After a CFIUS-induced retreat in 2024, the wafer-scale chip startup is back with an OpenAI deal in its pocket and a sharper bet against Nvidia.

Eighteen months ago, Cerebras Systems looked like a cautionary tale. The Sunnyvale-based AI chip startup had filed for an initial public offering in September 2024, only to see its plans stall under a national-security review of its largest customer.

By October 2025, the registration was withdrawn. The company, which makes the largest commercial silicon chip in the world, was not so much beaten as quietly removed from the field.

It is back now, and considerably more ambitious.

On Friday, Bloomberg reported that Cerebras is seeking to raise as much as $4bn in its IPO, at a valuation of roughly $40bn, citing people familiar with the matter. The figure is striking.

It is nearly five times the $8.1bn private-market valuation Cerebras commanded as recently as September 2025, and well above the $22bn to $25bn band industry analysts had expected when the company refiled its S-1 with the US Securities and Exchange Commission on April 17.

Cerebras intends to list on the Nasdaq Global Select Market under the ticker CBRS, with Morgan Stanley, Citigroup, Barclays, and UBS Investment Bank as joint book-running managers. If pricing holds, this would be the first major AI hardware listing of 2026 and one of the largest US technology debuts in recent memory.

Much of the new appetite traces to a single contract. In its updated prospectus, Cerebras disclosed a multi-year compute agreement with OpenAI worth more than $10bn over its term, covering up to 750 megawatts of inference capacity through 2028.

For a company that booked $510m in revenue in 2025, up 76 per cent on the prior year, the deal is transformative. It also explains, in part, the leap in valuation: Cerebras now has the kind of anchor customer that investors prize above almost any other signal in the AI infrastructure market.

The OpenAI agreement does not displace Nvidia, which remains the dominant supplier of the GPUs powering most of the industry’s training workloads. But it does give Cerebras a credible foothold in inference, the part of the AI stack where compute is run, not built, and where margin pressure is mounting fastest.

The company’s wafer-scale processors, physically dozens of times the size of an Nvidia H100, are designed for exactly that workload.

Cerebras’s first IPO attempt collapsed not over technology but over geopolitics. Abu Dhabi-based G42, an AI conglomerate with prior ties to Chinese partners, had become the company’s largest customer and a sizeable shareholder, accounting for 87 per cent of Cerebras’s revenue in the first half of 2024.

The Committee on Foreign Investment in the United States opened a review of G42’s stake, and Cerebras paused its filing. CNBC reported that CFIUS clearance was finally granted on 31 March 2025, after G42 divested its Chinese investments and agreed to restructure its Cerebras holding into non-voting shares.

The company’s S-1 reflects that hard-won outcome. G42’s economic stake remains, but its governance influence has been curtailed. Whether that will satisfy public-market investors who continue to scrutinise foreign exposure in US chip firms is a separate question. The risk factors section of the prospectus dwells on it at some length.

Between abandoning its first IPO and refiling, Cerebras did what well-capitalised growth companies do when the public window shuts: it raised more money privately. A $1.1bn Series G in September 2025 valued the company at $8.1bn.

A $1bn Series H in February 2026 lifted that figure to $23bn, almost a tripling in five months. The proposed $40bn IPO valuation, if achieved, would represent another 74 per cent step-up in roughly three months.

Investors will weigh that climb against the underlying numbers. Revenue of $510m is healthy for a hardware company at this stage, and the OpenAI commitment provides forward visibility that few rivals can match. But Cerebras remains heavily concentrated, both in customer mix and in a single architectural bet.

The wafer-scale design that distinguishes it from Nvidia is also what makes it hard to manufacture at volume, and the prospectus, by most analyst readings, does not yet show the gross-margin profile of a company that has solved that problem.

Cerebras is filing into a market that has been more cautious than the headlines suggest. Several large AI-adjacent listings in late 2025 priced below their indicative ranges, and bankers have privately warned that public investors are no longer willing to pay private-market multiples for unprofitable infrastructure plays. A $4bn raise at $40bn would test that thesis directly.

There is also the question of comparison. Nvidia’s market capitalisation has hovered above $3tn for much of the past year, and any new entrant in AI silicon will, fairly or not, be measured against it. Cerebras’s pitch is that it is not trying to replace Nvidia in training but to complement it in inference, an argument that has the virtue of being both technically defensible and commercially convenient.

If the roadshow goes well, pricing could come in mid-May, according to people familiar with the timetable cited by Investing.com and Bloomberg. If it does not, Cerebras will once again find itself a company whose technology is widely admired and whose path to public markets is less so.

The difference this time is that it has an OpenAI contract on the books and a CFIUS letter in the file.