OpenAI Cut Its Round to $20 Billion and Its Valuation by $150 Billion — Its Compute Plan Went the Other Way, to $856 Billion

The $600 billion figure now travelling alongside the funding cut is a February artifact. On Feb. 20 OpenAI told investors it would spend roughly $600 billion on compute through 2030, down from the $1.4 trillion Sam Altman had been touting. Two months later The Wall Street Journal reported the number had been raised to about $750 billion. By Sept. 18 the Financial Times put planned spending on computing power and infrastructure at about $856 billion through end-2030, alongside $278 billion of negative free cash flow between 2026 and 2030.

OpenAI's own language in April left no room for doubt about direction. On April 29 the company said it had already surpassed the 10 gigawatts of US infrastructure it committed to in January 2025, with more than 3 gigawatts added in the previous 90 days, and that the only responsible way to meet demand was to build more compute, faster. A company that says it cannot build fast enough does not trim its budget two weeks before a funding round.

Read in sequence, 2026 contains a spend plan that fell once in February and rose twice, a revenue forecast that also rose between February and September, and then a revenue run rate that came in below what investors had been told three weeks before the round shrank. The October revision is a financing revision, not a demand revision. The arithmetic of the round shows why that distinction matters: $856 billion spread over five years is roughly $170 billion of compute a year, the March round alone covered $122 billion of it, and the new $20 billion round covers about 12% of a single year.

The $150 billion valuation cut is 11%. The run-rate revenue cut, from $68 billion to $50 billion, is 26%. Those two numbers do not cancel, and the direction of travel is not what the headline implies: at the $1.4 trillion ask on a $68 billion run rate, OpenAI was priced at 20.6 times annualized revenue; at $1.25 trillion on $50 billion it is priced at 25 times. Part of the gap is definitional rather than operational — a person familiar with the matter said the $68 billion figure had included gross revenue from OpenAI's partners — but even on the company's own numbers the mark did not get cheaper.

Three things keep the demand shock smaller than the headline suggests. OpenAI's capacity commitments are contractual: Oracle's remaining performance obligations reached $664 billion at Aug. 31, and CoreWeave's order form with OpenAI commits up to roughly $6.5 billion through May 31, 2031. The run rate is still compounding fast, at 77% in the September quarter and 107% for the enterprise business. And a large part of the gap is a change in what gets counted.

That is the hinge of the episode. If roughly $170 billion a year of compute has to be funded from cash flow and credit rather than from a fresh equity raise, the marginal dollar of AI infrastructure is underwritten by balance sheets, and the risk premium moves from the chip order book onto those balance sheets. The listed companies carrying them have already reported what it costs.

Microsoft, Alphabet, Amazon, Oracle and CoreWeave together spent $607.1 billion on capital expenditure over the trailing twelve months and generated about negative $9.6 billion of combined free cash flow. None of that spending stops because OpenAI's equity round is smaller. What changes is who absorbs the cost of the next tranche, and at what price.

Oracle shows the arithmetic most cleanly. Remaining performance obligations reached $664 billion at Aug. 31, 2026 from $455 billion a year earlier, cloud infrastructure revenue was $7.4 billion in the quarter, and capital expenditure was $28.5 billion in the same three months. Trailing free cash flow is negative $28.7 billion, filled with $37.2 billion of net new debt, leaving net debt of $119.6 billion — about 3.2 times trailing EBITDA. Depreciation says what is coming: Oracle's quarterly charge nearly doubled year on year to $3.4 billion.

If OpenAI's compute budget is still rising, the physical chain should still be growing. It is, and it grew straight through the scare — which is the strongest single argument that this was never a demand revision.

The market's Oct. 8 reaction confirms that investors now trade this chain off OpenAI's numbers: NVIDIA fell 3%, Oracle nearly 6%, CoreWeave nearly 8%, AMD 4% and Broadcom 4% on the revenue disclosure. What repriced was the financing risk attached to each name, not the volume of hardware the build requires.

On Oct. 26 Anthropic filed to go public, and for the first time a frontier lab's compute commitments became public. The filing shows future obligations to pay providers about $417 billion for compute — data centers, equipment and cloud services — under deals that start between 2026 and 2027 and run from under a year to 20 years. Capacity delivered between 2028 and 2031 is expected to add another $125 billion to $150 billion. Anthropic has agreed to spend at least $542 billion on data centers and cloud computing, and separately agreed to pay SpaceX around $3.1 billion a month once capacity is delivered, up to $120.5 billion through 2029.

This is the part of the IPO calendar that matters for the listed chain. OpenAI's round is private, negotiated, and now priced $150 billion below its September ask. Anthropic's is public, backed by a filing that discloses $417 billion of obligations and an $80.7 billion half-year loss. One of those two numbers will anchor the other, and the anchor gets set in public.

The facts, kept separate from the rest. OpenAI cut its round target to $20 billion and its valuation ask to $1.25 trillion on Oct. 30, having reported an annualized revenue run rate of about $50 billion against the $68 billion investors had been told. Its compute plan stands at about $856 billion through 2030, the highest figure it has published. The five listed companies carrying the build spent $607.1 billion of capital expenditure over the trailing year against about negative $9.6 billion of combined free cash flow.

The inference is that this changes the funding mix for AI infrastructure rather than its volume. OpenAI's capacity commitments are contractual and its build plan has been raised, not cut, so the damage landed on financed balance sheets and left order books intact. The corollary is uncomfortable: with the revenue base down 26% and the price down 11%, the implied multiple on OpenAI's own revenue rose, which makes the next dollar of OpenAI equity more expensive to issue rather than less.