Cerebras Wants $26.6 Billion From Public Markets. OpenAI Is 87 Percent of the Business.

Cerebras wants Wall Street to believe it is a diversified AI infrastructure company. OpenAI is not cooperating with that story.

The chipmaker, which refiled its IPO on Monday with plans to raise $3.5 billion at a valuation of up to $26.6 billion Reuters, reported $510 million in 2025 revenue TechCrunch. OpenAI accounted for roughly 87 percent of it, according to the company's own S-1 disclosures Tech-Insider. The $20 billion multi-year deal Cerebras announced in January, under which OpenAI will draw up to 750 megawatts of compute through 2028 Reuters, is not a growth catalyst. It is the business.

Non-GAAP net income was $237.8 million TechCrunch. Under standard accounting, the same numbers produced a GAAP net loss of $75.7 million. The difference is stock compensation, depreciation, and one-time items, not two different businesses. The profit story and the loss story are the same story.

"The path to diversified revenue streams beyond two anchor customers" is how Bernstein semiconductor analyst Stacy Rasgon put it, quoted by Tech-Insider Tech-Insider. That is the polite version. The blunt version: Cerebras is selling a wafer-scale chip company at GPU-company valuations while being a one-customer infrastructure play.

Andrew Feldman, the CEO, is not selling shares in the IPO CNBC. He will hold 10.3 million shares worth up to $1.28 billion at the high end of the range. His incentives are aligned with the stock price going up, not with the public market's ability to exit.

The chip itself is not a story problem. Cerebras builds the Wafer-Scale Engine, a single silicon wafer machined into one chip. The WSE-3 is 46,225 square millimeters, roughly 57 times the die area of Nvidia's H100 Tech-Insider. It carries 44 gigabytes of on-chip SRAM and 900,000 AI-optimized cores. The architecture eliminates the high-bandwidth memory bottleneck that limits GPU inference at scale; a single WSE-3, by Cerebras's own specifications, delivers theoretical throughput equivalent to roughly 62 H100 GPUs. That is a different kind of chip ScienceArray.

For inference-heavy workloads, the kind OpenAI runs to serve ChatGPT to millions of users simultaneously, that matters. Feldman told the Wall Street Journal the company had taken the fast inference business from Nvidia at OpenAI TechCrunch. Nvidia did not comment.

The binding term sheet with AWS, announced in early 2026, adds a second large customer. Cerebras described it in the S-1 as a "substantial portion" of projected revenues Reuters. It is not yet deployed revenue. The difference matters: a signed contract and a running system are different things, especially in semiconductor infrastructure where deployment timelines stretch across quarters.

Independent benchmark data on the WSE-3 in production environments is thin. Cerebras publishes performance claims. Nvidia publishes performance claims. The third-party validation that infrastructure buyers typically require before committing multi-year capex budgets has not materialized in publicly available form arXiv.

Cerebras tried to go public once before, withdrawing its S-1 in October 2024 days after a $1 billion funding round Reuters. The cause was a national security review of G42, an Abu Dhabi-based conglomerate that held a minority stake and was simultaneously one of Cerebras's largest customers. U.S. officials were concerned that G42's ties to Chinese technology entities could give China access to advanced American AI chips. CFIUS cleared the investment structure in 2025. Cerebras filed again in April SEC.

G42 remains a shareholder. The risk factor did not disappear from the S-1; it survived the review with disclosures rather than restrictions.

The updated valuation range targets up to $26.6 billion Reuters, a step up from the $23 billion private valuation Cerebras announced in February, when Tiger Global led a $1 billion Series H round with participation from AMD, Coatue, and Benchmark CNBC. The valuation gap reflects updated IPO pricing and market appetite, not a corresponding change in the business fundamentals.

What the numbers reflect is a company with real revenue, genuine technical differentiation in a narrow use case, and a financial structure that is difficult to value independently of one customer relationship. The question the IPO asks public market investors to answer is whether OpenAI's inference needs in 2026 and 2027 represent a permanent business or a launching pad. Feldman is betting on the latter. The S-1 does not resolve it.

Cerebras is expected to price the offering in mid-May SEC. Morgan Stanley, Citigroup, Barclays, and UBS are lead underwriters.