Epoch looks at more types of data, including Usage, Compute Spend, and Valuation.
Epoch does not look at projections, either those made by the company itself or by extrapolating the current trend.
Differences in methodology, for the things we both consider:
Epoch uses reports of annualized revenue (the revenue in a month x12), while I focus on annual revenue (the amount of revenue a company received in a calendar year). I discuss this in the post. The results are similar for OpenAI, the only company I have decent revenue data for.
Epoch does not use estimates for the number employees made by market research organizations, like LeadiQ or RocketReach. Instead, they focus on media reports of the number of employees. This difference does have significant consequences.
The number of employees from market research organizations is often higher than the media reports used by Epoch.
For Anthropic, the media reports Epoch used have been lower than the data I used, but they are converging.
For OpenAI, the data were similar until the end of 2023 (8 years since the company was founded). After this, the media reports have been about a factor of 2 lower than the market research organizations.
I do not know why this discrepancy exists. The market research organizations do not publish their methodology, but they presumably have incentives to provide accurate information, and so their estimates should be reasonable. Media reports from organizations like TheInformation also have incentives to be accurate, and try to avoid publishing unverified numbers. For OpenAI in particular, Noam Brown has cautioned against the higher numbers.
If we accept Epoch's data for number of employees, then OpenAI is growing at a rate of 1.7x per year and Anthropic is growing at a rate of 3x per year. In 2027, OpenAI would have 7400 employees and Anthropic would have 14,000 employees. In 2030, OpenAI would have 37,000 employees and Anthropic would have 360,000 employees. In 2033, OpenAI would have 180,000 employees and Anthropic would have 9.4 million employees. In 2033, Anthropic would be larger than any company today, and larger than the entire US tech industry. This would be dramatic, although it still seems like this would be less of a limitation than receiving all global wealth as capital.
The same day this was posted, Sri Muppidi at The Information shared some new projections for OpenAI's finances for 2025-2030.[1] The graph includes 3 projections: R&D compute cost, Inference compute cost, and Revenue.
OpenAI is not projecting current growth rates to continue. Comparing to the previous data, these projection show a much lower rate of growth.
R&D compute cost is not just projected to grow slower than an exponential over the next 5 years, it is projected to grow slower than a linear trend. Or, if we treat this as a logistic curve that has had exponential growth in the past, OpenAI is projecting that we will reach the point of steepest growth next year, and that the leveling off will be apparent in 2-3 years.
Compared to the previous projections, these projections suggest that OpenAI will be cash flow positive by 2027, not 2029 - assuming that R&D compute and Inference compute constitute most of their total costs. They also suggest that OpenAI currently has enough capital to last until then.
The revenue projections are not that different from the previous ones ($55B vs $60B for 2027), so it seems as though OpenAI is now projecting that it will spend less over the next few years than it previously projected. I think that the main difference is that compute is not the total costs. If we include the unspecified other costs, the projection might be the same as before.
To get to "R&D costs amount to ~45% of total revenue in 2030", there is also around $40B of other research spending - which seems to be mostly "compensation costs."
The LLM adoption wave is underway, which means current AI companies are receiving valuations at very high multiples relative to ARR, and correspondingly high investment rounds. Current ARR only matters as a lower bound (with some non-instant deceleration that should be priced in), an ingredient for models of where it's going. Spending is going to mostly follow what can be raised, even if it's out of touch with current levels of ARR, so comparing them is little use. Once growth slows, spending will adjust.
The crucial question is when and where the growth will slow down, but this is extremely uncertain, especially because the levels of LLM capabilities attainable in the short term (and therefore TAM) remain extremely uncertain. For this, the relevant indicators are ARR growth, valuation to ARR ratios, and TAM estimates. Currently the TAM estimates seem to be the closest to indicating the time to a possible slowdown, if capabilities don't significantly improve. For example it would be hard to find much more than $100bn in revenue from programming or legal assistants if they are not truly autonomous (say $20K per year for 5M professionals). And so in 2026-2028 these kinds of markets should start to visibly saturate, at least as far as Anthropic-level 10x YoY ARR growth is concerned, if capabilities don't sufficiently improve.
OpenAI can probably achieve Meta/Google-style revenue just from monetizing free users, since they're already one of the biggest platforms in the world, with a clear path to increasing eyeballs through model progress+new modalities and use cases, and building up an app ecosystem (e.g. their widely rumored web browser). An anonymous OpenAI investor explains the basic logic:
The investor argues that the math for investing at the $500 billion valuation is straightforward: Hypothetically, if ChatGPT hits 2 billion users and monetizes at $5 per user per month—“half the rate of things like Google or Facebook”—that’s $120 billion in annual revenue.
