In June 2026, a leaked financial document from OpenAI caused a major stir in the tech circle. The document revealed that OpenAI's revenue reached $13.07 billion in 2025, representing an astonishing growth of 253% compared to $3.7 billion in 2024. However, alongside the revenue surge was an operational loss of $20.92 billion, with a net loss of approximately $8 billion.
Beneath the surface of prosperity, where ChatGPT's weekly active users exceeded 900 million and the company's valuation reached $852 billion, OpenAI's financial records revealed a harsh reality: in 2025, for every dollar the company earned, it spent $1.60. Is this "burning money for scale" model a unique growing pain for OpenAI on the road to Artificial General Intelligence (AGI), or is it a common ailment in the entire large model industry? By dissecting its cost structure and comparing it with financial data from leading companies like Anthropic and xAI, we may be able to uncover the true costs behind the current AI industry's prosperity.
The Cost Black Hole Behind $13 Billion Revenue: Where Did the Money Go?
To understand the logic behind OpenAI's losses, it's necessary to break down its total costs and expenses of $34 billion. In the leaked financial document, the largest expenditure was research and development costs, reaching $19.18 billion, which includes $10.59 billion paid to Microsoft. Following that were $7.5 billion in revenue costs (mainly for inference computation) and $5.73 billion in sales and marketing expenses.
In terms of growth rate, OpenAI's burning money efficiency has actually improved. In 2024, the company had to spend $2.37 for every dollar of revenue generated, while in 2025, this number dropped to $1.60. The revenue growth rate (253%) outpaced the total cost growth rate (172%). But this does not mean that cost pressures have eased; on the contrary, the ticket price of economies of scale is still rising sharply.
The $19.18 billion in research and development spending accounted for a staggering 147% of its annual revenue. In the large model field, R&D not only means wages for algorithm engineers but also substantial training computing resource consumption. To maintain its lead in model capabilities, OpenAI must continuously invest heavily in training the next generation of models. This investment is rigid, and if it slows down, the company risks losing its position against competitors.
The $7.5 billion in inference computation costs is also noteworthy. This cost is directly linked to user usage. With ChatGPT's weekly active users surpassing 900 million, it means a massive influx of inference requests directed at OpenAI's servers daily. Every conversation, every generation consumes real computational resources. Despite improvements in hardware performance, the demand for more complex and longer context interactions from users is growing even faster, leading to rising absolute values in inference costs.
Furthermore, the $5.73 billion in sales and marketing expenses also reflects the high cost of customer acquisition on the consumer side and expansion on the enterprise side for AI companies. In an environment where product homogeneity is becoming evident, maintaining brand presence and capturing enterprise client share requires substantial investment.
It is essential to clarify the metrics surrounding net losses. The leaked document shows that the net loss of approximately $8 billion in 2025 included about $30 billion in one-time non-cash accounting expenses, stemming from the fair value changes of convertible equity and warrant liabilities when OpenAI transitioned from a non-profit structure to a Public Benefit Corporation (PBC). Excluding this one-time factor, the actual operational loss was around $20.92 billion, with a net loss of about $8 billion. This distinction is crucial as it strips away accounting fluctuations caused by changes in financial structure and reveals the true consumption of the company's daily operations.
$17.2 Billion Structural Burden: Microsoft's "Invisible Cut"
Within OpenAI's cost structure, there is a colossal entity that cannot be avoided: Microsoft. According to the leaked document, OpenAI paid Microsoft a total of $17.2 billion in 2025, including $10.59 billion in research and development expenses, $6.047 billion in revenue costs, $527 million in sales expenses, and $42 million in administrative expenses.
This payment of $17.2 billion accounted for 50.5% of OpenAI's total annual costs, even exceeding its total annual revenue of $13.07 billion. Microsoft is not only OpenAI's cloud service provider but also an "invisible shareholder" deeply tied to OpenAI's cash flow through its computing power sharing. In the early stages of their collaboration, Microsoft's computing power support was key to OpenAI's rapid rise. However, as OpenAI's business scale expanded, this sharing model became a heavy structural burden.
According to a previously disclosed cooperation agreement, OpenAI is required to pay Microsoft a 20% revenue share, lasting until 2030. This means that as long as OpenAI continues to use Microsoft's Azure cloud services for training and inference, this expense will always be present. Before achieving positive cash flow, OpenAI must first pay Microsoft’s computing bills. This structure also explains why OpenAI needed to complete a massive $122 billion funding round in March 2026. In the absence of self-sustaining capabilities, external funding is the only way to maintain operations.
