OpenAI has escalated its projected compute expenditures to $750 billion, a move that intensifies pressure-testing on the AI infrastructure sector amid widening supply gaps.

On September 9, the New York Times' DealBook, citing an unnamed source familiar with the data, reported that OpenAI currently still perceives its compute capacity as "very insufficient."

The organization now anticipates approximately $750 billion in compute spending by 2030, representing a more than 25% upward revision from its earlier estimate of around $600 billion.

This projection underscores that the actual supply of chips, data centers, power, and network resources has yet to keep pace with the escalating demand generated by frontier model development and deployment.

For companies such as Oracle and CoreWeave, which have staked massive infrastructure expansion plans on AI demand, this signals both robust demand and a direct source of balance-sheet strain.

The $750 Billion Blueprint Reflects an Upgraded Procurement Scale

Wall Street observers noted that on July 22, the Wall Street Journal had reported that OpenAI, amid continuously rising AI compute demand, had already revised its future cloud infrastructure spending forecast to $750 billion—indicating the firm is persistently raising its long-term capital commitments.

OpenAI has previously entered into capacity agreements with multiple partners. Among these, the data center capacity contract signed with Oracle totals 6 gigawatts, with the majority still in the development phase.

OpenAI has also expanded its multi-year cloud computing agreement with Amazon Web Services to $138 billion over eight years, which includes 2 gigawatts of compute resources based on Trainium chips.

Additionally, OpenAI has pledged to increase its cloud spending with Microsoft Azure by $250 billion, though the specific timeline for that expenditure has not been disclosed. The accumulation of these substantial contracts reveals that OpenAI is securing long-term compute resources through a multi-supplier strategy.

Nevertheless, the expansion of compute procurement commitments means OpenAI must continuously coordinate the interplay between demand growth, infrastructure build-out, and capital deployment over the coming years.

Order Growth Coexists with Financial Strain

OpenAI's compute appetite provides long-term order support for cloud providers and the infrastructure supply chain, but the associated expansion costs equally test corporate balance sheets.

Taking Oracle as an example, its remaining performance obligations have surged to $638 billion, a 363% year-over-year increase. Of that total, $75 billion is tied to prepayments or arrangements where customers supply GPUs. Oracle's Cloud Infrastructure revenue reached $5.79 billion, up 93% year-over-year.

However, order accumulation is accompanied by hefty capital expenditures. Oracle reported negative free cash flow of $23.69 billion in its fourth fiscal quarter, capital expenditures of $55.66 billion, and projects net cash capital expenditures of approximately $70 billion for fiscal 2027. The company plans to raise roughly $40 billion through a combination of debt and equity financing, including a $20 billion at-the-market stock offering program.

CoreWeave faces a similar situation. The company posted negative free cash flow of $5.743 billion in the second quarter, with an order backlog of approximately $104 billion as of June 30. It has also set a target to exceed 8 gigawatts of active power capacity by 2030. Its expansion is primarily supported by debt, prepayments, and equity financing.

Supply Constraints Remain Concentrated in Chips, Power, and Data Centers

On the chip supply front, Nvidia CEO Jensen Huang recently stated that the supply chain is under strain, with current supply capable of meeting only about 70% of demand.

Meanwhile, Broadcom is advancing deployment of OpenAI's first-generation custom accelerator, Jalapeno, with plans to deploy 1.3 gigawatts by fiscal 2027. However, the company simultaneously noted that land, power, and data center building shells will determine the actual timeline for capacity realization.

This indicates that bottlenecks in AI infrastructure are not confined to GPU supply alone. Wafers, high-bandwidth memory, substrates, electricity, land, and data center construction capacity all have the potential to impact the final delivery of compute power.

OpenAI's latest assertion that compute remains insufficient reinforces the demand signals for AI infrastructure, while also presenting the market with a more practical yardstick: whether massive orders can be converted into usable, gigawatt-scale compute power as planned, and whether the companies involved can sustain the balance between financing and cash flow through a period of high capital expenditure.