AI Infrastructure Stress Test Is Here! NYT: OpenAI Projects $750 Billion Compute Spending by 2030, Yet Still Cries 'Compute Is Severely Insufficient'

The New York Times' DealBook reported on Wednesday (9th) citing sources that OpenAI still feels compute is 'very insufficient,' despite projecting around $750 billion in compute spending by 2030—over 25% higher than the previous estimate of about $600 billion. This indicates that actual supply of chips, data centers, power, and network resources has not kept pace with the growing demand for developing and deploying major large models. For companies like Oracle and CoreWeave, which are betting on AI demand, this serves as both a demand signal and a direct strain on their balance sheets.

As early as July 22, The Wall Street Journal reported that OpenAI had raised its forecast for future cloud infrastructure spending to $750 billion and signed capacity agreements with multiple partners: a 6GW data center capacity contract with Oracle, most of which remains in development; an expanded eight-year, $138 billion agreement with Amazon AWS, including 2GW of Trainium chip compute; and a commitment to increase cloud spending on Microsoft Azure by $250 billion. These large contracts indicate OpenAI is securing long-term compute through a multi-supplier strategy.

Compute demand brings long-term orders to cloud providers, but expansion costs also test balance sheets. Oracle's remaining performance obligations stand at $638 billion, up 363% year-on-year, with $75 billion related to customer prepayments or GPU arrangements. Its cloud infrastructure revenue reached $5.79 billion, up 93% year-on-year, but free cash flow in the fourth fiscal quarter was negative $2.369 billion, with capital expenditures totaling $55.66 billion. Oracle expects net cash outflows for capital spending of about $70 billion in fiscal 2027 and plans to raise approximately $40 billion, including $20 billion through at-the-market stock offerings.

CoreWeave is in a similar situation. The company reported negative free cash flow of $5.743 billion in Q2, with a backlog of approximately $104 billion as of June 30, and aims to achieve over 8 gigawatts of active power capacity by 2030. Its expansion relies heavily on debt, prepayments, and equity financing.

Supply constraints are concentrated in chips, power, and data centers. NVIDIA CEO Jensen Huang stated that the supply chain is under strain, with current supply meeting only about 70% of demand. Broadcom is advancing the deployment of OpenAI's first-generation custom accelerator, Jalapeno, targeting 1.3GW by fiscal 2027, but noted that land, power, and data center shells determine production timelines—indicating bottlenecks extend beyond GPUs. Wafers, HBM, substrates, power, and construction capacity could all impact delivery.

While OpenAI's latest statements reinforce demand signals, they also present a more realistic test: whether massive orders can translate into gigawatt-scale compute, and whether companies can maintain financing and cash flow balance during high capital expenditure cycles.