However, this might take a long time to fully realize, perhaps like 10 years?
OpenAI is a special case, they've captured the casual/free userbase. But my point is that ~$100bn per year is the current softcap (if capabilities don't significantly improve), about this much is feasible with either monetization of free users or AI assistants for highly paid knowledge workers, and probably all the LLM-enabled apps (on foundation model company APIs) also add up to something. Right now, it's not even clear if Anthropic or OpenAI will stumble first, since OpenAI's API business is weaker and their free user base already nears saturation, it can't keep growing very quickly for too long.
An interesting question is whether $250-500bn per year is feasible for one company within the current adoption wave, because that enables 10-20 GW training systems. There's a lot OpenAI might be able to do to monetize their users, and also they might be able to become competitive through API. The on-site casual users won't leave them at the drop of a hat even if their products are slightly worse than those of their competitors, while the API users will easily switch to them if their products become better. So apart from Google, OpenAI seems to be the best positioned for becoming able to build the largest training systems.
Frontier AI companies have seen rapid increases in the economic resources they have available to pursue AI progress. At some companies, the number of employees is at least doubling every year, and the amount of capital received is tripling every year. It is unclear whether this growth is sustainable. Is the revenue growing faster than the capital requirements? If it is, will revenue catch up before the capital available for AI investments runs out?
I do not think that there is enough publicly available data to answer these questions. It is plausible that frontier AI companies will run into significant economic limitations before Metaculus’s forecast for AGI in July 2033.
Similar Work
Epoch has an estimate for how various inputs to AI training runs could scale through 2030. Their work is distinct from this post because they focus on technical inputs (electric power, chip manufacturing, data, and latency), while this post focuses on economic inputs (labor and capital).
Epoch also has an estimate for revenue trends during 2023-2024. This post looks at a longer time scale and looks at other inputs in addition to revenue.
Parker Whitfill and Cheryl Wu have published an estimate for the number of employees at frontier companies over time.
Limitations
None of the frontier AI companies are independent, publicly traded companies. They are either independent companies that are not publicly traded (OpenAI, Anthropic, and xAI[1]), portions of a larger company (DeepMind and MetaAI), or not based somewhere with strong public reporting requirements (DeepSeek).
These companies are not required to make the data I am interested in publicly available. Instead, the information comes mostly from public statements from these companies. These are unlikely to be inaccurate (since then the companies could be sued for fraud), but they can be very imprecise or only selectively reported.
The quality of the data varies widely depending on what is being measured, and between the frontier AI companies. I think that enough of the data is high enough quality to draw some interesting conclusions from it.
A key assumption that I am making is that the amount of capital received is proportional to the current capital requirements at the frontier AI companies. Equivalently, these companies are spending most of their capital on furthering AI capabilities, rather than holding on to it for the future. If the amount of available capital slows down, this would translate into frontier AI companies spending fewer resources on developing AI.
I am not predicting what would happen if capital does run out. Does AI progress transition to a new exponential trend with a slower growth rate or is there an AI winter? Do the frontier AI companies continue unscathed or do some of them fail when the bubble pops? This post attempts to estimate whether the current trends are sustainable, not what would happen if they are not.
Employees
The growth rate for the number of employees at OpenAI is 2x per year, at Anthropic is 2.4x per year, and at DeepMind is 1.4x per year.
I do not predict that the growth in the number of employees will run into hard limits before 2033.
Sources
There are two main sources of data for the number of employees at the frontier AI companies.
The first source is when the companies publish the number of their employees themselves, or tell journalists who publish it. For example, during the conflict between OpenAI’s former Board of Directors and Sam Altman, newsorganizationsreported that over 700 out of OpenAI’s 770 employees threatened to leave with Sam Altman.
The other main source of data is from market research organizations like RocketReach, PitchBook, or LeadIQ. They regularly create estimates for the number of employees at lots of companies (among other data) and sell that data as part of their consulting service. The process by which they collect data is not always public, but it can include scraping LinkedIn and surveys sent to companies, in addition to the news reports. Most of my data comes from market research organizations through intermediaries who have made the data public. An example is the blog posts by SEO AI about OpenAI and Anthropic.
Some of these sources only say the month or year of their estimates, rather than the exact day. I have arbitrarily assigned these numbers to the middle of the year (July 1) and the beginning of the month. Changing when these are assigned can change the growth rate by 0.4x per year, with the strongest effect when half of the data is annual estimates.
