โŒ•

ONE CUSTOMER WAS 62% OF THE BUSINESS.

NOV 20, 2026 ยท SPLASH ยท BUSINESS ยท THE CONCENTRATION
CoreWeave, Scale AI and the concentration risk behind ARR

ONE CUSTOMER WAS 62% OF COREWeAVE'S ENTIRE BUSINESS. That customer was Microsoft — which also runs Azure, one of CoreWeave's competitors. If Microsoft sends one email, most of the revenue is gone. And CoreWeave is not unusual. Some of the fastest-growing companies in the world get almost all their money from two or three giants, and the risk is hidden inside three letters: ARR.

THE BET THAT WORKED

CoreWeave started as a crypto miner with GPUs. Then ChatGPT happened and those chips became the scarcest thing on earth.

Microsoft needed computing power faster than it could build its own data centres, so in 2023 it started signing huge contracts with CoreWeave. Revenue went from $16 million in 2022 to $1.92 billion in 2024, with more than $15 billion of contracts still to deliver.

The filing said the quiet part at the top: “A substantial portion of our revenue is driven by a limited number of customers.” Microsoft went from 35% of revenue in 2023 to 62% in 2024. The top two customers were 77%.

It worked anyway — and the reason is the second half of this page. Those were binding multi-year contracts where Microsoft pays whether it uses the capacity or not. If you need billions in loans to buy Nvidia chips and build data centres, that contract is the best thing you can show a bank.

THE DAY IT BROKE

June 2025 was supposed to be the best month in Scale AI's history. The company labels and prepares the training data that AI models learn from, and its customer list was basically everybody: Google, OpenAI, xAI, Meta.

12 June 2025: Meta agrees to pay $14.3 billion for 49% of the company, valuing it near $29 billion — more than double its value a year earlier. Founder Alexandr Wang leaves to run Meta's new superintelligence lab.

Now look at it from Google's side. You have been sending your most sensitive training data to an outside company, and your biggest rival just bought half of it and hired its chief executive.

Within days, Reuters reported that Google — Scale's biggest customer, planning to spend about $200 million that year — was moving to cut ties. OpenAI said it was winding down its work with Scale. xAI pulled back too.

Scale cut 500 contractors. Interim chief executive Jason Droege wrote that they had grown too fast and built too many layers of management, and that the restructuring would help win back customers who had slowed down.

The product did not get worse. The data was as good the day after the deal as the day before. What changed was trust — and trust is not in the revenue model.

THE OLD VERSION OF THE SAME STORY

2010. Zynga — FarmVille — made $597 million a year, almost all of it from people buying virtual goods inside its games.

When it filed to go public in 2011, the first risk factor on the list was one company: Facebook. In its own words, it made “substantially all of our revenue and players through the Facebook platform.”

Facebook was how people found the games, how Zynga marketed them, and how players paid. Every decision that mattered to Zynga got made in somebody else's conference room.

Then Facebook moved to mobile and changed the news feed, and the traffic dried up. Zynga went public at $10 a share late in 2011, fell below $3 the next year, and spent nearly a decade underwater.

TWO COMPANIES, ONE MILLION DOLLARS

Picture both doing a million a year.

The first sells $2,000 subscriptions to 500 small businesses. If one leaves, nobody notices.

The second has a single $1 million contract. Same revenue. If that customer leaves, the business is zero.

Every investor on earth pays more for the first, because the cash is predictable. That is why software gets the valuations it does: same product, thousands of customers, contracts that renew, every customer a replaceable slice.

When 60% or more of your revenue comes from one or two logos, that maths breaks. You stop looking like a software company and start looking like an agency with one relationship.

THE LAYER UNDERNEATH

Sometimes the concentration is not in the customer list at all.

An AI startup can show you 500 paying customers and still have a single point of failure one layer down: every feature running on one model provider's API, or every sale coming through one cloud marketplace.

The list looks spread out. The dependency does not.

THE THREE LETTERS

ARR — annual recurring revenue — is supposed to mean subscription money that renews every year, the kind you can count on. It is the number behind every viral funding announcement.

