Cerebras Systems is heading to the public markets with an unusual problem: its most important customer and financial backer is the same company that could destroy its business if the relationship falls apart. The AI chip startup announced its initial public offering on May 4 at a target valuation of $26.6 billion, planning to sell 28 million shares at $115 to $125 each. But beneath the headline numbers lies a web of dependencies that makes this one of the most structurally peculiar public offerings in recent AI history.

The relationship between Cerebras and OpenAI reads like a venture capital love story gone complicated. In December 2025, OpenAI loaned Cerebras $1 billion in working capital. In exchange, OpenAI received warrants for 33.5 million Cerebras shares at an exercise price described as "a fraction of a penny." Shortly after, in early 2026, Cerebras signed a multi-year agreement worth more than $20 billion with OpenAI to deliver 750 megawatts of computing capacity through 2028.

What makes this arrangement particularly striking is that OpenAI's leadership has become personally invested in Cerebras' success. Sam Altman, Greg Brockman, Ilya Sutskever, and Adam D'Angelo are all listed as angel investors in the company. This is not coincidental. OpenAI needs what Cerebras offers: an inference chip fast enough to power real-time applications at costs that do not require NVIDIA to approve every contract.

Why Does Inference Speed Matter More Than You Might Think?

To understand why OpenAI is willing to back Cerebras so heavily, you need to understand the difference between training and inference. Training is what happens when a company builds a large language model, like GPT-5. Inference is what happens every time someone uses ChatGPT, Gemini, or a code assistant. These are fundamentally different computational problems that favor different hardware architectures.

Cerebras' Wafer-Scale Engine 3 (WSE-3) was designed specifically for inference workloads. The chip is 46,225 square millimeters, compared to 814 square millimeters for NVIDIA's H100. It contains 900,000 cores with memory bandwidth seven thousand times greater than the H100. In practical terms, Cerebras claims the WSE-3 delivers 450 tokens per second running Llama 3.1 70B, roughly 20 times faster than NVIDIA GPU-based cloud solutions running the same model.

This performance advantage matters most for latency-sensitive enterprise applications where speed is not just a nice-to-have but a requirement. Financial risk models must respond in under 100 milliseconds. Real-time API calls in agentic workflows cannot tolerate delays. Medical diagnostics requiring immediate output cannot wait for batch processing. You cannot achieve sub-100 millisecond inference by stacking more H100s together because the architecture simply does not work that way.

How to Evaluate the Real Risks Behind Cerebras' IPO

• Customer Concentration: Two UAE-linked entities generated 86 percent of Cerebras' total sales in 2025, with Mohamed bin Zayed University of Artificial Intelligence accounting for 62 percent of revenue and G42 contributing 24 percent, creating extreme dependency on a small number of customers.
• Contract Termination Clauses: OpenAI's $20 billion computing contract contains termination provisions that could unwind Cerebras' entire business model if the company fails to deliver computing capacity on agreed timelines.
• Warrant Valuation Advantage: OpenAI's 33.5 million warrants are exercisable at a fraction of a penny per share, meaning at the IPO price of $125 per share, those warrants are worth approximately $4.2 billion despite the $1 billion loan investment.

The IPO prospectus acknowledges these risks directly. If Cerebras fails to deliver computing capacity on the agreed timelines, OpenAI can terminate part or all of the contract. The $1 billion loan could become repayable under certain circumstances. The IPO is, in part, a way for Cerebras to reduce that dependency before it becomes a liability in the public markets.

"The real concentration risk is not NVIDIA. It is that Cerebras has two customers who can each inflict existential damage," noted William Keating, an independent analyst.

William Keating, Independent Analyst

The customer concentration problem is particularly acute because Cerebras has essentially shifted its risk rather than reduced it. The company filed for its first IPO in 2024, but that attempt was delayed after national security reviewers scrutinized G42's ties to Chinese technology companies. The clearance eventually came in 2025. The second filing does not reduce the UAE exposure; it shifts the concentration from G42 down from 85 percent of 2024 revenue to 24 percent, while MBZUAI has become the dominant customer at 62 percent.

What Makes Cerebras Different From NVIDIA in the Inference Market?

The chip market that Cerebras is entering in 2026 looks fundamentally different from the one where NVIDIA became dominant. Training large models, which dominated AI infrastructure spending between 2020 and 2025, favors massive GPU clusters, parallelism, and NVIDIA's CUDA software ecosystem. Inference favors something entirely different: low latency, high throughput, and single-chip efficiency.

NVIDIA's H100 and Blackwell families win on training workloads, software ecosystem depth, and multi-modal flexibility. CUDA has sixteen years of tooling, libraries, and developer familiarity behind it. For any organization building or fine-tuning large models, NVIDIA remains the default and will likely remain so through the next hardware generation. Cerebras wins on inference latency for specific architectures. The WSE-3's single-chip design eliminates the inter-chip communication overhead that slows GPU clusters at low-batch inference.

For a financial institution running real-time compliance checks or an agentic workflow requiring synchronous model calls, the performance gap is meaningful and cannot be closed by adding more H100s. NVIDIA's Blackwell architecture is expected to narrow the inference speed difference on standard benchmarks, but it will not replicate the single-chip latency profile that Cerebras offers.

The IPO's oversubscription tells part of the story. Banks are fielding roughly $10 billion in orders for the shares, nearly three times the target, suggesting the price will close above the stated range. What this reflects is institutional appetite for AI chip supply chain diversification. Investors want NVIDIA alternatives to exist. Whether Cerebras is the right vehicle is a separate question from whether the demand for that vehicle is real.

The real question for investors is whether Cerebras can maintain its independence and deliver on its promises while managing the complex relationship with OpenAI. The company has solved one problem by going public, but it has not solved the fundamental concentration risk that comes from having two customers who can each inflict existential damage to the business.