Data
OpenAI has the most extensive data available, with multiple time series from market research organizations going back as far as 2015, in addition to a scattering of media or self-reported estimates. They currently have about 7400 employees, and are growing at a rate of 2.0x per year.
Anthropic has a similar density of data, but has existed for a much shorter amount of time. They currently have about 2300 employees, and are growing at a rate of 2.4x per year.
DeepMind has existed for even longer than OpenAI, but has sparser data. Half of my data points come from a single source,[2] with the others coming from media reports. They currently have about 6700 employees and are growing at a rate of 1.4x per year.
MetaAI, xAI, and DeepSeek do not have enough data for me to be comfortable estimating a trend. The most recent estimate for each of them has fewer employees than OpenAI, Anthropic, or DeepMind today.
Figure 1: Estimates for the number of employees at frontier AI companies vs time. All of the companies have some estimates, but only OpenAI, Anthropic, and DeepMind have enough data for me to include a trendline. DeepMind has long had the most employees, but they have recently been passed by OpenAI.
Projections
I will now recklessly extrapolate these trends forward in time.
The target dates for this extrapolation are: July 2027, when AI 2027 predicts that AGI will be announced, July 2030, a target date for an Epoch investigation, July 2033, when Metaculus predicts general AI, and July 2047, the aggregate forecast for 50% chance of HLMI from the 2023 AI Impacts survey.
In July 2027, these trends predict that OpenAI will have about 26,000 employees, and Anthropic and DeepMind will each have about 11,000 employees.
In July 2030, these trends predict that OpenAI will have about 210,000 employees, Anthropic will have about 160,000 employees, and DeepMind will have about 30,000 employees.
In July 2033, these trends predict that OpenAI will have 1.8 million employees, Anthropic will have 2.4 million employees, and DeepMind will have 80,000 employees. Changing the methodology can change these projections by roughly a factor of 2.
These projections are large, but not completely absurd. OpenAI & Anthropic would be larger than the largest software companies today: Microsoft at 228,000 and Google at 186,000 employees. They are similar to the largest companies today: Walmart at 2.10 million and Amazon at 1.56 million employees. The tech industry currently employs 6.4 million people, so these 3 companies would employ about 2/3 of the current US tech workforce.
While I expect that this growth will cause shortages of people with particular skills, I do not think that there are hard limits for frontier AI companies finding new employees before 2033.
In July 2047, these trends predict that OpenAI will employ 32 billion people, Anthropic will employ 600 billion people, and DeepMind will employ 8 million people. These numbers are ridiculous. If AGI is first developed in 2047, we should not expect that the number of employees at frontier AI companies will have continued to grow at the rate they are growing today.
Capital
Frontier AI companies are receiving a large and rapidly increasing amount of capital. OpenAI, Anthropic, and xAI have received tens of billions of dollars, and this is growing at between 2.2x and 3.5x per year.
The availability of capital could be a significant constraint for these companies in the near future.
Sources
The available sources for capital received are better than the sources for the number of employees, for some of the frontier AI companies.
Most of the funding rounds for the private companies have been made public. They are conveniently conglomerated by organizations like Tracxn. Funding rounds are also discrete events, so I do not have to worry about when an estimate was made. I am more confident in my estimates for capital trends for OpenAI, Anthropic, and xAI than in my estimates for employee trends.
DeepMind and MetaAI are parts of larger companies and have received most of their capital internally. These financial transfers are not publicly available, so I cannot do this analysis for them.
DeepSeek is in China, and so has different reporting requirements. A US Congressional report estimated that they received $420 million from the hedge fund that owns them, as of April 2025. I have found no other sources with data about DeepSeek’s capital.
Data
OpenAI has received about $62 billion in total capital. This number is growing at a rate of 2.2x per year. Their most recent funding round, $40 billion on March 31, 2025, looks to be above the trend. However, this capital has not yet all been received: $10 billion was received when this was announced, and $30 billion will be received by the end of the year.[3] If you split the capital up accordingly, it is on trend.
Anthropic has received about $30 billion in total capital. This number is growing at a rate of 3.5x per year. They also have the most recent funding round: $17 billion on September 2, 2025.
xAI has received about $22 billion in total capital. This number is growing at a rate of 3.3x per year. The trends for Anthropic and xAI look remarkably similar to each other.
Figure 2: Estimates for the total capital received by three of the frontier AI companies vs time. OpenAI has more capital, but Anthropic and xAI are growing faster.
