We use media reports of OpenAI and Anthropic revenue to estimate those companies’ revenues across 2023 and 2024. These are typically reports of figures from investor materials, or statements from company leadership. For OpenAI, we found media reports of total revenue in 2023, as well as 7 instances of reported monthly revenue. For Anthropic, we found 4 instances of reported monthly revenue. This data is shown below.

For our estimates of Google DeepMind revenues, we use publicly available third-party estimates of monthly visits to chat interface webpages, monthly active users for chat applications, monthly mobile app downloads, and monthly visits to the API documentation.

We estimate monthly revenues for Anthropic from January 2023 to December 2024 by fitting exponential growth to the four reported monthly revenue figures. We sum the estimated monthly revenues to obtain our half-yearly revenue estimates.

For OpenAI, we account for a jump in revenue due to the February 2023 launch of ChatGPT paid subscriptions. For August 2023 onwards, we fit exponential growth. For February to August 2023, we separately fit exponential revenue growth to match the total 2023 revenue when the entire year is summed. As with Anthropic, we sum the monthly revenues to obtain half-yearly revenues.

We model uncertainty in reported revenues for the date of each report (normal distribution with 90% CI of ±~2 weeks) and the revenue (normal with 90% CI of ±30%). We model additional uncertainty for Anthropic, since there is a gap of over 1 year between the first and second revenue report. We sample a range of plausible piecewise interpolations between these data points, allowing between 10% and 90% of revenue to occur at a date between 10% and 90% of the way through the gap. For all uncertainty estimates, we run 1,000 Monte Carlo samples, re-fitting or re-interpolating in each sample, then take percentile bands across the resulting revenue trajectories to form our confidence intervals.

Google DeepMind reported a revenue of £1.5B (equivalent to $1.9B) for 2023. However, it’s unclear how to infer revenue from external customers from this figure, because it represents “research and development remuneration from other group undertakings”. Because of this, we instead estimate Google DeepMind’s customer revenue using web traffic and mobile app usage data.

To estimate Google DeepMind’s subscription revenue at the end of 2024, we use three different proxies. First, we leverage monthly active users data for the Gemini, ChatGPT, and Claude mobile apps from DataReportal. Second, we use mobile app download statistics from Business of Apps and Backlinko. Third, we analyze chatbot webpage monthly visits from Similarweb. For each case, we assume Google DeepMind derives comparable subscription revenue per user interaction as OpenAI and Anthropic, within a 5x factor of uncertainty. We take the geometric mean of these estimates to form our final estimate for subscription revenue.

For API revenue estimation, we use developer documentation webpage visits from Similarweb as a proxy for API usage. We assume Google DeepMind derives comparable API revenue per developer webpage visit as OpenAI and Anthropic. As we did for the subscription revenue, we take the geometric mean of these estimates to form our final estimate for API revenue. We then add the subscription and API revenue estimates together to get our total revenue estimate for the end of 2024.

To extrapolate Google DeepMind’s revenue through 2023-2024, we calculate the monthly growth rate of each proxy metric for Google DeepMind. We apply this growth rate to our December 2024 estimates to generate a complete monthly revenue series. We assume subscription revenues were zero before April 2024, as Gemini’s first paid plan launched in February 2024 with a two-month free trial.

We use a similar Monte Carlo approach for DeepMind, sampling 1,000 times over key input parameters (usage metrics, subscription fractions, etc.). We model our uncertainty over the real value of revenues per proxy unit as a 90% confidence interval between ⅕ and 5x of the value for OpenAI and Anthropic. We also model uncertainty over the growth rate, sampling noise from a 90% confidence interval between 0.8 and 1.2 of the estimated growth rate. For each sample, we compute a revenue trajectory, then form confidence intervals by taking percentiles across all samples.

Our revenue estimates for OpenAI and Anthropic align reasonably well with public estimates made by others – the disagreement between our estimates and the estimates we found vary from 0% to 15%. On the other hand, our revenue estimates for Google DeepMind disagree with other public estimates by 48% to 85%, although it’s unclear whether these estimates are focused on external customer revenues, so this might not be a like-for-like comparison.

• We obtain our revenue figures from media reports, which are generally based on comments from OpenAI and Anthropic representatives, as well as leaked financial documents in some cases. We explicitly model uncertainty in the timing and accuracy of these reports.

• We focus our analysis on revenues from external customers, omitting “implicit” internal revenues. For example, Google may generate significant additional search engine revenue by using Google DeepMind’s models in its “AI Overview” product, but we do not take this kind of company-internal revenue into account. These internal revenues might be much larger than external customer revenues for Google DeepMind.

• Similarly, we only include revenues generated by models developed in-house. For example, Microsoft Copilot and Azure sell access to OpenAI and other companies’ models, and are perhaps the largest source of AI revenues (reportedly generating $13B in annualized revenues as of January 2025). There may be considerable revenue accruing to providers like AWS and Google, whose cloud infrastructure is used to serve third-party models.

• Google DeepMind’s revenues are estimated using web traffic and app usage proxies. While we explicitly model uncertainty in revenue per unit of each proxy, there are likely biases between organizations. For instance, many third party developers have replicated the OpenAI chat endpoint API for their own models, which may drive additional web traffic to the OpenAI developer documentation that does not translate into revenue.

• We believe that no other AI companies had direct revenues of over $100M in 2024. The table below shows our analysis for other prominent companies.