From Ritu Raj | Product & Market Analysis
CoreWeave Customer Concentration: Top Client Fell From 71% to 36% in a Year
On this page
CoreWeave's largest customer supplied 71% of its revenue in the second quarter of 2025 and 36% in the second quarter of 2026. That is the cleanest measure of CoreWeave customer concentration in the filings, and it reads like a dramatic de-risking. Convert the percentages into dollars and a different story appears: the top customer's quarterly bill barely moved, while two newer buyers grew around it.
Key takeaways
- CoreWeave's top customer share roughly halved in a year. Customer A fell from 71% of Q2 2025 revenue to 36% of Q2 2026 revenue, while a second customer rose to 26%, according to the company's own 10-Q filings.
- The dilution came from growth around the top customer, not from that customer shrinking. At 36% of $2,575 million, Customer A billed about $927 million in Q2 2026, against about $861 million at 71% a year earlier.
- Meta's two disclosed commitments are almost the size of CoreWeave's entire debt stack. Meta committed up to $14.2 billion in September 2025 and a further $21 billion in March 2026, a combined $35.2 billion against $35.6 billion of debt principal.
- Two customers still carry the interest bill. Strip the two largest customers out of Q2 2026 and $978 million of revenue remains, against $640 million of net interest expense in the same quarter.
The short answer. CoreWeave's revenue still depends heavily on a few buyers. Microsoft was 67% of 2025 revenue. By Q2 2026 the largest customer was 36% and the top three were 72%. OpenAI and Meta have disclosed commitments worth up to $57.6 billion, so concentration has spread across three names rather than disappeared.
What CoreWeave's filings actually disclose about customers
CoreWeave reports customer concentration in two places, and they follow different rules. Reading both is the only way to build a usable picture.
The quarterly 10-Q carries a table in Note 2 headed "Significant Customers". It lists every customer above 10% of revenue as Customer A, B, C or D, with no names attached. It also gives each lettered customer's share of accounts receivable, which is money billed but not yet collected.
The annual 10-K does name names, but only in the risk factors. For 2025 it states plainly that the company recognised "approximately 67% of our revenue from our top customer, Microsoft". It adds that no other customer reached 10% that year, and that OpenAI and Meta are expected to be significant customers in future periods.
There is one sentence in every quarterly note that a careful reader cannot skip. The filings say the customer references A through D "may represent different customers than those reported in a previous period". In plain terms, Customer A in March 2026 is not promised to be the same company as Customer A in June 2026.
That caveat shapes everything below. Where this post attaches a name to a letter, it says so and says why. Where the filing does not support a name, the letter stays a letter.
CoreWeave customer concentration, filing by filing
Laid end to end, the disclosures show a business that became more dependent on one customer through 2025 and then broke that dependence in the first half of 2026. The table below pulls every period we could verify against a primary filing.
| Period | Largest customer | Other customers at 10% or more | Who is named |
|---|---|---|---|
| FY2023 | Top three combined 73% | Not split in the 10-K text | None in the 10-K passage |
| FY2024 | 62% | Top two combined 77% | Microsoft, per S-1 coverage |
| Q1 2025 | 72% | None | Lettered only |
| Q2 2025 | 71% | None | Lettered only |
| FY2025 | 67% | None | Microsoft, in the 10-K |
| Q1 2026 | 45% | Customer B 20% | Lettered only |
| Q2 2026 | 36% | Customer B 26%, Customer C 10% | Lettered only |
Quarterly figures come from Note 2 of the Q1 2026 and Q2 2026 10-Q filings, which carry the 2025 quarters as comparatives. Annual figures come from the FY2025 10-K. The FY2024 Microsoft figure is from coverage of the March 2025 S-1, not a document we opened directly.
The 2025 peak is the period most coverage still quotes. A customer at 71% or 72% of revenue is close to a single-tenant business, and the 10-K confirms that customer was Microsoft for the year as a whole.
The 2026 break is real and it is fast. Within two quarters, the largest share fell by 35 percentage points and two new customers crossed the 10% line. I read that as the most important change in CoreWeave's risk profile since the IPO, more important than any single financing.
The grey block is the detail the headline misses. Customers below the 10% threshold held 28% of revenue in Q1 2025 and 28% again in Q2 2026. The long tail did not grow as a share. The concentration simply spread across three large names instead of one.