Projections
I will now recklessly extrapolate these trends forward in time.
In July 2027, these trends predict that OpenAI will have received about $280 billion in capital, Anthropic will have received about $270 billion, and xAI will have received about $230 billion. The total amount of venture capital under management in the US is currently about $1.2 trillion. These three companies would account for over 60% of US venture capital. The current pool of venture capital is rapidly running out. Frontier AI companies have already begun looking for other sources of capital. They have sold conventional debt and are receiving investments from sovereign wealth funds.
In July 2030, these trends predict that OpenAI will have received about $3 trillion in capital, Anthropic will have received about $12 trillion, and xAI will have received about $8 trillion. For comparison, the total amount of capital in sovereign wealth funds is currently $14.3 trillion, the total size of the US bond market is currently $58.2 trillion, and the total size of the US stock market is currently $62.2 trillion. These three companies would account for more than all of the capital in sovereign wealth funds, or more than a third of the US bond market or the US stock market. Figuring out whether this is an unreasonable amount of capital would require trying to figure out how much of the US bond & stock markets are potentially divertible into AI companies, which is beyond the scope of this post.
In July 2033, these trends predict that OpenAI will have received about $34 trillion in capital, Anthropic will have received about $520 trillion, and xAI will have received about $310 trillion. For comparison, the total global wealth is about $500 trillion. These three companies would account for more than all of the current total global wealth.
There is no point in extrapolating these trends to July 2047 other than to gawk at how many zeros there are.
It seems important to emphasize that these are trends in inputs to AI, not outputs from AI. This is an estimate of the amount of capital received by frontier AI companies, which I assume tracks the amount spent to produce and deploy AI systems at scale.[4] It is not an estimate of the amount of value produced by AI systems.
In order for these trends to continue through 2033, most of the wealth required to continue AI development has to be newly generated. If frontier AI companies are not generating a significant fraction of global GDP,[5] they will have run out of money before 2033.Determining when exactly this will occur would require estimating how much of global wealth is available to be invested in AI, which is beyond the scope of this post. Most global wealth is not liquid or is otherwise unavailable for investing in AI. I turn instead to trends in the revenue generated by AI.
Revenue
The revenue data publicly available for most frontier AI companies is terrible. A growth rate of 3x per year is maybe reasonable.
Some frontier AI companies have also made public projections for their future revenue, and these projections are lower than what the trends would suggest.
Sources
None of the frontier AI companies are legally required to publish revenue data. OpenAI, Anthropic, and xAI are privately owned (not publicly traded on the stock market). DeepMind and MetaAI are both part of larger companies. DeepSeek is in China. These data are only public if companies choose to share them.
There are multiple ways of reporting revenue. I am focusing on annual revenue: the actual amount of revenue generated over the course of a year. To give it a specific date, I assign it at the middle of the year (July 1). Another common thing to report is annualized revenue: the revenue generated in a month (or quarter), multiplied by 12 (or 4). This can be helpful for tracking trends that might be changing rapidly. I am not using it because companies report annualized revenue for some months but not others, and I expect that there is selection bias.
OpenAI
OpenAI is the only frontier AI company that has published decent revenue data. Their nonprofit is legally required to publish their precise revenue, although that is a small part of their total organization. Their for profit company has also published annual revenue from 2020-2024, and has a reasonable looking projection for 2025.
OpenAI generated $3.7 billion in revenue in 2024 and projects $12.7 billion in 2025. It is growing at a rate of 3.2x per year.
Figure 3: Revenue of OpenAI’s nonprofit and for profit entities vs time. OpenAI’s projections for 2029 are substantially below the current trend.
OpenAI has made revenue projections for 2029: either $100 billion or $125 billion and it will be cash flow positive by then. These projections include a chart, which I have used to estimate their revenue projections for each year from 2026-2030.
This is a surprising projection. If the current trend continues, OpenAI would generate $1.4 trillion in revenue in 2029. This would also be the first year when projected revenue would cover projected capital requirements.
OpenAI is publicly projecting that its revenue growth will slow. They are projecting a revenue growth rate of 1.5x per year, not 3.2x per year.
Either OpenAI's projections are accurate and their revenue growth will be slower than the trends described above, or their public projections are inaccurate.
Accurate prediction:
In this case, OpenAI's current revenue growth will not continue. Since OpenAI is also claiming that it will be cash flow positive, it is predicting that the growth rate in capital it receives will slow.
