Cerebras Systems went public on Thursday. The chip company priced its IPO at $185 a share late Wednesday, opened on Nasdaq under the ticker CBRS at $350, briefly halted trading after running above $385, and closed the day at $311.07, up 68% from the offer price. Market capitalisation at the bell was over $66 billion. The deal raised around $5.6 billion, making it the biggest IPO of the year so far.
This is the same company that filed to go public in 2024 and pulled the offering in 2025. Between those two attempts, the questions that scared the first set of bankers (customer concentration, the UAE-connected G42 relationship, and chief executive Andrew Feldman's guilty plea for accounting fraud at a previous company in the 2000s) have not been answered so much as overwhelmed by AI demand.
What changed in eighteen months
Asked by Fortune what was different about Cerebras now versus 2024, Feldman gave a list rather than an argument: bigger customers, more deployments, more mature technology, a better roadmap. The concrete answer is two contracts. In January, Cerebras signed a deal worth more than $10 billion with OpenAI to furnish computing capacity (OpenAI co-founders Sam Altman and Greg Brockman are also significant shareholders). In March, it announced a multiyear partnership with Amazon Web Services on undisclosed terms.
Those are the two anchors. Karl Alomar, managing partner at M13, named the problem out loud in an emailed comment: "Cerebras has proven demand from hyperscalers, but they need to prove they can manufacture at volume, deliver on time, and support multiple customers at the same operational level Nvidia does." He pointed at recent softness in the secondary market for Cerebras shares as evidence the question is still live. "Investors want to see concrete customer wins and production schedules, not just OpenAI."
That is, essentially, the same critique that killed the 2024 attempt, just with bigger names in the customer column.
What Cerebras actually sells
The product is the Wafer-Scale Engine, a single silicon die roughly the size of a children's picture book. It packs four trillion transistors and 44 GB of on-die SRAM, which is the relevant number: keeping model weights and activations in fast memory next to the compute is what makes the WSE good at fast inference, the use case Cerebras is now leaning into rather than training.
Feldman's pitch on customer fit is three layers: AI-native startups and labs that use the chips through Cerebras's cloud, enterprise customers like banks that will access them through AWS data centres, and overseas government and private buyers who want the silicon on premises. "You can switch from a workload on Nvidia to a workload on Cerebras in about 10 keystrokes," he told the Wall Street Journal. That is a statement about software portability, not architecture parity. It is also the kind of claim that gets tested in the first deployment, not the IPO roadshow.
The Benchmark case
Eric Vishria, the Benchmark general partner who backed Cerebras early, offered the bullish frame: demand for compute and inference is not slowing, and once it exists, fast tends to win. "How big is the market for slow anything? Slow internet, slow search, slow streaming, ultimately everything gets really fast, and fast tends to win." The implication is that an inference chip optimised for latency is a structural bet on a category that has not been priced yet, rather than a substitution play against Nvidia.
Feldman put a number on it that is intentionally ridiculous: 47 million software engineers globally, $100,000 of tokens a year each, $5 trillion from one use case. He is not claiming Cerebras will capture that. He is claiming the addressable market for fast inference is so loosely bounded that conventional sizing exercises are not load-bearing. That is a useful pitch for an IPO. It is also exactly the kind of framing that becomes a liability if the second wave of customers is slow to land.
The Figma comparison
Software firm Figma jumped 250% on its IPO last year and is now down more than 80% from that first-day peak. The Cerebras pop is smaller in percentage terms but the structural risk is the same: when a deal prices at $185, opens at $350, and closes above $311, the bankers either left billions on the table or the public market is doing price discovery the private market refused to do. The two readings have very different implications for where the stock settles in twelve months.
Many institutional investors who wanted shares in the offering got nothing, according to people familiar with the deal, forcing them to buy in the public market and adding to the first-day pressure. Allocations to funds that did get in were materially smaller than requested. That dynamic helps the print on day one. It does not help the float on day sixty.
What today did not resolve
Cerebras now has a public-market valuation that requires it to scale a manufacturing and customer pipeline that has historically been the company's weakest argument. The 2024 concerns (concentration, G42, Feldman's history) have not been retired. They have been temporarily outvoted by the OpenAI deal, the AWS partnership, and an AI-IPO window that is about to widen further. OpenAI and Anthropic are both eyeing public offerings in the second half of this year. xAI, which Elon Musk's SpaceX recently acquired, is aiming for June.
If Cerebras lands two or three more named hyperscaler-scale customers in the next twelve months and ships on schedule, the IPO will look underpriced. If the next customer takes longer than that to surface, or if the OpenAI capacity ramp slips, the stock will retrace into a single sentence about a $66 billion valuation built on one contract and a story. The first day told us almost nothing about which of those is the right read. It just told us the market wanted to find out.
