Cerebras falls behind on first OpenAI capacity tranches as data center deals lag

Cerebras Systems is behind schedule on the first capacity tranches it owes OpenAI. The cause is not the chips. It is the buildings that are supposed to house them.

The AI chipmaker has pinned the delay on data center contracts that are taking longer than planned. Cerebras has described the pace of securing facilities as moving at “the speed of real estate.”

What Cerebras promised OpenAI

The two companies signed a Master Relationship Agreement on December 24, 2025. Under it, OpenAI agreed to buy 750 megawatts (MW) of inference capacity from Cerebras.

The 750 MW is split into three equal tranches. The first 250 MW is due by the end of 2026. Another 250 MW follows by the end of 2027, and the final 250 MW by the end of 2028.

The total value of the agreement is estimated to exceed $20 billion. The deal also includes bonuses tied to performance metrics. OpenAI holds an option for an additional 1.25 gigawatts of capacity by 2030.

Money, revenue and a very public balance sheet

OpenAI did more than sign a purchase order. In January 2026, it extended Cerebras a $1 billion working-capital loan to support the buildout.

Cerebras went public in May 2026 and trades on the Nasdaq under the ticker CBRS.

Cerebras reported a surge in cloud revenue for Q3 2026. It also raised its full-year sales guidance, citing broader demand across its business despite the infrastructure constraints.

Its backlog now exceeds $25 billion, with a significant portion tied to the OpenAI agreement.

Why real estate is the bottleneck

Cerebras has made clear that infrastructure and data center buildouts remain the critical constraint in meeting OpenAI’s rising demand.

To ease the squeeze, Cerebras is pursuing new data center sites across North America and Europe. Among them is a 165 MW facility in Finland that the company is working to secure.

What this means for Cerebras, OpenAI and investors

For Cerebras shareholders, the core issue is execution risk. The company has demand, a large backlog and a marquee customer. What it needs now is physical capacity coming online on schedule, since delays in data center availability could push out when backlog turns into recognized revenue.

The first deadline is the end of 2026 for the initial 250 MW. Cerebras is already reporting that it is behind on that early capacity.

The contract structure adds another layer. Because the deal includes performance-linked bonuses, delivery timing and quality could affect how much of the estimated value Cerebras ultimately captures. The 1.25 GW option for 2030 is also not guaranteed.

Cerebras raised guidance and grew cloud revenue while still falling behind on its biggest contract. That combination suggests demand is not the binding constraint in AI compute right now. Power, sites and contracts are.