Help World Market Atlas grow
Ordering a company valuation also supports the project’s development.
Support the project
Help develop World Market Atlas — an independent project for individual investors. Your support helps us produce analyses and improve our tools.
OpenAI and Anthropic’s losses: the economics of the frontier-model race and the challenge of comparing financial data
Data as of 23 September 2026. Media-reported figures, historical results and forecasts are distinguished throughout the text.
Abstract
This article examines operating losses, cash consumption and future infrastructure commitments at two leading private companies developing frontier artificial intelligence models: OpenAI and Anthropic. The analysis shows that there is no single, fully comparable and publicly verified figure defining each company’s “loss”. Available sources use different measures: operating loss, net loss, negative free cash flow, cash burn and adjusted operating income.
According to shareholder documents described by The Information, OpenAI reportedly burned $3.7 billion in cash in the first quarter of 2026 while generating $5.7 billion in revenue. Reuters noted that it could not independently confirm these figures. A forecast reported by the Financial Times presents an even larger scale: OpenAI reportedly expects cumulative negative free cash flow of $278 billion in 2026–2030. This is a forecast, however, not a loss already incurred. Reuters
For Anthropic, media sources indicate approximately $5.6 billion of cash burn in 2024 and a planned reduction to $3 billion in 2025. In 2026, Anthropic reportedly forecast its first positive adjusted operating result. This does not automatically imply positive net income or positive free cash flow. Reuters
Keywords: OpenAI, Anthropic, artificial intelligence, cash burn, operating loss, free cash flow, compute, frontier models.
1. Introduction
The development of generative artificial intelligence has produced companies with exceptionally fast revenue growth, but also an unprecedented need for investment in data centres, processors, energy, transmission networks, cloud infrastructure and highly skilled employees. OpenAI and Anthropic are among the leading participants in this race.
Assessing their financial position is more difficult than analysing traditional listed companies. Both remain private businesses and do not publicly release a complete, standardised set of consolidated financial statements comparable to the annual reports of US-listed companies. Most information about their losses therefore comes from investor documents, shareholder presentations, material obtained by the media or statements by management.
In several cases, Reuters explicitly stated that it could not independently confirm figures first published by The Information or the Financial Times. A methodologically sound formulation is therefore not “OpenAI definitely lost a particular amount”, but “according to documents described by a particular outlet, the company reportedly recorded a particular result”. Reuters
2. Research method and source hierarchy
The article uses documentary analysis. Sources are divided into three groups:
Unsigned social-media posts, unverified online compilations and figures for which an original source could not be established were excluded from the analysis.
3. “Loss” is not a single concept
Discussions of OpenAI and Anthropic frequently mix several different financial measures.
Operating loss arises when the costs of a company’s core business exceed its revenue. These costs include infrastructure, research and development, employee compensation, sales and administration.
Net loss also incorporates non-operating items, such as financing costs, changes in the value of equity instruments, taxes or one-off restructuring effects. It can therefore be substantially higher or lower than operating loss.
Cash burn indicates the reduction in a company’s cash resources associated with funding operations and investment. It is not the same as an accounting loss.
Negative free cash flow means that cash flows generated by the business are insufficient to finance its investments. For AI laboratories, both ongoing computing costs and multiyear agreements for data-centre capacity may be important.
Adjusted operating income may exclude certain expenses, such as stock-based compensation. A positive adjusted operating result is therefore not equivalent to positive net income.
Another issue is the use of run-rate revenue. This extrapolates the latest pace of sales rather than measuring revenue earned over a full twelve months. Anthropic defined this metric by multiplying consumption-based sales over the previous 28 days by 13 and monthly subscription revenue by 12. According to Reuters Breakingviews, Anthropic’s GAAP revenue from inception through the end of 2025 exceeded $5 billion, while its reported run-rate had already reached $19 billion in February 2026. Both figures may be true, but they measure entirely different things. Reuters
4. OpenAI — revenue growth with very high capital consumption
4.1. First-quarter 2026 figures
According to shareholder documents described by The Information, OpenAI reportedly:
Reuters published this information on 16 June 2026, while noting that it could not independently confirm the report. Reuters
The ratio of cash burn to revenue was approximately:
(3.7 ÷ 5.7) × 100% ≈ 64.9%
Reported cash burn thus corresponded to about 65 cents for every dollar of quarterly revenue. This ratio should not be interpreted as an operating margin: cash burn may include investment, payments arising from earlier commitments and changes in working capital.
