OPENAI
The compute bill is catching up fast
OpenAI’s finances are back in the spotlight as IPO rumours grow.
Leaked documents shared with tech writer Ed Zitron give a clearer look at what the company is earning, and how much it’s spending to keep its models running.
The documents show Microsoft received $493.8m in revenue-share payments from OpenAI in 2024.
In the first three quarters of 2025, that jumped to $865.8m.
OpenAI is widely reported to give Microsoft around 20% of its revenue as part of their multi-billion-dollar partnership, though neither side has confirmed that number.
Microsoft also reportedly sends OpenAI about 20% of the revenue from Bing and the Azure OpenAI Service, since both rely on OpenAI’s models.
A source told TechCrunch that the leaked figures only reflect Microsoft’s net revenue share, meaning Microsoft has already deducted whatever it paid OpenAI from Bing and Azure royalties.
Because Microsoft doesn’t break out those numbers publicly, it’s hard to know the full scale.
Using the 20% estimate, the leaked payments suggest OpenAI made at least $2.5bn in 2024 and $4.33bn in the first nine months of 2025.
Other reports say 2024 revenue was closer to $4bn.
CEO Sam Altman has also said OpenAI’s yearly revenue is already “well more” than $13bn and could end the year above a $20bn annualised run rate, with a long-term projection of $100bn by 2027.
The spending side is where the pressure shows. Zitron estimates OpenAI spent about $3.8bn on inference, the compute needed to run models, in 2024.
That rose to roughly $8.65bn in the first nine months of 2025.
In brief:
Leaked figures point to strong revenue growth, but also very high compute costs.
Training is largely covered by Microsoft credits, but inference costs are mostly cash.
The numbers add to questions about whether the AI sector’s current economics can hold.
Where the money’s actually going
Previous reports suggest total compute spend in 2024 was around $5.6bn, with a “cost of revenue” of $2.5bn in the first half of 2025.
OpenAI mostly relies on Microsoft Azure for compute, but it also works with CoreWeave, Oracle, AWS and Google Cloud.
According to TechCrunch, most training costs are covered by Microsoft credits, but inference costs are paid in cash.
Taken together, these numbers suggest OpenAI may currently be spending more to operate its models than it earns, a point that fuels wider debate about whether today’s AI market is sustainable.
This is what happens when your hobby is “training God-level models.” - MG