Bernstein: OpenAI's Compute Commitment Revenue Outpaces Billings, Downside Risk Low for Oracle and Microsoft
nashnova research
Bernstein estimates OpenAI's infrastructure commitments at roughly $614 billion through 2030, but revenue growth is outrunning spend, limiting downside for Oracle and Microsoft — the real risk lies in speculative builds once capacity scarcity fades.
$750 billion in contracts — what's the real bill?
The widely cited "$1.5 trillion-plus" figure is closer to a long-term vision. Bernstein rebuilt the model bottom-up: if every contract delivered on time — which the report calls "extremely unlikely" — total spend through 2030 would be roughly $614 billion.
This means → the actual bill will be lower still, cushioned by three layers: chips are likely vendor-financed over roughly six-year hardware lives; AMD and Cerebras deals carry warrants — a financial instrument that vests based on stock price — making those chips potentially near-free for OpenAI; and self-built data-center shells (~$20 billion, phased 2028–2032) are long-life assets front-loaded in the model.
In plain terms = signing $750 billion in contracts does not mean paying $750 billion in cash by 2030 — installments, financial instruments, and time spread the load far thinner.
Can revenue growth keep outrunning the compute bill?
OpenAI told investors its annualized revenue approached $50 billion by late September, with year-on-year growth of 77%. That projects to over $80 billion next year — above Bernstein's estimate of 2027 infrastructure spend.
By 2028, infrastructure spend is forecast to rise ~58% to $139 billion. As long as revenue sustains 60%+ growth, it keeps outpacing spend.
This means → OpenAI's leverage and payment capacity are scaling in tandem with its commitments. Bernstein notes OpenAI has beaten every revenue target so far; at the industry level, global inference spend must eventually reach 5–20× training spend — the market is still early.
How hard are these contracts across different suppliers?
Oracle's core deal is ~$300 billion, structured as multiple five-to-six-year capacity tranches. Bernstein estimates actual recognizable revenue at ~$312 billion, with 57% booked before Oracle's fiscal 2030. OpenAI's share of Oracle's remaining performance obligations sits below 50% and will keep declining.
Microsoft plays three roles — investor, model licensee, and capacity provider. OpenAI has signed for an incremental $250 billion in Azure services, each tranche a six-year contract. Amazon holds two deals totaling ~$138 billion, including up to 2 GW of Trainium chips.
CoreWeave's contracts total ~$22.4 billion, but Bernstein stresses the contract language reads as a "ceiling," not committed revenue. This reflects a stark gap in contract rigidity: hyperscalers lock in multi-year commitments, while independent compute lessors hold flexible frameworks.
Are these commitments truly non-cancellable?
Two cancellations are already on record: OpenAI pulled out of an Oracle 2.1 GW expansion in Abilene, Texas, ceding an adjacent 900 MW site to Microsoft/Crusoe — builder Crusoe found a replacement tenant within roughly two weeks. A Norway Stargate project was likewise handed off to Microsoft after OpenAI's exit.
In plain terms = "announced" and "non-cancellable contract" are very different things — top labs can adjust or exit published projects, and replacement demand is currently filling the gaps fast.
This means → for suppliers like Oracle and Microsoft, a single project cancellation is not systemic risk — capacity gets absorbed by other industry buyers.
Where is the real risk?
Bernstein concludes that AI labs are driving construction beyond actual near-term need; supply-demand normalization could arrive within two to three years.
This reflects a risk that sits not in the headline commitment figures but in the speculative builds and low-quality business models that emerge once capacity scarcity eases — Bernstein reiterates its skepticism of the neocloud model (independent compute lessors).
This means → labs will increasingly rely on hyperscale cloud partnerships, colocation agreements, and self-build, pressuring the long-term competitive position of standalone compute lessors. Bernstein rates CoreWeave underperform with a $54 target, Oracle outperform at $325, and Microsoft outperform at $722.
市场有风险，内容仅供研究参考，不构成投资建议。