OpenAI Expects to Burn $278 Billion Over Five Years, with Computing Power and Infrastructure as the Largest Single Item, Total Estimated Expenditure Around $856 Billion
Internal presentation documents from OpenAI indicate that the cumulative negative free cash flow from 2026 to 2030 is expected to reach $278 billion. Revenue is projected to rise from $36 billion this year to $350 billion by 2030, with a total revenue of about $840 billion over five years, which still does not cover expenses.
Computing power and infrastructure are the largest single items, with cumulative expenditures expected to be around $856 billion during the same period. The document originates from a non-public presentation related to computing power transactions in July. The company completed a $122 billion financing round in March, with a valuation of $852 billion; at the current consumption rate, this money is expected to run out by 2028. The latest figures have narrowed compared to the previous forecast of about $305 billion in negative cash flow in May.
A new round of financing has entered early discussions. Some investors are approaching at a valuation of about $1.2 trillion, approximately 41% higher than the March valuation, with the scale and final pricing still variable. This round in March was anchored by commitments from Amazon of up to $50 billion, Nvidia and SoftBank each at $30 billion, etc., with $35 billion from Amazon contingent on an IPO or reaching its defined AGI milestone; an additional approximately $3 billion came from individual investors through bank channels. OpenAI declined to comment to the Financial Times.
The IPO timeline has been pushed back. The company originally planned to advance this fall, having submitted confidential filing materials in June, but CEO Sam Altman later stated that there would be no IPO this year, citing safety concerns. During the same period, the company lowered prices to capture market share, with competitors including Anthropic and lower-cost open-source weight models. Cloud and chip contracts are still being increased, with the funding gap directly cited as a reason for the next round of private equity.
Who is buying and who is selling: Buyers are sovereign capital, cloud vendors, and chip manufacturers willing to lock in model distribution rights at a trillion-dollar valuation; sellers are computing power, cabinets, and electricity for the next five years. The event-driven aspect comes from the competition for training and inference capacity, not current profits. Funds are flowing from equity financing to suppliers like Oracle, Microsoft, Amazon, Nvidia, AMD, and CoreWeave. Beneficiaries are infrastructure providers with long-term contracts, while OpenAI's balance sheet, which must refinance before 2028, and public shareholders who would take on the cash burn curve if listed, are under pressure.
Public reports have not provided a breakdown of free cash flow by year, nor disclosed the subscription amount for the $1.2 trillion round; verifiable figures include $278 billion, $856 billion, $122 billion, and $852 billion from internal planning, as well as the postponement of the IPO from this fall.
Source: Public Information
ABAB AI Insight
OpenAI's financing history is about continuously converting "model leadership" into "prepaid computing power." Microsoft's early multi-billion dollar binding moved training to Azure; subsequently, SoftBank, Amazon, and Nvidia turned checks into cabinet and GPU orders. The $122 billion in March no longer resembles traditional venture capital but rather looks like a lifeline from cloud and chip companies to their largest customer. Amazon's $35 billion is contingent on an IPO or AGI conditions, effectively locking in the next round of funding after the narrative is fulfilled.
The flow of money is not for R&D salaries but for data center rents and chip prepayments. Oracle's approximately $300 billion cloud contract, Azure's additions, and AWS and CoreWeave expansions have transformed OpenAI from a software company into a major buyer of electricity and silicon. The motivation is to dominate the inference entry point: whoever fills capacity first can use price reductions to push Anthropic and open-source models below the cost curve. Strategically, it is about trading equity dilution for time, using time to develop the next generation of models, and then using those models to achieve a higher valuation to fill the 2028 gap.
The benchmark is not subscription monetization like Salesforce but rather a WeWork-style capacity pre-purchase layered on top of Netflix's early content deficit, closer to the capital-intensive stages of nuclear energy or semiconductor manufacturing. Anthropic has also raised its valuation to near or above OpenAI's range, indicating that pricing power in the lab has been rewritten by "who can sign for computing power." The industry is transitioning from the late stage of expansion to a control phase: model differences are narrowing, and what is truly scarce is electricity, land, and long-term contracts.
Structural judgment belongs to capital concentration overlaying technological substitution. The mechanism is: software with near-zero marginal replication costs is bound by training and inference with extremely high marginal costs; whoever can continuously concentrate capital at the billion-dollar level can create barriers to entry in computing. Pricing power has temporarily returned from the application layer to chips, clouds, and the power grid. Revenue can increase tenfold, but as long as infrastructure bills rise faster, free cash flow will remain negative—this is the core constraint of this capital path, not a footnote to the revenue story.
ABAB News · Cognitive Laws
Revenue increases tenfold, bills increase elevenfold, still a loss-making machine
Models sell dreams, chips collect cash
Valuation supplements cash flow, time sold to power plants