AI giants get 80% of enterprise revenue from 1% of customers, and most of those are kept alive by VC
OpenAI and Anthropic get 80% of their enterprise revenue from 1% of customers, and many of those are AI companies burning VC-subsidised tokens; the only ones willing to pay real money are tech itself, and when VC stops, the trillion-dollar compute bill blows up first.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
80% of enterprise revenue rides on 1% of customers
Ramp's enterprise spending data shows that roughly 80% of OpenAI's and Anthropic's enterprise customer revenue comes from the top 1% of customers, and that ratio has not improved at all over the past three years. Ramp chief economist Ara Karazian points out that this 1% is heavily skewed toward the tech industry and AI product/service companies, a level of concentration he has not seen in any other software category he tracks. The host adds the data's boundary: it excludes players like Microsoft and the big banks, but it is still representative of the overall state of AI industry spending. In other words, what actually props up the enterprise revenue of the two leading model makers is not a broad mass market, but a small handful of extremely free-spending customers.
— Ed ZitronThe big AI spenders are funded by VC, not revenue
This spending structure holds up largely because AI startups are kept burning money on VC transfusions. The common play is to sell subscriptions but subsidise tokens: a user on a $20-a-month subscription can burn $30, $40 or even $100 worth of tokens, with the gap covered by capital. That is, the money OpenAI and Anthropic receive from these customers depends on those customers being able to keep raising. The host admits this is his intuition: the loudest voices on AI Twitter are basically VC-backed startups, and the tokens they burn internally writing code may far exceed what outside customers pay for. So-called real demand is, at least in large part, actually propped up by ‘the ability to raise’.
— Ed ZitronThe only ones willing to pay big for AI are tech itself
Lay this data out and the conclusion is more awkward than most people want to admit: AI has occupied every media outlet and every meeting room for three straight years, almost every company has tried it, plenty have approved big AI budgets, but the ones actually willing to spend big turn out to be, it seems, only the tech industry itself. Industries outside tech and AI barely pay high prices for large models, and there is no sign of any reversal. Worse, this small set of leading customers is itself unstable — they could easily switch to cheaper self-built open-source models, or eventually move to on-device models. The host points out that this customer structure is the real underlying exposure of OpenAI and Anthropic.
— Ed ZitronCloud vendors have staked their fortunes on two model companies
The risk does not stop at the two labs themselves. The compute purchase commitments on OpenAI and Anthropic exceed $1.1 trillion, and they are take-or-pay contracts — no matter how much is actually used or earned, the agreed amount must be paid. The compute revenue of Microsoft, Google, Amazon, Oracle and CoreWeave is all betting that these two can pay. The figures the host cites: OpenAI contributes about 70% of Microsoft's FY26 AI revenue; UBS estimates OpenAI and Anthropic will account for 48% of Google Cloud's total revenue next year, corresponding to roughly $84 billion to $100 billion; Barclays also gives 2027 forecasts of about $40 billion for AWS and at least $50 billion for Azure. And both companies are still losing billions of dollars a year.
— Ed ZitronThe compute bill hits a reset wall in 2027
Why hasn't the bubble burst yet? Because the mortgage-style structure hasn't reached its repayment peak. These compute contracts usually start with a small prepayment, the data centre capacity isn't online yet, and the real payment obligation gets heavier month by month starting next year. The host cites an analogy from a Groundbreaker article: it's like a mortgage's low teaser rate expiring — the rate was extremely low at signing, and after the reset the monthly payment balloons to something unsustainable; borrowers all assumed at signing that house prices would keep rising and that they could refinance or sell, but when everyone tries to run at the same time, no one can get out. OpenAI's and Anthropic's enormous compute commitments are the subprime of the AI bubble. The pretty big orders leave the cloud vendors carrying huge revenue backlogs, while the ability to pay rests on a belief in continued growth and VC money that keeps arriving.