Converting CoreWeave's concentration percentages into dollars
Percentages hide direction. A customer can fall from 71% to 36% of revenue while paying you more money, if the total grows faster than its spend. Multiplying each share by reported revenue shows which happened here.
| Period | Total revenue | Customer A share | Customer A, implied | Customer B, implied |
|---|---|---|---|---|
| Q1 2025 | $982 | 72% | About $707 | Below 10% |
| Q2 2025 | $1,212 | 71% | About $861 | Below 10% |
| FY2025 | $5,131 | 67% (Microsoft) | About $3,438 | Below 10% |
| Q1 2026 | $2,078 | 45% | About $935 | About $416 |
| Q2 2026 | $2,575 | 36% | About $927 | About $670 |
Our arithmetic: reported revenue multiplied by the disclosed percentage. Shares are rounded to whole numbers in the filings, so each implied figure carries an error of roughly 0.5% of quarterly revenue, about $13 million in Q2 2026.
The top customer stood still while the business doubled
If Customer A is the same company throughout, its quarterly spend rose from about $861 million in Q2 2025 to about $935 million in Q1 2026. It then held at about $927 million. Total revenue grew 112% over the same year.
That pattern fits Microsoft's 67% share for full-year 2025, which works out to roughly $3.4 billion, or about $860 million a quarter on average. The fit is suggestive, not proof, because the filing reserves the right to reshuffle letters. Still, a flat top customer is exactly what the 10-K hinted at when it named OpenAI and Meta, not Microsoft, as the customers expected to matter in future periods.
Customer B is now the growth engine
Customer B went from below 10% of revenue in Q2 2025 to about $416 million in Q1 2026 and about $670 million in Q2 2026. That is a 61% rise in a single quarter. Customer C appeared at 10%, about $258 million, for the first time in Q2 2026.
Receivables add a useful wrinkle. At 30 June 2026, Customers A and B each accounted for 32% of net accounts receivable, even though B supplied only 26% of revenue. A customer whose share of unpaid bills runs ahead of its share of revenue is either ramping fast, paying on longer terms, or both. Neither reading is alarming on its own, and both are worth a second look in the next filing.
Microsoft, OpenAI and Meta: what each has signed
The lettered tables tell you how concentrated revenue is. The named disclosures tell you who the concentration is with. They come from different documents and they measure different things, so the comparison needs care.
Microsoft: the largest payer with no disclosed contract value
Microsoft is the only customer CoreWeave has named with a revenue percentage. It was 67% of 2025 revenue, and coverage of the S-1 put it at 62% of 2024 revenue. The 10-K risk factors also list Microsoft among the counterparties whose non-payment or non-performance would hurt the business.
What the filings do not give is a Microsoft contract value. There is no 8-K for a Microsoft order form comparable to the OpenAI and Meta filings. That gap is why this post cannot say how much of the $103.7 billion backlog Microsoft holds, and I would treat any precise figure circulating online as unsourced.
OpenAI: up to $22.4 billion across three orders
CoreWeave disclosed an initial OpenAI agreement worth up to $11.9 billion in March 2025. An expansion worth up to $4 billion followed in May 2025, and a third order worth up to $6.5 billion in September 2025. The company's 25 September 2025 release put the combined value at "up to approximately $22.4 billion".
The September 8-K gives the third order's term as running through 31 May 2031. The phrase "up to" appears in every OpenAI disclosure, and it matters: these are ceilings on what OpenAI has committed to pay, not floors. How OpenAI funds commitments of this size across every supplier is examined in the piece on OpenAI's compute commitments against its revenue.
Meta: two orders worth up to $35.2 billion
Meta signed an order form in September 2025 under which it "initially committed to pay" up to approximately $14.2 billion through 14 December 2031. The order carries an option to expand through 2032. That language comes from the 8-K filed on 30 September 2025.
Six months later, CoreWeave's Q1 2026 results reported "multiple new agreements with Meta, including a new $21 billion commitment signed in March." Together that is $35.2 billion of disclosed Meta commitments. On the evidence available, Meta is the strongest candidate for Customer B, but no filing says so and I am not going to assert it.
The rest of the named list is thinner. The same Q1 release announced a multi-year agreement with Anthropic, with no value disclosed, and expansions with Cohere, Jane Street and Mistral. Jane Street then made a $1 billion strategic investment in CoreWeave, a customer funding its supplier in a pattern covered in the explainer on circular AI deals.
Concentration inside the backlog is falling faster than revenue concentration
Revenue tells you who pays today. Remaining performance obligations, or RPO, tell you who has signed to pay later. RPO is contracted revenue not yet recognised, and CoreWeave's stood at $103.7 billion at 30 June 2026, with 41% expected to be recognised within 24 months.
The disclosed OpenAI and Meta maximums add up to $57.6 billion. Set against the June RPO, that is about 56% of the backlog tied to two named customers, on the most generous reading. The true share is lower, because OpenAI's 2025 orders have been delivering since last year and part of their value is already recognised revenue rather than backlog.
