← The economics of intelligence

What do we know about OpenAI’s finances?

OpenAI is still private, so it does not publish 10-Ks or 10-Qs. It reportedly filed for an IPO confidentially, which means we cannot see the filing yet. The numbers below come from documents shared with shareholders and reporting about those documents. They are not audited public filings.

Q1 2026

Revenue$5.7B
Cost of serving the models$3.5B
Gross profitabout $2.2B
R&D, including model training$8.6B
Operating loss$9.3B
Cash burn$3.7B
Net loss$21.3B

Source: The Information, based on documents OpenAI shared with shareholders.

How did it lose $21.3B?

OpenAI made money on serving its models. Customers paid $5.7B and serving them cost around $3.5B, leaving about $2.2B in gross profit.

But that gross profit did not cover R&D, employees, stock compensation, sales, marketing and administration.

Revenue                         +$5.7B
Serving the models              -$3.5B
                                -------
Gross profit                     $2.2B

R&D                              -$8.6B
Other operating expenses         -$2.9B
                                -------
Operating loss                  -$9.3B

Mostly accounting charges      -$12.0B
                                -------
Net loss                        -$21.3B

The entire $8.6B of R&D was not necessarily spent on frontier training. Training is included, but OpenAI has not disclosed the exact share.

What was the accounting loss?

About $12B came from a change in the estimated value of convertible investor rights and a warrant. If OpenAI becomes more valuable, the shares it has promised to someone else also become more valuable. Accounting treats that increase as a loss, even though OpenAI did not pay out $12B in cash that quarter.

It is still a real economic cost because those rights can become shares and dilute the other owners. But it is different from cash leaving the bank.

Who holds those rights?

The confidential financial documents do not show us exactly how the $12B charge is divided. The two clearest public examples are:

Microsoft owns around 27%. Current and former employees and other investors own the remaining roughly 36%.

Sources: SoftBank, OpenAI’s structure and the Microsoft partnership.

What this tells us

Inference itself was not losing money. OpenAI reported a gross margin of around 39%. The problem is that the resulting gross profit was nowhere near enough to cover R&D and the rest of the company.

At the same 39% margin, OpenAI would need roughly $22B of quarterly revenue to produce $8.6B of gross profit. That would only equal the reported R&D expense. It would not yet pay for the rest of the company.

This is only an intuition. R&D contains more than training, margins can change and we do not know which revenue came from subscriptions, APIs, ads or enterprise contracts.

What we still do not know