This would have a significant impact on AI forecasts. In particular, manyforecastsinvolveextrapolating exponential trends.[6] If current trends rely on maintaining this exponential growth of inputs, then those forecasts seem dubious.
Inaccurate prediction:
OpenAI might also continue their current exponential growth rate (or grow faster). This is consistent with a qualitative prediction they made: that projected revenue would surpass projected capital requirements in 2029. It is inconsistent with their numerical predictions.
Anthropic
Anthropic’s data is terrible. I have found multiple sources with very different estimates. The revenue for 2024 seems to be between $200 million and $1 billion, and is growing at a rate of between 2.6x and 11.5x per year. This is not a small range. In particular, it includes the growth rate for capital received (3.5x per year), so it is unclear whether these projections suggest that Anthropic will ever be profitable.
Figure 4: Revenue of Anthropic vs time, according to three different sources. There is huge uncertainty.
Anthropic has also made projections for 2025 and 2027. Their 2025 projection is $2.2 billion in revenue. Their “base case” projection for 2027 is $12 billion, and they say that revenue “could reach as high as” $34.5 billion. If I assume exponential growth between 2025 and 2027, they are projecting an exponential growth rate of between 2.7x and 7.8x per year.
Other Frontier AI Companies
DeepMind has not released revenue data since 2020. That data included “how much Alphabet pays for internal services, and that can be completely arbitrary.”
I don’t know of any revenue data that has been released for MetaAI.
xAI has only released rapidly changing projections for annual revenue. In February of this year, they projected annual revenue for 2025 to be $100 million. In June, after generating $50 million in revenue in the first quarter, they projected annual revenue for 2025 to be $1 billion, and $14 billion in 2029.[7]
DeepSeek has released a theoretical calculation of what their revenue could be: $200 million, along with a theoretical cost-profit ratio of 545%.
I distrust all of these numbers.
Epoch’s Estimates
Epoch has also estimated the growth rate in revenue for the frontier AI companies. They focused on the years 2023-2024. They estimate that Anthropic and DeepMind[8] had revenue of single digit billions per year and OpenAI had revenue of about $10 billion per year, in April 2025. The growth rate for each of these three companies’ revenue is estimated to be about 3x per year. They also argue that no other AI company had revenue exceeding $100 million in 2024.
The biggest difference in methodology is that Epoch uses reported annualized revenue. This allows them to focus on a shorter time period, although I don’t think that there is a solution for the selection bias in reporting. DeepMind doesn’t even report annualized revenue, so Epoch created a proxy based on the number of users.
I think that Epoch’s estimates are reasonable, and that they adequately express uncertainty. A growth rate of 3x per year is consistent with my estimate for OpenAI, and is within my large uncertainty for Anthropic and DeepMind.
Conclusion
The economic resources being used for AI are growing rapidly.
The number of employees at OpenAI and Anthropic is at least doubling every year, and increasing by 40% every year at DeepMind. The amount of capital received by OpenAI is doubling every year, and more than tripling every year for Anthropic and xAI. The growth of revenue generated is highly uncertain, but might be tripling every year.
It is unclear how sustainable this is. Frontier AI companies could run out of US venture capital in mid 2027 and could run out of all global wealth in 2033. Substantial new wealth generated by AI is necessary to maintain exponential trends. Revenue might catch up to capital requirements for OpenAI in 2029, although they project a much lower revenue then. For Anthropic, neither Epoch’s revenue estimate nor their own baseline projections are growing as fast as their capital received (although other estimates exist). Other frontier AI companies have too little data to make the comparison.
Either frontier AI companies will have to generate more revenue than their own projections suggest, or the amount of capital available to be invested in AI will not be able to continue its current exponential growth for much longer.
Thank you to the many people who volunteered as mentors for scholars at MATS! We would also like to thank our 2025 donors, without whom MATS would not be possible.
If you are interested in becoming a MATS scholar, applications for the Winter 2026 cohort are now open until October 2.
xAI merged with the social media company X (formerly Twitter) in March 2025, so it could also be considered to be in the second category. In that merger, xAI had a higher valuation.
Epoch estimates that frontier AI companies should spend roughly equal amounts on training and inference. If frontier AI companies are using most of the capital they acquire on training and inference, then each would account for somewhat less than half of these companies’ current budget.
The annual capital requirements for these three companies would be greater than the current global GDP of $111 trillion. I do not know what the rest of the economy would look like in order for this to work.
This estimate excludes revenue gained by integration of DeepMind’s products within other Google products, like including Gemini with Google search results.