The company's own backlog concentration figure has moved even faster. On the Q3 2025 earnings call, CFO Nitin Agrawal said no single customer represented more than approximately 35% of revenue backlog, as reported by DatacenterDynamics. That was down from about 50% the previous quarter and about 85% at the start of 2025. The same call put more than 60% of backlog with investment-grade customers.
That is a secondary report of a spoken figure, and it has not been restated in any filing we read. It is also a year old. Backlog has nearly doubled since then, and the Q2 2026 results noted more than $25 billion of net new commitments in early Q3 that were not attributed to any customer. Treat the 35% as the last public data point, not the current one.
Setting CoreWeave customer concentration against the $35.6 billion debt stack
Concentration matters for CoreWeave in a way it does not for a software company, because almost every dollar of capacity is financed with debt. The site's breakdown of CoreWeave's $35.6 billion debt stack covers the facilities and the spread ladder. This section asks a narrower question: how much of the interest bill depends on which customers.
Net interest expense was $640 million in Q2 2026. Revenue was $2,575 million. That means interest absorbed about 25 cents of every revenue dollar before a single GPU was powered or depreciated. The scenario table removes customers one at a time.
| Scenario | Revenue left, $ millions | Interest as a share of revenue |
|---|---|---|
| All customers, as reported | $2,575 | 25% |
| Without Customer A | About $1,648 | 39% |
| Without Customer B | About $1,905 | 34% |
| Without Customers A and B | About $978 | 65% |
| Without Customers A, B and C | About $720 | 89% |
Illustrative only. Customers do not vanish overnight, contracts run for years, and capacity freed by one customer can be re-let. The table shows dependence, not a forecast.
What the secured lenders already price
The lenders worked this out before anyone wrote it up. CoreWeave's secured facilities are each backed by specific customer contracts, and the debt post shows their spreads ranging from SOFR plus 2.25% to SOFR plus 5.50% depending on the counterparty and contract length. In effect, the loan market prices concentration contract by contract.
That has a consequence for how you read the headline number. A falling top-customer share does not lower CoreWeave's borrowing cost by itself. What lowers it is the credit quality of whoever replaces the top customer. Swapping one investment-grade hyperscaler for an AI lab with heavy losses can raise funding costs even as the concentration ratio improves. The timing of when those secured loans come due is covered in the CoreWeave debt maturity ladder.
Where this argument is weakest
This post stacks arithmetic on top of lettered disclosures. Three parts of it deserve less weight than the tables imply, and one competing view genuinely beats it.
The letters may not be who you think
The dollar continuity of Customer A depends on A being the same company in every period. The filings explicitly refuse to promise that. If a different customer took the top spot in early 2026, the "flat top customer" reading in this post is wrong, and Microsoft's spend may have fallen rather than plateaued. The Q3 2026 10-Q will not settle it either, because the letters will still be letters.
The "up to" commitments carry a similar problem. A maximum contract value is a ceiling set by the customer's option to buy, not a schedule of payments. Adding three OpenAI ceilings and two Meta ceilings produces a number nobody has promised to pay in full.
Concentration in strong hands can be a feature
The strongest counter-argument is that concentration with Microsoft and Meta is safer than diversification across dozens of AI startups. Both are among the most creditworthy companies in the world, and the CFO's statement that more than 60% of backlog sat with investment-grade customers supports that view.
I think that argument wins for Microsoft and Meta and loses for OpenAI, whose ability to pay depends on continued fundraising. It also says nothing about renewal. A creditworthy customer can pay every invoice and still decline to renew in 2029, which is the risk the secured lenders are visibly pricing.
What to check in CoreWeave's Q3 2026 filing
CoreWeave's next 10-Q, covering the quarter to 30 September 2026, will add one more row to every table in this post. Four lines in it will tell you whether the de-concentration is continuing.
First, the Customer A share. A figure below 36% continues the trend, while a figure above it suggests the Q2 low was a timing effect. Second, the count of customers above 10%. Three in Q2 was a record, and four would mean the base is genuinely broadening.
Third, the receivables split. If Customer B's share of receivables keeps running ahead of its revenue share, ask why in plain language on the earnings call. Fourth, any 8-K naming the counterparty behind the more than $25 billion of early Q3 commitments. A new named customer of that size would reshape the backlog more than any percentage in the 10-Q.
The same four lines are worth comparing with other suppliers' disclosures. Nvidia's own concentration is tracked in the analysis of Nvidia's customer concentration. Backlog definitions across the clouds differ in ways set out in the comparison of cloud backlog and RPO.