— Ed ZitronNVIDIA sells cards on its customers' debt
The very top of the chain is just as dangerous. The reason hyperscalers and neoclouds buy more GPUs is the demand brought by OpenAI and Anthropic; even those who build data centres to sell to third parties are only referencing the revenue backlog on the books of companies like CoreWeave and Nebius — and those companies' customers are mainly OpenAI, Anthropic, or the cloud giants behind them. More glaring still, NVIDIA's customers can no longer buy GPUs out of operating cash flow alone; every purchase depends on debt remaining available. Broadcom's latest earnings also show that Anthropic and OpenAI will become its two largest customers. The host closes this layer with irony: ‘Everything is normal, everything is fine, no one needs to panic.’
— Ed ZitronBloomberg pops a new mega-deal, and no one worries
The host points out that sell-side analysts and the financial media will reassure you: the smartest people and the most powerful companies don't burn money for no reason, AI demand is real, and sceptics are just cherry-picking data. But just as his own monologue is about to end, a news item pops up on the Bloomberg terminal: Crusoe signs a cloud computing contract worth about $13 billion with Jane Street. The irony is that Jane Street is itself both a major CoreWeave customer and a CoreWeave shareholder — string those identities together and it is exactly a sample of the concentration risk he has been describing all episode. He judges it a catastrophic capital misallocation, and also the biggest collective failure to report in the history of journalism.
— Ed ZitronIn their own words · checked verbatim
Per data from fintech firm Ramp, 80% of OpenAI and Anthropix Enterprise revenues come from 1% of their customers, a number that hasn't improved over the last three years.
Ed Zitron1:03
Most businesses have been given the green light to spend a bunch of money on AI. And in the end, it seems that the only people the tech industry can get to spend money on AI is the tech industry itself.
Ed Zitron4:08
These are, from what I can tell, take-or-pay agreements where they agree to buy that compute capacity regardless of how much capacity they actually end up using and how much revenue they actually bring in.
Ed Zitron5:09
In other words, OpenAI and Anthropic's massive compute commitments are the subprime mortgages of the AI bubble. They signed big, beautiful deals that helped hyperscalers and neoclouds post massive revenue backlogs under the belief that nothing bad would ever happen.
Ed Zitron7:11
Oh, and NVIDIA's customers are no longer able to buy its GPUs through cash flow alone, so all of those purchases are contingent on their constantly availability of debt.
Ed Zitron8:11
I am literally looking at my fucking Bloomberg terminal, and what just popped up says, Crusoe signs roughly $ 13 billion Jane Street deal for cloud computing. ... Jane Street's a major customer of CoreWeave and an investor in CoreWeave.
Ed Zitron9:12
Figures
| Share of OpenAI/Anthropic enterprise revenue from top 1% of customers | 80%, unchanged over the past three years | 1:03 |
| Total compute purchase commitments by OpenAI + Anthropic | Over $1.1 trillion (take-or-pay) | 4:08 |
| OpenAI's share of Microsoft AI revenue (FY26) | About 70% | 5:09 |
| OpenAI + Anthropic projected 2027 spending on AWS (Barclays) | About $40 billion | 5:09 |
| OpenAI + Anthropic projected 2027 spending on Azure (Barclays) | At least $50 billion | 5:09 |
| Expected 2027 revenue for the three cloud vendors from OpenAI/Anthropic (low-end figure) | Over $174 billion | 5:09 |
| Value of the Crusoe–Jane Street cloud computing contract | About $13 billion | 9:12 |
Glossary
- concentration risk
- Revenue or investment overly concentrated in a few customers, businesses or assets, so that it collapses once they leave.
- take-or-pay agreement
- A long-term purchase agreement in which the buyer must pay the agreed amount regardless of how much it actually uses.
- neocloud
- A new generation of cloud vendors specialising in renting out GPU compute, such as CoreWeave.
- revenue backlog
- Future revenue already contracted but not yet delivered, often used to prove that demand is real.
How to listen
AI founders, VCs and enterprise tech buyers: if you want to know how fragile the revenue of the big model vendors and cloud giants really is, this episode lays out the transmission chain clearly.