Sometimes it means something else. TechCrunch looked into it in May 2026 and found at least three different things being called ARR: real live subscription contracts; one big month of usage multiplied by twelve; and CARR — contracted ARR — which counts signed paperwork that has not launched and has not collected a dollar.

In April 2026, Scott Stevenson, co-founder and CEO of the legal AI company Spellbook, called it out publicly on X as “a huge scam”. TechCrunch then talked to insiders: one former employee described companies signing big enterprise free year-long pilots and counting what the contract might be worth after the pilot as ARR today. One investor said a startup's CARR was 70% higher than the revenue actually coming in. Another said everyone has a company monetising CARR as ARR.

To be fair, tracking signed contracts is not wrong — firms like Bessemer use ARR to measure enterprise momentum. The abuse is pitching unbilled pilots as live revenue.

Jamin Ball of Altimeter has a name for the weakest version: experimental recurring revenue. Big companies run three-month AI trials out of a temporary innovation budget, partly so they can tell their board they are doing AI. The moment the finance chief tightens the budget, it is gone.

Which is how two Fortune 500 free one-year trials, cancellable with 30 days' notice, become “$5 million ARR” on a pitch deck.

WHY YOU ONLY HEAR ABOUT IT AFTER AN IPO

Under US accounting rules, any customer that makes up 10% or more of revenue has to be disclosed. That is the only reason the world knows about Microsoft and CoreWeave.

Private startups do not have to tell anyone. So the concentration is invisible until the filing — or until a customer leaves.

AND THE BIGGER QUESTION

If two giant buyers are your business, do you even have product-market fit?

Real product-market fit is when lots of different buyers pull the same product out of your hands without you building something custom for each one. A seven-figure contract proves you solved a painful problem for one company. It does not prove the next hundred have that problem.

SPLASH'S READ

Strip the acronyms and this is the same story as every other page in this lane: the number in the headline is not the number that survives.

CoreWeave survived its concentration because the contracts were binding and the chips were scarce — and it then used that one giant buyer to go sign an $11.9 billion, five-year deal with OpenAI that pushed Microsoft below half its backlog. That is how you use a whale: they pay for the build, then you go find the next one.

The question this video ends on is the only one that matters: how many emails would it take to make that number disappear? If the answer is one, you do not have revenue. You have a relationship, and relationships end.

AD SPACE ยท RESERVED
MORE ON THE BOARD โ–ธ
๐Ÿ’ผ All Business ยท The Nvidia Guarantees
Sources & claims: Sources. CoreWeave: IPO filing figures and the disclosure that a substantial portion of revenue is driven by a limited number of customers; revenue of $16 million in 2022 rising to $1.92 billion in 2024, with more than $15 billion of contracted backlog at filing; Microsoft at 35% of revenue in 2023 and 62% in 2024, with the top two customers at 77% (S-1; reporting at the March 2025 IPO, priced at $40 a share and raising $1.5 billion); the Financial Times report that Microsoft walked away from some commitments, which CoreWeave denied; the $11.9 billion five-year OpenAI agreement. Scale AI: Meta's 12 June 2025 agreement to pay $14.3 billion for 49% at a valuation near $29 billion, with Alexandr Wang joining Meta (Scale AI's own announcement; CNBC); Google's roughly $200 million planned spend and Reuters' report that it moved to cut ties within days, with OpenAI winding down and xAI pulling back; 500 contractors cut and the interim-CEO memo from Jason Droege (Scale AI, 18 June 2025). Zynga: $597 million of 2010 revenue; the 2011 S-1 naming Facebook as its first risk factor and stating that Facebook made “substantially all of our revenue and players”; the $10 IPO price and the fall below $3. ARR definitions: TechCrunch, 22 May 2026, and Scott Stevenson, co-founder and CEO of Spellbook, whose April 2026 post called the practice “a huge scam”; Jamin Ball of Altimeter for “experimental recurring revenue.” The 10% customer-disclosure rule is US accounting requirement, described here generally. Video source: Gabriel Jarrosson, “One Customer Leaves. The Startup Dies.” (lA85qZzj_Zw) — read in full and mined for the sequence; every figure above was checked against the sources named here, and three names the transcript mangled (Droege, Ball, Spellbook) were corrected against the record. Related: Revel and the Nvidia guarantees.