Bernstein breaks down OpenAI's compute commitments: revenue growth outpaces the bill, with low downside risk for suppliers such as Oracle Corporation (ORCL.US) and Microsoft Corporation (MSFT.US)

As AI labs sign up for a substantial share of AI compute capacity under construction, concerns about customer concentration risk continue to heat up. In this report, Bernstein uses OpenAI's annual contractual commitments as a proxy for all AI labs and reaches a conclusion contrary to the panic narrative: the commitments are large in scale but layered over time and not all irrevocable; while capacity and related costs grow, OpenAI's leverage and ability to pay are also growingfor major suppliers such as Oracle Corporation (ORCL.US) and Microsoft Corporation (MSFT.US), downside risk is low. Bernstein maintains "Outperform" ratings on Microsoft Corporation and Oracle Corporation (price targets of US$722 and US$325, respectively) and an "Underperform" rating on CoreWeave (price target US$54): it has no concerns about its existing backlog but doubts its ability to renew contracts and sign incremental contracts with strong economics. The true composition of the US$750 billion in commitments Bernstein first clarifies the scope: the media-circulated figure that "OpenAI has committed more than US$1.5 trillion" is closer to a long-term vision than to specific contracts. Two more reliable data points are: in October 2025, OpenAI said Stargate had secured nearly 7 GW of planned capacity, with more than US$400 billion of investment over the next three years, and that it would lock in US$500 billion and 10 GW of commitments ahead of schedule by the end of 2025; a July 2026 report said OpenAI's cloud infrastructure spending through 2030 would be approximately US$750 billion. Bernstein's bottom-up reconstruction estimates that if all contracts are delivered on timewhich the report explicitly says is "highly unlikely"OpenAI's infrastructure spending through 2030 would be approximately US$614 billion; on an annual basis, the peak of currently committed spending occurs in 2031, at approximately US$184 billion. Signing US$750 billion of contracts does not equal paying US$750 billion of cash before 2030, for three reasons: first, chips from NVIDIA Corporation, Broadcom Inc. and AMD are likely to be vendor-financed and paid in installments over an approximately six-year hardware life, with delivery timing and volumes both variable; second, agreements with AMD and Cerebras carry warrants that vest in tranches based on AMD's share price, and depending on the share price these chips could be nearly free to OpenAI; third, the self-built Project Camellia data center shell (approximately US$20 billion, 3.2 GW in Georgia, phased into operation from 2028 to 2032) is a long-lived asset that could have been recognized in stages or financed otherwise, but is treated in the report only as an upfront charge. Revenue outpaces the compute bill The key data point on the demand side is: OpenAI reportedly told investors that annualized revenue was close to US$50 billion at the end of September, with total revenue growing 77% year over year. On that basis, OpenAI's revenue next year will exceed US$80 billion, higher than the report's estimated 2027 infrastructure spending (after excluding the US$15 billion Camellia shell). The report estimates that OpenAI's infrastructure spending will grow approximately 58% year over year to US$139 billion in 2028 (US$129 billion excluding the shell); as long as revenue maintains growth of 60% or more, revenue will continue to outpace infrastructure spending. Bernstein notes that OpenAI has exceeded every revenue target to date; and at the industry level, global annual inference spending must ultimately reach 5 to 20 times training spending, and it is still early. Supplier tiers: whose contracts are the firmest Oracle Corporation's relationship with OpenAI centers on a US$300 billion agreement, composed of multiple tranches of 5- to 6-year capacity contracts (signed in 2025, effective from 2026, and coming online progressively through 2030, with Bernstein expecting the bulk in 2028). Bernstein estimates that Oracle Corporation can actually recognize approximately US$312 billion of revenue, 57% of which will be booked before its fiscal 2030; OpenAI accounts for less than 50% of Oracle Corporation's RPO, and that share will continue to decline as other businesses grow. Microsoft Corporation's relationship with OpenAI is complex: as an investor, model user and capacity supplier, Microsoft Corporation previously announced that OpenAI had contracted to purchase an incremental US$250 billion of Azure services, with each tranche a 6-year contract corresponding to hardware life. Amazon.com, Inc. has two agreements: a US$38 billion 7-year agreement signed in 2025, and a US$100 billion 8-year agreement signed in 2026 as part of the partnership (including up to 2 GW of Trainium chips). CoreWeave (CRWV.US)'s total contract value with OpenAI is approximately US$22.4 billion (US$11.9 billion in March 2025 with a US$350 million equity investment, US$4 billion in May, and up to US$6.5 billion in September), but Bernstein emphasizes that its contract language is a "ceiling" rather than committed revenue. Although Alphabet Inc. Class C (GOOGL.US) has repeatedly been reported to have capacity and TPU agreements with OpenAI, there is a lack of contract evidence verifiable in financial reports, and Bernstein has excluded it from the model. Commitments are not monolithic The report specifically cites two cases of cancellation: OpenAI had reached an agreement with Oracle Corporation for a 2.1 GW expansion in Abilene, Texas, then withdrew and ceded the adjacent 900 MW site to Microsoft Corporation/Crusoe, and the builder Crusoe found a replacement tenant in about two weeks; the Norway Stargate project announced in July (an Nscale and Aker partnership) was likewise taken over by Microsoft Corporation after OpenAI exited. These cases show a significant gap between "announced" and "irrevocable contract." Bernstein judges that AI labs are pushing construction of capacity beyond what is actually needed, and that supply-demand normalization may occur within the next 2 to 3 years; the real risk to the industry is not the scale of commitments from leading labs, but the speculative construction and low-quality business models spawned once capacity scarcity eases. Bernstein reiterates its skepticism about the neocloud business model, believing that labs will increasingly rely on hyperscale cloud partnerships, hosting agreements and self-build.