OpenAI's commitment to purchase 750 megawatts of computing capacity from Cerebras, valued at over $20 billion, has created a backlog so large it represents nearly 29 times the chip maker's expected 2026 revenue. This single agreement reveals just how expensive it is to keep large language models like GPT running at scale, and why infrastructure companies are suddenly worth tens of billions of dollars.

In December 2025, Cerebras Systems signed a master relationship agreement with OpenAI that committed the ChatGPT maker to purchase 750 megawatts of computing capacity for AI inference, with an option to buy an additional 1.25 gigawatts by the end of 2030. The deal was valued at more than $20 billion, and it arrived on Cerebras's books almost immediately. By the end of 2025, the company had $24.6 billion in remaining performance obligations, the accounting term for contracted work not yet delivered. That figure grew only slightly to $25.4 billion by June 2026.

The backlog is enormous, but it converts into actual revenue slowly by design. Cerebras expects to recognize only about 22 percent of the $25.4 billion, roughly $5.6 billion, over the 24 months ending June 30, 2028. Another 43 percent should arrive between months 25 and 48, with the remainder coming later. The slow conversion reflects the reality that Cerebras is still building the data center capacity it has sold. Deployment happens in stages from 2026 through 2028, with more than 600 megawatts of capacity already live or under contract for delivery by the end of 2027.

Why Is Building AI Infrastructure Taking So Long?

The delay between signing a contract and recognizing revenue highlights a fundamental challenge in the AI boom: the physical infrastructure required to run these models is enormous and takes years to build. Cerebras announced a new 165-megawatt data center in Finland just this week, and the company is expanding manufacturing capacity more than tenfold in 2026. OpenAI is even helping to finance the build-out, advancing Cerebras a $1 billion working capital loan in January to accelerate construction.

This infrastructure spending is not unique to Cerebras. The company's second-quarter results showed adjusted revenue grew 103 percent year over year to $209.9 million, and its inference cloud business nearly quadrupled. Management raised its full-year outlook to $880 million to $890 million in adjusted revenue. Yet despite this explosive growth, the company is still posting operating losses and trades at roughly 58 times expected 2026 revenue, a valuation that depends entirely on the company executing on its backlog commitments.

How Does OpenAI's Deal Reshape the AI Supply Chain?

The OpenAI agreement reveals a critical shift in how AI companies are structured. Rather than building all their own data centers, OpenAI is outsourcing inference capacity to specialized infrastructure providers. This allows OpenAI to focus on model development while companies like Cerebras handle the physical build-out. The arrangement also spreads financial risk; Cerebras bears the capital expenditure burden, while OpenAI commits to long-term purchasing agreements that guarantee revenue.

Cerebras is not alone in its customer concentration. In 2025, Mohamed bin Zayed University of Artificial Intelligence accounted for 62 percent of the company's revenue, and Group 42 accounted for another 24 percent. In the second quarter of 2026, three customers each accounted for at least 10 percent of revenue, representing 76 percent of total revenue combined. This concentration matters because it means a small number of buyers are driving the entire AI infrastructure boom.

Steps to Understanding AI Infrastructure Economics

• Recognize the scale: A single OpenAI deal is worth more than $20 billion and represents nearly 29 times a major chip maker's annual revenue, illustrating the unprecedented capital requirements of modern AI systems.
• Understand the timeline: Infrastructure contracts are recognized as revenue over years, not quarters, because data centers must be physically built and deployed in stages before revenue can be recorded.
• Track customer concentration: A handful of customers like OpenAI, universities, and investment firms account for the vast majority of AI infrastructure spending, creating both opportunity and risk for suppliers.

The backlog is evidence of extraordinary demand for AI computing power, and arguably the best reason to watch Cerebras closely. However, the company must execute flawlessly on its commitments, building enormous capacity on time and on budget, much of it for a single buyer whose needs could change. The valuation reflects not just current revenue, but the assumption that Cerebras will successfully convert its $25.4 billion backlog into actual profits over the next several years.

Meanwhile, OpenAI's expansion into advertising is creating a new revenue stream that could help offset these massive infrastructure costs. On August 31, 2026, OpenAI expanded ChatGPT Ads to the Middle East and North Africa, and the company announced that its advertising business had reached a $1 billion annualized revenue run rate in under 200 days. Ads appear alongside relevant conversations, with the platform using context to decide which ad fits, and they are clearly labeled and separate from ChatGPT's answers.

The advertising expansion shows OpenAI is diversifying its revenue beyond subscriptions and API access. Ads currently show only to Free and Go users, while Pro, Business, Enterprise, and Education users have an ad-free experience. This means the reach is primarily general users rather than decision-makers at large companies, but it still represents a significant new income stream that could help fund the company's infrastructure spending.