The Same AI Dollar Is Being Counted Two or Three Times
Much of AI's booming sales is the same customer payment moving from app to model to cloud.
20VC with Harry Stebbings · Anthropic's Super Bowl Ad: Who Won & Lost? | Sierra Hits $150M ARR: Is Customer Support Too Crowded?Add up the revenue headlines in artificial intelligence and you will double count the same money.
The problem is that one customer payment travels up a stack, and each layer reports part of it as its own sales. The customer pays an app. The app pays for the AI model it uses. The model maker pays the cloud service that runs it, and for the chips underneath. Each company books revenue, but a fresh dollar entered only once.
Mike Cannon-Brookes, the co-founder of software company Atlassian, described how it works with his own bills. When Atlassian uses models from Anthropic, a leading AI lab, it does not usually pay Anthropic directly. It pays Amazon Web Services, the cloud rental business, and Amazon then pays Anthropic. So if a customer spends on Atlassian software, part of that payment shows up as Atlassian revenue, then as Amazon revenue, then as Anthropic revenue.
For example, Anthropic reports $1 billion in revenue and Cursor, a popular coding assistant built on models like Anthropic's, also reports $1 billion. It is largely the same sales recorded twice at different layers of the stack.
That matters when judging forecasts. Anthropic has projected about $150 billion in annual recurring revenue, often shortened to ARR, by 2029 in its most optimistic case. The figure for OpenAI for the same year was put at about $180 billion in the discussion. Together that is $330 billion to $380 billion, in a world where total software sales are about $700 billion and Microsoft does about $200 billion in revenue.
As one of the other participants noted, Anthropic growing from about $4.5 billion today to $150 billion would mean adding essentially another Microsoft. With OpenAI at a similar size, that would be two more Microsofts to be fed from the same pool of customer spending, unless technology budgets grow much larger.
Questions people ask
If revenue is double counted, does that mean AI profits are double counted too?
No. Revenue stacks, but profit does not. When Atlassian pays Amazon and Amazon pays Anthropic, each company keeps only the difference between what it charges and what it pays. The costs pass upward along with the revenue, so the stacked profit is much smaller than the stacked sales.
Why don't app companies just buy models directly and avoid the middle layer?
Many buy through a cloud provider because the cloud already handles computing, security, billing, and reliability. Amazon, Microsoft, and Google offer access to several leading models in one place, and let a company switch models or balance traffic by cost, speed, and quality. Direct deals exist, but routing through the cloud is often simpler.
Could AI budgets still grow enough to support those big forecasts?
Possibly, if software takes over work now done by people. The global market for consulting and services work is roughly $1 trillion, much larger than software. If AI tools replace some of that labor, companies could shift those wages into software budgets, which would expand the total available market.