At CoreWeave it was $15.1 billion at the end of 2024, $60.7 billion at the end of 2025, $98.8 billion at 31 March 2026 and $103.7 billion at 30 June 2026.

CoreWeave has disclosed backlog concentration once, verbally. On the Q3 2025 call on 10 November 2025, the CFO said no single customer represented more than approximately 35% of revenue backlog, down from about 50% the quarter before, as reported by DatacenterDynamics. Several outlets put the figure at roughly 85% at the start of 2025.

Now run the disclosed ceilings against the June 2026 RPO. Meta's roughly $35.2 billion is about 34% of $103.7 billion. OpenAI's $22.4 billion is at most about 22%, and less once you strip out revenue already recognised since 2025. Microsoft's remaining commitment is not disclosed at all.

My read is that Meta is now CoreWeave's largest backlog customer while Microsoft remains its largest revenue customer. If that is right, the company's revenue line and its contracted future rest on different counterparties. Anyone modelling CoreWeave from the 10-Q revenue table alone is underweighting Meta, and that is the most underappreciated fact in the CoreWeave customer story.

A second caution sits inside the RPO itself. At 30 June 2026, 41% was expected to be recognised within 24 months, 39% in months 25 to 48, and the rest in months 49 to 78. The Meta orders run to late 2032, so a large share of them sits in the back of that schedule.

Setting CoreWeave's top customers against the $35.6 billion debt stack

Customer concentration is a sales metric at a software company. At CoreWeave it is a credit metric. The company finances GPUs with loans secured on specific customer contracts, a structure laid out in the analysis of CoreWeave's $35.6 billion debt stack. The customer is the collateral.

Three ratios fall out of the filings. Disclosed OpenAI and Meta ceilings of about $57.6 billion equal 56% of June 2026 RPO and about 1.6 times debt principal. That is a strong position on paper, and it is why lenders keep funding.

The near-term ratio is tighter. CoreWeave reported $6,235 million of recourse debt and $1,278 million of non-recourse debt due within 12 months, about $7.5 billion together. Its top two customers paid about $2.93 billion in the first half of 2026. At that pace, the debt due inside a year equals roughly 15 months of what its two largest buyers pay.

None of that signals distress. It does mean that one large customer slowing its ramp, or declining to renew, moves CoreWeave's refinancing math directly. The timing of those maturities is set out in the CoreWeave debt maturity ladder.

Counterparty quality matters more than counterparty share

If I were lending to CoreWeave, I would care less about the 62% than about who stands behind it. A contract with a company that publishes audited cash flows is a different asset from one with a private company still raising capital to meet its commitments. The 10-K draws that line itself by singling out OpenAI as private.

Lenders already price it. The March 2026 facility secured on contracted services with a large enterprise customer was rated investment grade. Later facilities backed by shorter contracts priced far wider. Concentration in a strong counterparty is cheap to finance. Concentration in a weaker one is not, however diversified the percentage looks. How coverage of interest by each revenue dollar has trended is tracked in the post on CoreWeave interest coverage per revenue dollar.

Where this argument is weakest

This post leans on inference more than most on this site. Here is where it could be wrong, and where the opposing view genuinely wins.

The anonymised labels can break every dollar trend

CoreWeave states that Customer A in one quarter may not be Customer A in another. If Microsoft's share fell below Customer B's at some point, the labels could have swapped, and the "flat dollars" finding would describe two different companies stitched together. I think that is unlikely, because the receivables data and the dollar continuity line up. Unlikely is not proven.

The rounding matters too. Each share is a whole percentage, so each implied dollar figure carries about $13 million of error either way. That does not change the direction of any finding here, but it rules out precision below that level.

Concentration is the business model, not a defect

The strongest counter-argument is that a neocloud is supposed to look like this. It builds capacity to order for a handful of buyers who need more compute than anyone else can deliver. The 10-K says plainly that it expects concentration "to continue in future years because of the long-term nature of contracts with those customers."

On that view, three investment-scale buyers on take-or-pay contracts averaging about five years is a feature. Committed contracts provided over 98% of 2025 revenue. A thousand small customers on monthly terms would be riskier collateral, not safer. I accept most of this. My disagreement is narrower: the risk sits at renewal and in counterparty quality, and the headline percentage measures neither.

Finally, the contract ceilings are ceilings. "Up to" language means the actual spend can land below the stated figure. That cuts against both the bull case and my Meta-versus-debt comparison.

What to read in CoreWeave's Q3 2026 filing

CoreWeave's Q3 2026 report has not been published as of 3 November 2026. When the 10-Q lands, four lines will tell you more about customer concentration than the headline revenue.

First, the Note 2 table. Watch whether Customer A's implied dollars break above the $930 million plateau, and whether Customer B overtakes it. Second, the receivables split, which will show the next large contract ramping a quarter before revenue does.

Third, any 8-K naming a new order form. CoreWeave added more than $25 billion of net new customer commitments early in Q3, per its August earnings release, and none was in the June backlog. Whoever signed it changes the backlog ranking above. Fourth, the RPO timing split, which shows how much of that backlog arrives before the next wave of debt matures. The same discipline applied to a chip supplier is in the read of Nvidia's customer concentration disclosures.

Frequently asked questions

How to use this before your next compute or credit decision

If you buy GPU capacity from CoreWeave or any neocloud, ask your account team one question this week. Which counterparty's contracts secure the facility that financed the capacity you sit on? A provider whose debt is backed by one hyperscaler's contracts behaves differently at renewal from one backed by a private lab's commitments.

If you hold or analyse the credit, rebuild the Customer A dollar line yourself when the Q3 10-Q lands. It takes four multiplications. If the largest customer's quarterly bill stays near $930 million while backlog keeps climbing, the company's future depends on its newer buyers, and Meta's renewal terms become the number to find.

Related reading