In June 2026, a leaked OpenAI financial document sent shockwaves through the tech world. The document showed that OpenAI's revenue reached $13.07 billion in 2025, a staggering 253% increase compared to $3.7 billion in 2024. However, alongside the soaring revenue was an operating loss of up to $20.92 billion, with a net loss of approximately $8 billion.

Beneath the prosperous surface of ChatGPT surpassing 900 million weekly active users and a company valuation of $852 billion, OpenAI's books reveal a harsh reality: in 2025, for every dollar the company earned, it spent $1.60. Is this "burn cash for scale" model a unique growing pain on OpenAI's path to Artificial General Intelligence (AGI), or a common affliction of the entire large model industry? By dissecting its cost structure and conducting a horizontal comparison with the financial data of leading companies like Anthropic and xAI, we may be able to see the true cost behind the current AI industry boom.

The Cost Black Hole Behind $13 Billion in Revenue: Where Did the Money Go?

To understand OpenAI's loss logic, one must first dissect the composition of its $34 billion in total costs and expenses. In this leaked financial document, the largest expense item was R&D costs, reaching $19.18 billion, which included $10.59 billion paid to Microsoft. This was followed by $7.5 billion in cost of revenue (primarily for inference computing) and $5.73 billion in sales and marketing expenditures.

In terms of growth rate, OpenAI's cash-burning efficiency has actually improved. In 2024, the company needed to spend $2.37 for every $1 of revenue generated, and by 2025, this figure dropped to $1.60. Revenue growth (253%) outpaced total cost growth (172%). But this does not mean cost pressure has eased; on the contrary, the ticket price for the scaling law is still rising sharply.

The $19.18 billion in R&D expenditure accounted for as much as 147% of its annual revenue. In the large model field, R&D means not only the salaries of algorithm engineers but also massive consumption of training compute power. To maintain a lead in model capabilities, OpenAI must continuously invest huge sums in training the next generation of models. This investment is rigid; once slowed, it could lose its position in the competition with rivals.

The $7.5 billion in inference computing costs cannot be ignored either. This part of the cost is directly linked to user usage. ChatGPT's weekly active users exceeding 900 million means a massive influx of inference requests hits OpenAI's servers every day. Every conversation, every generation, consumes real computing resources. Although hardware performance is improving, user demand for more complex, longer-context interactions is growing faster, causing the absolute value of inference costs to continue climbing.

Furthermore, the $5.73 billion in sales and marketing spending also reflects the high cost for AI companies in consumer acquisition and enterprise expansion. At a time when product homogenization trends are emerging, maintaining brand presence and capturing enterprise customer market share requires substantial financial investment.

What needs special clarification is the caliber of the net loss. The leaked documents show that the 2025 net loss included approximately $30 billion in one-time non-cash accounting charges, stemming from changes in the fair value of convertible interests and warrant liabilities when OpenAI converted from a non-profit structure to a Public Benefit Corporation (PBC). Excluding this one-time factor, the actual operating loss was approximately $20.92 billion, with a net loss of about $8 billion. This distinction is crucial because it strips away the book fluctuations caused by financial structure changes and restores the true consumption of the company's daily operations.

A $17.2 Billion Structural Burden: Microsoft's "Invisible Cut"

In OpenAI's cost structure, there is an unavoidable behemoth: Microsoft. According to the leaked documents, OpenAI's total payments to Microsoft in 2025 amounted to $17.2 billion, including $10.59 billion in R&D expenditure, $6.047 billion in cost of revenue, $527 million in sales expenditure, and $42 million in administrative expenditure.

This $17.2 billion payment accounted for 50.5% of OpenAI's total annual costs, even exceeding its $13.07 billion annual revenue. Microsoft is not only OpenAI's cloud service provider but also an "invisible shareholder" deeply binding OpenAI's cash flow through compute revenue sharing. In the early cooperation, Microsoft's computing support was key to OpenAI's rapid rise. But as OpenAI's business scale expanded, this revenue-sharing model became a heavy structural burden.

According to previously disclosed cooperation agreements, OpenAI must pay Microsoft a 20% revenue share, lasting until 2030. This means that as long as OpenAI uses Microsoft's Azure cloud services for training and inference, this expenditure will follow like a shadow. Before achieving positive cash flow, OpenAI must first fill Microsoft's compute bill. This structure also explains why OpenAI needed to complete a massive $122 billion financing round in March 2026. Unable to rely on its own cash generation, external blood transfusion is the only way to sustain operations.