Burning Money Efficiency Rankings: OpenAI vs Anthropic vs xAI
Is high R&D investment and high losses a unique phenomenon for OpenAI? Turning to two other leading AI companies, the answer is no.
According to the IPO S-1 document filed by SpaceX, Elon Musk's xAI reported revenue of $3.2 billion in 2025 but incurred an operating loss of $6.4 billion, with capital expenditures reaching $12.7 billion. If we calculate the burning money efficiency, xAI spends $3 for every dollar earned, with a loss/revenue ratio of 200%, much higher than OpenAI's 160%. To bet on trillion-parameter models, xAI built the Colossus data center in just 122 days, with capital expenditures exceeding the total capital expenditures for SpaceX's Starlink and rocket businesses. This indicates that in the race for economies of scale, xAI has taken a more extreme asset-heavy bet than OpenAI.
Another major competitor, Anthropic, presents a different path. According to an official announcement, Anthropic's annualized revenue (ARR) reached $9 billion by the end of 2025 and soared to $47 billion by May 2026. Its core growth engine, Claude Code, had an annualized revenue exceeding $2.5 billion as of February 2026.
However, behind the rapid growth lies cost pressure. According to The Information, Anthropic's gross margin in 2025 was only 40%, 10 percentage points lower than expected, due to inference costs being 23% higher than anticipated. Regarding losses, media reports suggest that its EBITDA losses also reached billions of dollars. Due to the lack of precise audit documents, we cannot ascertain the actual total net loss for Anthropic, but the 40% gross margin and unexpected increases in inference costs reveal similar industry-wide pressures.
Comparing the data of the three companies shows that in 2025, the combined operational losses of OpenAI, xAI, and Anthropic exceeded $30 billion. Burning money for scale is not an isolated incident, but the norm in the current large model competition. The difference lies in the choice of business path. Anthropic relies on a multi-cloud strategy by utilizing AWS, Google, and Azure without building its own data centers, taking a light asset approach, and monetizing through Claude Code at a high premium on the enterprise side; xAI firmly controls computing power infrastructure, betting on computing power monopoly; while OpenAI occupies a middle ground, relying on Microsoft's computing power while having a large consumer user base.
900 Million Weekly Active Users and 5.6% Conversion Rate: A Pressure Test for Monetization Limits
The vast user base is OpenAI's core moat, and also an important support for its $852 billion valuation. However, financial data reveals another side of this moat.
Out of ChatGPT's 900 million weekly active users, the paid users number around 50 million, resulting in a conversion rate of approximately 5.6%. With an estimated revenue of $13.07 billion, the annual contribution of a single paid user (ARPU) is about $261. This means that over 800 million free users are consuming computational power without generating direct revenue.
In the context of persistent high inference costs, the computational power consumption of free users has become a significant burden. How to improve conversion rates and ARPU is the direct challenge faced by OpenAI. Compared to Anthropic's strategy, this pressure is even more apparent. In response to cost pressure, Anthropic chose to double the pricing on its top model APIs, introducing tiered charging strategies like Claude Fable, turning top AI capabilities into "luxury goods" to filter high-net-worth enterprise clients.
Meanwhile, OpenAI currently maintains a basic subscription model at $20 per month. While this model has aided rapid user base expansion during the user growth phase, during the stage requiring cost structure optimization, it inevitably faces pressure to raise prices or implement further tiered charging.
Who Will Pay the Bill for Economies of Scale?
This leaked financial record from OpenAI has peeled back a corner of the shiny facade of the AI industry. Earning billions yet losing billions is not only OpenAI's current state but also a common predicament faced by leading companies such as xAI and Anthropic. High R&D investments and high inference costs constitute two significant mountains in large model competition.
Massive financing provides a cushion for this money-burning model. OpenAI's $122 billion financing completed in March 2026, along with Anthropic's valuation reaching $965 billion in May of the same year, indicates that the capital market is still willing to pay for economies of scale. However, the patience of capital is limited.
Whether AI companies can break free from the mire of losses depends on whether they can achieve a steep reduction in marginal costs. Early SpaceX reduced launch costs by over 90% through rocket reusability, thereby changing the economics of the aerospace industry. Whether the AI industry can replicate this path depends on whether the costs of inference computing power can significantly drop through specialized chips, model compression, or architectural innovation. Until then, high R&D and high losses will remain the main melody of the AI industry. What determines whether AI tools can continue to iterate is not the brilliance of algorithms, but the cost structure hidden in the financial records.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
