Bernstein: OpenAI’s computing spending rises year by year as cloud providers’ existing contracts and renewal prospects diverge

When do massive computing commitments become spending?

Bernstein uses OpenAI as a proxy for examining AI labs’ computing commitments: contract values span multiple years, data centers come online in stages, and some agreements can be adjusted or canceled. Long-term commitments cannot all be treated as near-term payments.
The firm estimates that if all contracts included in its model are delivered on time and none are canceled, OpenAI’s infrastructure spending through 2030 will total 614 billion dollars. Based on existing commitments, annual spending will peak at approximately 184 billion dollars in 2031. Project delays or cancellations would change the timing and total amount of spending.
The model assumes some chip purchases are financed by suppliers and amortized over approximately 5–6 years; the number and timing of chip deliveries may also change. Data center hosting and hardware purchases are calculated separately, helping avoid counting the same batch of customer-provided chips twice in computing contracts.

According to media reports cited in the report, OpenAI told investors that annualized revenue at the end of September 2026 was approximately 50 billion dollars, with revenue up 77% year over year. Annualized revenue reflects the run rate at that time; actual full-year revenue must be calculated separately.
Bernstein extrapolates from this that revenue could exceed 80 billion dollars in 2027, above its estimate of infrastructure spending that year. This comparison, however, excludes the 15 billion dollars of upfront investment in the Project Camellia data center shell.
The firm’s model shows infrastructure spending rising approximately 58% year over year in 2028 to 139 billion dollars, or 129 billion dollars excluding construction of the Camellia shell. If OpenAI’s revenue growth reaches 60% or more and there are no new data center contracts, revenue growth will continue to outpace infrastructure spending growth.

Oracle’s main agreement with OpenAI is worth approximately 300 billion dollars and consists of multiple batches of capacity contracts, each lasting 5–6 years. Bernstein estimates that most of the capacity could come online in 2028. The firm believes OpenAI accounts for less than 50% of Oracle’s remaining performance obligations.
Microsoft’s new 250 billion dollars Azure services commitment with OpenAI is also being implemented in batches, with each contract lasting approximately six years. Bernstein believes Microsoft’s global resource footprint makes its revenue less likely to be affected by project delays than that of newer cloud providers and hosting companies.
CoreWeave’s contracts with OpenAI total up to approximately 22.4 billion dollars; Bernstein notes that this figure is the contractual maximum. The firm is relatively confident in the existing backlog, but is skeptical about future renewals or securing new contracts on favorable economic terms amid increased competition.

Bernstein believes Oracle and Microsoft have strong bargaining positions with both suppliers and customers, and expects both companies’ OpenAI contracts to be renewed at least once more. Its model still includes only the initial contract terms; smaller or shorter-term contracts may not be renewed.
The firm expects computing capacity built in response to AI labs’ demand could exceed those labs’ own needs, and that supply and demand could normalize over the next 2–3 years. Easing computing scarcity will affect speculative construction and weaker business models, while major cloud providers’ platform tools and customer relationships help make their services stickier.