Cash-Burning Efficiency Rankings: OpenAI vs Anthropic vs xAI

Is high R&D spending and high losses a phenomenon unique to OpenAI? Turning our gaze to two other leading AI companies, the answer is no.

According to the IPO S-1 filing submitted by SpaceX, Musk's xAI had revenue of $3.2 billion in 2025, but an operating loss of up to $6.4 billion, with capital expenditures reaching $12.7 billion. Calculating cash-burning efficiency, xAI needs to spend $3 for every $1 earned, with a loss/revenue ratio as high as 200%, far exceeding OpenAI's 160%. To bet on trillion-parameter models, xAI built the Colossus data center in just 122 days, with its capital expenditure even exceeding the combined capital expenditure of SpaceX's Starlink and rocket businesses. This indicates that on the track of pursuing the scaling law, xAI has adopted a heavier asset bet than OpenAI.

The situation for another major competitor, Anthropic, presents a different path. According to official announcements, Anthropic's annualized revenue (ARR) reached $9 billion at the end of 2025 and soared to $47 billion in May 2026. Its core growth engine, Claude Code, had an annualized revenue exceeding $2.5 billion in February 2026.

However, behind the high-speed growth also lies cost pressure. According to The Information, Anthropic's gross margin in 2025 was only 40%, 10 percentage points lower than expected, because inference costs were 23% higher than anticipated. In terms of losses, media reports indicate its EBITDA loss magnitude is also in the billions of dollars. Due to the lack of exact audit documents, we cannot know Anthropic's actual total net loss, but the 40% gross margin and higher-than-expected inference costs expose the same industry-wide common pressure.

Comparing the data of the three companies side by side reveals: in 2025, the combined operating losses of OpenAI, xAI, and Anthropic exceeded $30 billion. Burning cash for scale is not an isolated case but the norm in current large model competition. The difference lies in the choice of business path. Anthropic does not build its own data centers, relying on a multi-cloud strategy of AWS, Google, and Azure, taking an asset-light route, and achieving high-premium monetization on the enterprise side through Claude Code; xAI firmly holds computing infrastructure in its own hands, betting on computing power monopoly; OpenAI falls somewhere in between, relying on Microsoft's computing power while possessing a massive consumer user base.

900 Million Weekly Active Users and a 5.6% Conversion Rate: A Stress Test of the Monetization Ceiling

The massive user base is OpenAI's core moat and an important support for its $852 billion valuation. But the financial data reveals the other side of this moat.

Among ChatGPT's 900 million weekly active users, paying users are approximately 50 million, a conversion rate of about 5.6%. Roughly estimated based on $13.07 billion in revenue, the annual revenue contribution per paying user (ARPU) is about $261. This means that over 800 million free users are consuming computing power without generating direct revenue.

Against the backdrop of persistently high inference costs, the computing power consumption of free users has become a huge burden. How to improve conversion rates and ARPU is a direct challenge facing OpenAI. This pressure is more apparent when comparing Anthropic's strategy. Facing cost pressure, Anthropic chose to double the pricing for its top-tier model API, launching tiered pricing strategies like Claude Fable, turning cutting-edge AI capabilities into "luxury goods" to screen for high-net-worth enterprise customers.

OpenAI, however, currently maintains a basic subscription model of $20 per month. This model helps rapidly expand the user base during the expansion phase, but in a stage where the cost structure needs optimization, it inevitably faces pressure to raise prices or further tiered pricing.

Who Pays the Bill for the Scaling Law?

OpenAI's leaked ledger tears open a corner of the AI industry's glossy exterior. Annual revenue of tens of billions yet a net loss of eight billion is not only OpenAI's current situation but also a common dilemma faced by leading companies like xAI and Anthropic. High R&D investment and high inference costs constitute the two major mountains in large model competition.

Massive financing provides a cushion for this cash-burning model. The $122 billion financing completed by OpenAI in March 2026, and Anthropic reaching a valuation of $965 billion in May of the same year, both indicate that the capital market is currently still willing to pay for the scaling law. But capital's patience is limited.

Whether AI companies can escape the quagmire of losses depends on whether they can achieve a sudden drop in marginal costs. Early SpaceX reduced launch costs by over 90% through rocket reuse, thereby changing the economics of the aerospace industry. Whether the AI industry can replicate this path depends on whether inference computing costs can be significantly reduced through specialized chips, model compression, or architectural innovation. Until then, high R&D spending and high losses will remain the main theme of the AI industry. What determines whether AI tools can continue to iterate is not the stunning nature of the algorithms, but the cost structure hidden in the ledger.