Three days after suggesting it might target up to $4bn at $40bn, the AI chipmaker has marketed 28 million shares at $115-$125 each, lining the deal up with its February private valuation rather than above it.

Cerebras Systems was preparing to seek up to $4bn at a roughly $40bn valuation in its long-awaited initial public offering. On Monday, the company filed an updated prospectus that priced the deal more cautiously: 28 million shares at a marketed range of $115 to $125 each, raising as much as $3.5bn at the high end.

At that range, the implied valuation lands closer to $26.6bn, well below the figure circulating last week and roughly in line with the $23bn private mark Cerebras carried after its February Series H.

It is, by any reading, a more grounded number. Whether it is grounded enough is the question the order book will answer.

Per the updated prospectus filed on Monday, 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.

The 28-million-share offering carries an underwriter option of an additional 4.2 million shares, which would lift gross proceeds by roughly $525m at the top of the range. CNBC notes that chief executive Andrew Feldman, the company’s co-founder, is not selling shares in the IPO; he will hold 10.3 million shares post-deal, worth as much as $1.28bn at the high end.

The financial profile underlying the offering is, by AI-hardware standards, well-developed. Cerebras’s most recent quarter showed revenue up roughly 76 per cent year-on-year to $510m, with $87.9m in net income.

The company’s Master Relationship Agreement with OpenAI, signed in January, promises up to 750 megawatts of inference capacity through 2028 in a contract Cerebras has previously valued at more than $20bn over its term. That deal, more than any single demonstration of wafer-scale technology, has anchored the IPO narrative since.

Why the valuation came down?

The shift from a possible $4bn raise at $40bn to a marketed $3.5bn at roughly $26.6bn is not, by IPO standards, dramatic. It is also not nothing. There are three plausible readings.

The first is execution. Filing at a wider range and a more sober valuation reduces the risk of an undersubscribed book on launch day, and it gives the deal room to price up if demand is strong. After two years in which AI hardware listings have priced below initial ranges as often as above, that discipline is sensible.

The second is market context. AI software multiples have been compressing through April, with Palantir down roughly 30 per cent year-to-date and Citi cutting price targets across the AI-platform space.

While Cerebras is hardware rather than software, hardware listings have not been immune to that broader recalibration. Pricing inside the now-public reference points for memory and compute is, again, sensible.

The third is concentration risk. Public-market investors, in their diligence on the prospectus, will have noted that OpenAI is the dominant customer in the deal. The economics, if OpenAI executes as planned, are formidable.

If OpenAI’s compute demand recalibrates, even modestly, Cerebras’s near-term revenue trajectory recalibrates with it. That risk does not go away at $26.6bn; it is, however, easier to underwrite at that figure than at $40bn.

What it would mean if it prices

If Cerebras prices at the high end of the range, the offering would be one of the largest US technology IPOs of 2026, ahead of any new entrant in AI silicon since Nvidia’s reignited public market presence three years ago.

Reuters and IFR have flagged the deal as a roughly $3bn raise in Nasdaq IPO terms; the precise number will depend on whether the underwriter overallotment is exercised and on the final pricing. Either way, this is the largest first-day public capital event in dedicated AI training and inference hardware to date.

It would also be a public test of a thesis Cerebras has built carefully over a decade: that its wafer-scale chip, an architecturally radical alternative to packing many smaller GPUs onto a board, is now commercially defensible at scale. The OpenAI contract is the most visible vote of confidence in that thesis. The IPO, if it prices in line, is the second.

The road from filing to first trade is not a formality. Cerebras’s earlier IPO attempt, in 2024, collapsed after the Committee on Foreign Investment in the United States opened a review of G42’s stake in the company; the eventual restructuring of that stake into non-voting shares cleared the way for the current attempt.

Public-market investors will have read that section of the risk factors closely. They will also have noted Cerebras’s customer concentration, the extent to which margin will depend on the cost of advanced packaging, and the strategic question of whether Nvidia’s next-generation accelerators erode the use cases Cerebras’s wafer-scale design currently serves best.

None of those is fatal. They are, however, the substance behind the lower valuation. Cerebras has filed a serious prospectus into a market that is, just at the moment, taking AI hardware more seriously than AI software. The rest is up to the order book.