Market concerns over customer concentration risk continue to build as AI labs sign contracts for a substantial share of AI compute capacity still under construction.

In its report, Bernstein used OpenAI's annual contract commitments as a proxy for all AI labs and reached a conclusion that runs counter to the panic narrative: the commitments are large, but they are layered over time and not all are non-cancellable; while capacity and related costs grow, OpenAI's leverage and ability to pay are also growing—for major suppliers such as Oracle (ORCL.US) and Microsoft (MSFT.US), downside risk is low.

Bernstein maintained an Outperform rating on Microsoft and Oracle with price targets of $722 and $325 respectively, and an Underperform rating on CoreWeave with a price target of $54: it has no concerns about its existing backlog, but doubts its ability to renew contracts and sign incremental deals with strong economics.

The True Composition of the $750 Billion Commitment

Bernstein first clarified the terminology: the widely circulated claim that OpenAI has committed more than $1.5 trillion is closer to a long-term vision than to specific contracts.

Two more reliable data points: in October 2025, OpenAI said Stargate had secured nearly 7 gigawatts of planned capacity, with more than $400 billion of investment over the next three years, and that it would lock in $500 billion and 10 gigawatts of commitments ahead of the end of 2025; a July 2026 report said OpenAI's cloud infrastructure spending through 2030 would be about $750 billion.

After rebuilding its model from the bottom up, Bernstein estimated that if all contracts were delivered on time—the report explicitly says this is extremely unlikely—OpenAI's infrastructure spending through 2030 would be about $614 billion; on an annual basis, the peak of currently committed spending occurs in 2031 at about $184 billion.

Signing a $750 billion contract does not equal paying $750 billion in cash before 2030 for three reasons: first, chips from Nvidia, Broadcom and AMD will likely be vendor-financed, paid in installments over an approximately six-year hardware life, with delivery timing and volumes not fixed; second, agreements with AMD and Cerebras include warrants that vest in tranches based on AMD's share price, meaning those chips could be nearly free to OpenAI depending on the stock price; third, the self-built Project Camellia data center shell (about $20 billion, 3.2 gigawatts in Georgia, phased into operation from 2028 to 2032) is a long-lived asset that could have been recognized in installments or financed in other ways, but the report only treats it as an upfront charge.

Revenue Outpaces the Compute Bill

The key demand-side data point is that OpenAI reportedly told investors its annualized revenue was approaching $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 $80 billion, above the report's estimated 2027 infrastructure spending (excluding the $15 billion Camellia shell).

The report estimates OpenAI's 2028 infrastructure spending will grow about 58% year over year to $139 billion (excluding the shell, $129 billion), and as long as revenue maintains growth of 60% or more, revenue will continue to outpace infrastructure spending.

Bernstein noted that OpenAI has exceeded every revenue target so far; and at the industry level, global annual inference spending must eventually reach 5 to 20 times training spending, and it is still early days.

Supplier Tiering: Whose Contracts Are the Most Solid

Oracle's relationship with OpenAI centers on a $300 billion agreement, made up of multiple batches of five- to six-year capacity contracts (signed in 2025, effective from 2026, and coming online through 2030, with Bernstein expecting the bulk in 2028).

Bernstein estimates Oracle's actual recognizable revenue at about $312 billion, of which 57% will be booked before its fiscal 2030; OpenAI accounts for less than 50% of Oracle's RPO, and that share will continue to decline as other businesses grow.

Microsoft's relationship with OpenAI is complex: as an investor, model user and capacity supplier, Microsoft previously announced that OpenAI has contracted to buy an incremental $25 billion of Azure services, with each capacity batch under a six-year contract corresponding to hardware life.

Amazon has two agreements: a $38 billion seven-year agreement signed in 2025, and a $100 billion eight-year agreement signed in 2026 as part of the partnership (including up to 2 gigawatts of Trainium chips).

CoreWeave (CRWV.US) has total contracts with OpenAI of about $22.4 billion ($11.9 billion in March 2025 plus a $350 million equity investment, $4 billion in May, and up to $6.5 billion in September), but Bernstein stressed that its contract language is a cap rather than committed revenue.

Although Google (GOOGL.US) has repeatedly been reported to have capacity and TPU agreements with OpenAI, there is a lack of contract evidence that can be verified in financial reports, and Bernstein has excluded it from its model.

Commitments Are Not Set in Stone

The report specifically lists two cases that have already been cancelled: OpenAI had reached an agreement with Oracle for a 2.1 gigawatt expansion in Abilene, Texas, then withdrew and gave the adjacent 900 megawatt site to Microsoft/Crusoe, and the builder Crusoe found a replacement tenant in about two weeks; the Norway Stargate project announced in July (a partnership between Nscale and Aker) was also taken over by Microsoft after OpenAI withdrew.

These cases show there is a significant distance between an announcement and a non-cancellable contract.

Bernstein judges that AI labs are pushing to build more capacity than is actually needed, and supply-demand normalization could occur within the next two to three 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 that will emerge once capacity scarcity eases.

Bernstein reiterated its skepticism about the neocloud business model, arguing that labs will increasingly rely on hyperscaler partnerships, hosting agreements and self-building.