Frequently asked questions
Who are CoreWeave's biggest customers?
Microsoft was CoreWeave's largest customer in 2025, at about 67% of revenue, according to the company's 10-K. OpenAI and Meta are the other two customers CoreWeave has disclosed contracts with, at up to $22.4 billion and up to $35.2 billion respectively. The quarterly filings label customers by letter only, so the current ranking cannot be confirmed by name.
What percentage of CoreWeave revenue comes from Microsoft?
About 67% of CoreWeave's 2025 revenue of $5.1 billion came from Microsoft, per the FY2025 10-K. Coverage of the IPO filing put Microsoft at 62% of 2024 revenue. For 2026, the 10-Q filings show the largest customer at 45% in Q1 and 36% in Q2, but they do not name that customer, so the 2026 Microsoft share is not disclosed.
How much is CoreWeave's contract with OpenAI worth?
CoreWeave says its contracts with OpenAI total up to approximately $22.4 billion. That figure combines an initial agreement worth up to $11.9 billion in March 2025, an expansion worth up to $4 billion in May 2025, and a further order worth up to $6.5 billion in September 2025. All three are maximum values rather than guaranteed payments.
How much has Meta committed to CoreWeave?
Meta committed up to about $14.2 billion through December 2031 under an order form filed in September 2025, with an option to expand through 2032. CoreWeave then reported a new $21 billion Meta commitment signed in March 2026. Combined, the two disclosed Meta commitments come to $35.2 billion, close to CoreWeave's $35.6 billion of debt principal at June 2026.
Is CoreWeave customer concentration a risk to its debt?
Yes, though less than a year ago. In Q2 2026 net interest expense was $640 million against $2,575 million of revenue. Removing the two largest customers would leave about $978 million, so interest would absorb 65% of what remained. Secured lenders price this directly, charging wider spreads on facilities backed by shorter or weaker customer contracts.
What is CoreWeave's revenue backlog and who holds it?
CoreWeave reported $103.7 billion of remaining performance obligations at 30 June 2026, with 41% due within 24 months. The company does not break backlog down by customer in its filings. On the Q3 2025 call, management said no single customer held more than about 35% of backlog. Disclosed OpenAI and Meta maximums total $57.6 billion.
Where to start
If you hold CoreWeave debt or equity, or buy capacity from it, build the dollar table yourself before the Q3 filing lands. It takes ten minutes: multiply each lettered share by quarterly revenue and line the results up by period. The direction of Customer A's dollars tells you more than its percentage.
If you run a business with a customer above 30% of revenue, apply the same scenario table to your own numbers. Remove the top customer, then the top two, and see what share of fixed costs the remainder covers. That one line is the honest measure of how concentrated you are.
References
- CoreWeave, Inc., Form 10-K for fiscal year 2025, filed 2 March 2026. Used for Microsoft's 67% share, top-two and top-three shares for 2024 and 2023, 2025 revenue, and risk factor language on OpenAI and Meta.
- CoreWeave, Inc., Form 10-Q for the quarter ended 31 March 2026, filed 8 May 2026. Used for Q1 2026 and Q1 2025 customer shares, receivables concentration and revenue.
- CoreWeave, Inc., Form 10-Q for the quarter ended 30 June 2026, filed 12 August 2026. Used for Q2 2026 and Q2 2025 customer shares, receivables, RPO of $103.7 billion and balance sheet debt.
- CoreWeave, Inc., Form 8-K on the Meta order form, filed 30 September 2025. Used for the $14.2 billion Meta commitment and its term.
- CoreWeave, CoreWeave Expands Agreement with OpenAI by up to $6.5B, 25 September 2025. Used for the $11.9 billion, $4 billion and $22.4 billion OpenAI figures.
- CoreWeave, CoreWeave Reports Strong First Quarter 2026 Results, 7 May 2026. Used for the $21 billion Meta commitment, the Anthropic agreement and the Jane Street investment.
- DatacenterDynamics, CoreWeave Q3 earnings show revenue backlog doubled to $55.6bn, November 2025. Secondary report of the CFO's backlog concentration statement. Upgrade to the call transcript when available.
Weakest point in the source base: the quarterly filings never name customers, so every link between a letter and a company is inference. The FY2024 Microsoft share and the 35% backlog figure rest on secondary coverage. The $35.6 billion debt principal and $640 million of Q2 net interest are taken from the Q2 2026 10-Q as tabulated in our debt post; the balance sheet carrying value is $35.1 billion. Last verified 3 November 2026, before CoreWeave's Q3 2026 filing.
Related reading