OpenAI's Five-Year Computing Power Spending Reaches $856 Billion
According to a report by the Financial Times on September 18, an internal confidential document from OpenAI shows that the company expects its cumulative spending on computing power and infrastructure to reach $856 billion by the end of 2030. This is the largest single expenditure listed in the document and marks the first time the public has seen the company's actual spending plans, which are significantly higher than previously disclosed figures.
The nature of this document is not a general business forecast but an "underwriting document"—provided in July 2026 to an unnamed infrastructure partner to assess whether it is worthwhile to commit to supplying power, chips, and infrastructure capacity for OpenAI in the coming years. The core logic of the document sets the "availability of capacity" as a hard constraint for business development, rather than "market demand being sufficient." Revenue growth is described in the document as a "default assumption," with the real operational challenge being how to raise and implement the $856 billion in infrastructure spending.
The document indicates that OpenAI expects revenue of about $36 billion in 2026, growing to about $350 billion by 2030, with cumulative revenue of about $840 billion over five years. However, even so, the company anticipates a cumulative free cash flow gap of $278 billion (i.e., net cash consumption) from 2026 to 2030. This cash flow gap forecast has narrowed from the $305 billion disclosed by the company in May this year, but the absolute scale remains substantial.
OpenAI completed a round of financing of $122 billion in March this year, but based on the current spending pace, this funding could be exhausted as early as 2028, meaning the company will need a new round of large-scale financing to maintain operations. It is reported that investors are currently in discussions regarding a new valuation for the company, with discussions around $1.2 trillion, while OpenAI hopes to achieve a higher valuation level.
Notably, OpenAI secretly submitted an IPO application in June this year, originally planning to launch the IPO within 2026, but later voluntarily postponed this plan, citing "concerns about the pace of cutting-edge AI development and doubts about how the public market would price a company with such significant expected losses."
The $856 billion internal figure sharply contrasts with the public figures previously disclosed by the company: in February this year, OpenAI told investors that its computing power spending target for 2030 was about $600 billion; subsequently, the Wall Street Journal reported that the company had raised this target to $750 billion; now, the leaked internal document shows that the actual planned figure has further climbed to $856 billion—an increase of over 40% in just over six months.
From an industry chain perspective, the recipients of this massive computing power expenditure are infrastructure and chip suppliers such as NVIDIA, Oracle, and SoftBank-related companies. OpenAI has signed long-term "take-or-pay" procurement commitments with these suppliers to secure priority allocation of capacity; the essence of this arrangement is to transform OpenAI's yet-to-be-realized revenue growth expectations into a basis for suppliers to lock in and expand capital expenditures. The beneficiaries are these suppliers who have received long-term locked orders, which provide certainty for their stock prices and capital expenditure plans; while the pressure falls on OpenAI's own balance sheet—if actual revenue growth does not meet the aggressive assumption of $350 billion, the company will face the dual risks of expanding cash flow gaps and breaching contracts with suppliers. The launch of new models in July this year has already led to a year-on-year revenue growth of about 20%, becoming key recent evidence supporting this optimistic forecast.
Source: Public Information
ABAB AI Insight
OpenAI's disclosed computing power spending targets have been continuously revised upward over the past six months, showing a clear historical trajectory—initially set at about $600 billion in February 2026 for investors, then raised to $750 billion according to the Wall Street Journal, and now revealed by the Financial Times to have reached $856 billion in internal documents. This pattern of "external figures consistently being surpassed by internal real numbers" is highly similar to the historical trajectory of several tech companies repeatedly raising budgets during capital expenditure cycles, often indicating that actual capacity procurement and order locking have outpaced public financial communications.
This document essentially reveals a clear path of capital flow: OpenAI uses its unrealized revenue growth expectations (from $36 billion to $350 billion) as leverage to sign long-term "take-or-pay" capacity commitments with suppliers like NVIDIA, Oracle, and SoftBank-related companies, which in turn expands their capital expenditures and capacity construction. This arrangement partially shifts the infrastructure investments that should be supported by OpenAI's own cash flow onto the suppliers' balance sheets, effectively exposing suppliers to the risk of "paying in advance for the customer's yet-to-be-realized revenue," and whether this funding can be repaid smoothly ultimately depends on OpenAI's ability to secure new financing on time and whether actual revenues can keep pace with the projected curve.
A comparable historical case is the common "take-or-pay" long-term procurement agreements between operators and equipment manufacturers during the telecom industry's 3G and 4G construction cycles—operators use unrealized user growth and traffic revenue expectations to obtain early production expansions from equipment manufacturers, and when actual demand falls short of expectations, both operators and manufacturers often find themselves in financial distress. The current AI industry is in an expansion phase where leading model companies collectively bet on "revenue will grow exponentially" and sign massive long-term infrastructure contracts based on this assumption, similar to OpenAI and Anthropic, which is simultaneously racing towards a $2 trillion IPO valuation, both betting on capital expenditures far exceeding current actual revenue scales for the future.
This incident essentially reflects that global AI infrastructure capital is highly concentrated towards a few leading laboratories and their supplier chains: the $856 billion five-year spending plan means OpenAI has placed almost all of its capital operation focus on whether it can continuously obtain external financing and supplier capacity commitments, which in turn relies on whether the market is willing to continue to provide funding based on continuously rising expected valuations (such as $1.2 trillion or even higher). Mechanistically, this concentration of capital can be maintained because only a very few companies recognized by the market as "capable of leading the commercialization of AGI" can persuade investors and suppliers to jointly bet on an unverified revenue curve; once this curve deviates significantly, whether it be a delayed IPO or forced capital expenditure cuts, will become key signals to test whether the valuation logic of the entire industry holds.
ABAB News · Cognitive Law
- External figures always lag behind internal real numbers.
- Supplier willingness to lock orders is the true support for valuation.
- The day the revenue curve fails to meet expectations is when the bills truly come due.