4.2. Non-public figures for 2025
In June 2026, Fortune described copies of OpenAI financial statements obtained by journalist Ed Zitron and the Financial Times. According to these materials, the company’s 2025 revenue was reportedly about $13.07 billion, costs and expenses were $34 billion, and operating loss was approximately $20.92 billion. Fortune
These figures are important, but they should not be presented as a publicly confirmed, audited result available on the same basis as a listed company’s report. They come from copies of documents obtained by the media. In this article, they therefore serve as supporting material rather than an undisputed public financial statement.
If the figures are accurate, operating loss would correspond to approximately 160% of revenue, and total costs would be about 2.6 times sales. This scale suggests that rapid revenue growth was not yet sufficient to cover research, infrastructure, model-serving, distribution and organisational expansion costs.
4.3. Forecast through 2030
The most significant figure concerning OpenAI’s future funding needs comes from a company presentation reportedly reviewed by the Financial Times. According to Reuters, OpenAI forecasts:
All these figures are forecasts from an internal presentation, not realised results. OpenAI did not directly comment on the report before Reuters published it. Reuters
The plan would imply that, even with very rapid sales growth, the company anticipates a multiyear period of negative cash flow. In this respect, OpenAI’s strategy resembles infrastructure development: the company secures computing capacity in advance, before revenue generated by those assets fully covers their costs.
4.4. Financing losses
On 31 March 2026, OpenAI officially announced that it had raised $122 billion in committed capital at a post-money valuation of $852 billion. The company also stated that it was generating approximately $2 billion in monthly revenue. This information comes directly from the business, but it does not confirm profitability. OpenAI
Capital raised from investors is neither sales revenue nor profit. It helps cover negative cash flows, fund research and enter agreements for future supplies of computing capacity.
5. Anthropic — from high cash burn to reported adjusted profitability
5.1. Cash consumption in 2024–2025
In February 2025, Reuters, citing The Information, reported that Anthropic:
Anthropic did not respond to Reuters’ request for comment at the time. Crucially, the $3 billion figure was a forecast for 2025, not a subsequently confirmed annual result. Reuters
It is therefore inappropriate to write that “Anthropic lost $3 billion in 2025”. The correct formulation is that, in February 2025, the company reportedly projected approximately $3 billion of cash burn for the full year.
5.2. First reported forecast of a positive operating result
In May 2026, Reuters Breakingviews reported that Anthropic generated $4.8 billion in first-quarter revenue and expected $10.9 billion in the second quarter. The company also reportedly forecast $559 million in adjusted operating profit for that period, which would make it the first profitable quarter under this measure. Reuters
The projected adjusted operating margin was approximately:
(0.559 ÷ 10.9) × 100% ≈ 5.1%
According to Reuters, this measure included the full cost of training new models but excluded stock-based compensation. It was therefore not equivalent to unadjusted operating profit or net income. Reuters
In September 2026, the Financial Times reported that Anthropic had reportedly given shareholders a forecast of positive adjusted operating income for a second consecutive quarter. Reuters again noted that it could not independently verify the information, and Anthropic did not respond before publication. Reuters
These forecasts suggest a possible improvement in Anthropic’s operating economics, but they do not yet demonstrate that the company has achieved:
5.3. Why Anthropic and OpenAI cannot be compared without adjustments
Reuters Breakingviews noted that Anthropic’s revenue was not reduced by the share owed to distribution partners. OpenAI, by contrast, reportedly presented revenue after deducting Microsoft’s share. Consequently, $1 of revenue reported by Anthropic does not necessarily correspond to $1 reported by OpenAI. Reuters
A similar problem arises on the cost side. Anthropic’s adjusted operating result excluded stock-based compensation. For a rapidly expanding technology company, such compensation can be economically significant even when it does not cause an immediate cash outflow.
Available figures therefore do not support a responsible conclusion that Anthropic is categorically more profitable than OpenAI. One can only say that Anthropic reported a faster path to positive adjusted operating income.
5.4. Anthropic’s financing and infrastructure commitments
In February 2026, Anthropic officially announced a $30 billion funding round at a post-money valuation of $380 billion. It reported run-rate revenue of $14 billion at the time. Anthropic
In April 2026, Anthropic announced an agreement with Amazon involving a commitment to spend more than $100 billion on AWS technologies over ten years and secure up to 5 GW of computing capacity. Amazon simultaneously committed to invest $5 billion, with the possibility of investing another $20 billion, on top of its earlier $8 billion. Anthropic
The $100 billion is not Anthropic’s current loss. It is a multiyear contractual commitment to be settled in the future. It nevertheless shows that even a company reporting positive adjusted operating income must secure enormous amounts of capital for future infrastructure.
6. Summary of available figures
The table shows that the available figures do not form a consistent comparative series. Some entries represent historical results, others forecasts, while definitions of revenue and profitability also differ.
