Leaked financial documents show OpenAI is losing billions of dollars a year
Leaked internal documents indicate OpenAI is on track to lose around $38.5 billion a year at its current run rate. The figures reviewed by Ars Technica show that despite strong revenue growth, the costs of running the AI industry's flagship company are scaling far faster than its income.
On the revenue side, ChatGPT subscription revenue is approaching the $12 billion band. API revenue adds roughly $5 billion more. Total annual revenue sits in the $17-18 billion range.
The cost side is heavier still. Compute infrastructure — anchored by Microsoft Azure, with follow-on contracts at SpaceX, Oracle and Google — clears $45 billion a year. That is nearly three times revenue.
The secondary load is research staffing and account-security operations. Training-time compute and data acquisition budgets sit inside this cluster. OpenAI is reported to be budgeting $1-3 billion per model in recent training runs.
The cash-flow picture is sharply negative. With the funding raised in its latest investor round, the company appears to have a 3-4 year runway — assuming the cost curve does not bend down meaningfully.
The strategy debate at this point splits two ways. In one camp, OpenAI bets on the so-called scaling laws: revenue will keep growing in proportion to GPU capacity and margins will turn positive within the next two years. In the other, critics argue that marginal gains in model quality are showing diminishing returns while costs keep climbing faster.
Price sensitivity among enterprise customers has risen. TechCrunch and The Information have reported that large SaaS firms are renegotiating OpenAI contracts by floating counter-bids from Anthropic Claude or Google Gemini. That pressure is squeezing margin.
SPAC or IPO routes are on the table. A side note in the leak points to banker discussions firming up for the first half of 2027. A reported valuation above $500 billion is a tough argument for the IPO of a company not yet profitable.
The Microsoft front has also tightened. Limited Azure capacity, and Microsoft's increased investment over the past six months in its own MAI models, has turned the OpenAI-Microsoft relationship into a point of strategic tension. The leaked documents show contract renegotiations started late.
Ultimately, the leak doubles as a heat map for the AI industry as a whole. If the market leader is burning $38 billion a year, how do smaller players stay sustainable? For investors, this may be the moment to revisit financial models — because revenue growth glitters, but the balance sheet is fragile.
What will resolve the picture is not the headline loss number but how compute pricing evolves over the next 18 months. If GPU costs drop the way they did for storage in the late 2000s, the curve bends; if they stay flat, the cash burn keeps compounding.
Read next
Why is DuckDB so fast? Columnar storage, vectorised execution and a modern analytical engine design
The open-source analytical database DuckDB has become a standard component in data-science tool stacks over the past three years. A new technical post highlighted on Hacker News walks through the three core design choices behind the engine's speed — columnar storage, vectorised execution and zero-copy integration.
Android developer verification is coming: Google's why, when and which app stores guide
Google has set out the timeline and scope for mandatory Android developer verification in a new planning document covered by Ars Technica. Certification will extend beyond the Play Store to the other major third-party stores and will introduce a new compliance framework for sideloaded APKs.
Microsoft uncovers a self-propagating backdoor that steals cryptocurrency: how the threat works
Microsoft's Threat Intelligence team has documented a lightweight, self-propagating backdoor designed to steal wallet keys. According to Ars Technica, the malware spreads as a compressed payload, mimics system processes to bypass existing security tools and systematically scans the typical file paths where crypto wallets are stored.
Amazon to sell its AI chips outside AWS in a more direct challenge to Nvidia
According to TechCrunch, Amazon is preparing to sell the Trainium and Inferentia chips it designs and manufactures inside AWS directly to enterprise customers as well. The move is being read as the first concrete enterprise-grade alternative to Nvidia's Hopper/Blackwell line in market share.
The smartphone era created an attention crisis: how the slow-tech movement is responding
The steady stream of notifications that smartphone economics built into daily life has spawned a debate stretching from academic research to product design. TechCrunch reports that a new movement known as "slow tech" is trying to make conscious, low-dose use — rather than constant attention disruption — the default design frame.