7. Main drivers of losses
7.1. Frontier-model training costs
Research by Ben Cottier and co-authors indicates that the amortised training cost of the most compute-intensive frontier models increased by an average of approximately 2.4 times per year from 2016. The authors estimated that computing hardware accounted for approximately 47–64% of model development costs in the projects studied, while research and development staff — including equity compensation — accounted for approximately 29–49%. Energy itself represented a much smaller share, usually 2–6%. arXiv
The study does not present OpenAI’s or Anthropic’s actual accounts. It is a cost model based on hardware, energy, cloud and staffing data. Its significance lies in showing that training costs extend well beyond electricity purchases. The most important categories are processors, servers, networks, infrastructure depreciation and the work of highly paid research teams.
7.2. The cost of serving users
After training is complete, a model continues to incur costs for every user query. This inference cost includes processors, memory, networks, data centres and energy.
The Stanford AI Index reported that the inference cost of a system achieving roughly GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024. Hardware costs fell by an average of approximately 30% per year, while energy efficiency improved by approximately 40% annually. Stanford HAI
Lower unit costs do not necessarily produce lower total expenditure. If user numbers, response lengths, agent usage and demand for more advanced models grow faster than hardware efficiency, the total inference bill may continue to increase.
7.3. The race for computing capacity
AI laboratories secure computing capacity years in advance. This reflects limited availability of advanced accelerators, transformers, grid connections, cooling systems and suitably equipped data centres.
OpenAI’s projected infrastructure expenditure of $856 billion through the end of 2030 and Anthropic’s commitment to spend more than $100 billion on AWS technologies over ten years show that frontier-model development is beginning to resemble an infrastructure business rather than software development alone. Reuters
7.4. Research costs and rapid model obsolescence
A new model may require months of research, experimentation and safety testing. Its technological advantage may nevertheless erode quickly when competitors introduce their next generation of models.
This causes accelerated economic depreciation of research and development expenditure. A model may remain technically functional but require a price reduction or replacement by a newer system. This risk increases the pressure to keep incurring further training costs.
8. Discussion: are losses a sign of weakness?
Large losses do not automatically mean that a company’s business model is ineffective. During rapid expansion, a business may deliberately accept negative cash flow to acquire customers, develop technology and secure future infrastructure.
At the same time, the ability to raise enormous amounts of capital does not demonstrate sustainable profitability. Investor funding merely postpones the point at which revenue must begin to cover:
inference costs + partners’ shares + research and training + employee compensation + infrastructure + other costs
The greatest risk is a situation in which model prices fall faster than training and serving costs. Competition between OpenAI, Anthropic, Google, Meta, xAI and open-model providers may exert downward pressure on prices. In such an environment, rapid user growth does not necessarily translate into proportionate margin growth.
Conversely, increasing automation of programming, document analysis, customer service and administrative processes may generate far more valuable revenue than simple chatbot interactions. Economically, the key will therefore be not token volume alone, but the ability to charge for solving a specific business problem.
9. Research limitations
The principal limitation is the absence of complete, publicly available financial statements prepared on a consistent basis for both companies. Four problems follow.
First, some figures come from internal documents described by the media. Second, historical results are sometimes compared with future forecasts. Third, revenue definitions differ between companies. Fourth, adjusted results may exclude economically significant items such as stock-based compensation or distribution partners’ shares. Reuters
The available material therefore does not allow fully comparable EBITDA, EBIT, net-income or free-cash-flow margins to be calculated for the two companies.
10. Conclusions
OpenAI remains a company with exceptionally large capital requirements. The latest report concerning the first quarter of 2026 indicates $3.7 billion in cash burn against $5.7 billion in revenue, but Reuters did not independently confirm those figures. The multiyear forecast of $278 billion in negative free cash flow suggests that management expects aggressive infrastructure financing to continue at least through the end of the decade. Reuters
Anthropic also incurred very high costs, reflected in reported cash burn of approximately $5.6 billion in 2024. The 2026 figures, however, suggest a transition to positive adjusted operating income. This is not equivalent to sustained net profit: the measure excluded stock-based compensation, among other items, and the treatment of partners’ shares differed from OpenAI’s approach. Reuters
The available evidence supports three main conclusions:
Both companies’ losses should be understood as part of a wider race to control computing capacity. In this model, compute is becoming not merely a current expense but a strategic productive resource. The ultimate economic value of OpenAI and Anthropic will depend on whether revenue generated by infrastructure secured today exceeds the future cost of financing, maintaining and continually upgrading it.
Bibliography
Posts reflect the author’s personal views as of publication. They are not investment recommendations.