The $23 Billion Question Cerebras Has to Answer Before Its Stock Ever Trades

If the AI compute gold rush is as transformative as the industry claims, why does Cerebras need public market capital so urgently?

That is the question the company's S-1 filing — submitted April 17 — forces on anyone willing to read past the headline numbers. Cerebras Systems Inc. is seeking a roughly $23 billion valuation on the Nasdaq under ticker CBRS — its second attempt at going public after withdrawing a 2024 filing that collapsed under a U.S. national security review of its Abu Dhabi investor base. The company makes the Wafer-Scale Engine 3, a single-silicon chip with four trillion transistors, 900,000 AI cores, and 21 petabytes per second of memory bandwidth — roughly 58 times larger than Nvidia's B200 GPU. For inference workloads — generating tokens from a trained model, the dominant cost at scale — Cerebras claims speeds up to 25 times faster than GPU clusters. These are not demo numbers. They are not slideware.

But here is the rest of the filing. In 2025, Cerebras generated $510 million in revenue, up 76 percent from $290 million the prior year. That growth is real. What is not real is the $237.8 million in GAAP net income that makes the top line look like a company crossing into profitability. That number is almost entirely a $363.3 million one-time accounting gain from restructuring a liability held by Group 42, the Abu Dhabi technology conglomerate that was both a Cerebras investor and one of its largest customers. Strip out that gain and stock-based compensation, and the company posted a non-GAAP operating loss of $75.7 million — wider than the $21.8 million loss in 2024. The business is losing more money at higher revenue. That is not the trajectory of a company whose moment has arrived.

The concentration in that revenue tells the rest of the story. Two Abu Dhabi entities accounted for 86 percent of 2025 sales: the Mohamed bin Zayed University of Artificial Intelligence contributed 62 percent, and G42 contributed 24 percent. Meanwhile, U.S.-billed revenue declined 34 percent year-over-year, falling from $282.7 million to $187.6 million even as total revenue grew 76 percent. Export controls did not just shape the 2024 IPO withdrawal. They are already reshaping the customer mix in real time.

To offset that concentration, Cerebras executed what the S-1 describes as a Master Relationship Agreement with OpenAI: a commitment to deploy 750 megawatts of inference compute capacity by 2028, with an option to expand to two gigawatts, valued at up to $20 billion over the contract term. OpenAI also advanced Cerebras a $1 billion working capital loan and received warrants for 33.4 million shares — vesting on milestones including full delivery and a market cap above $40 billion — that would give it roughly 10 percent ownership if fully exercised. The world's most prominent AI lab has bet substantially on Cerebras succeeding. That validation is real. But it has not yet converted to recognized revenue; it sits in backlog alongside $24.6 billion in contracted future performance obligations against a $510 million annual run rate.

The company also signed a binding term sheet with Amazon Web Services to integrate Cerebras CS-3 hardware into the Bedrock managed inference service, pairing AWS's Trainium3 prefill chip with Cerebras decode acceleration. If the integration ships at scale through a mainstream cloud API, developers who cannot operate custom silicon infrastructure gain access to a machine that runs 70-billion-parameter models at roughly 2,000 tokens per second versus approximately 130 tokens per second on a comparable H100 cluster. That is a different kind of bet than buying access to a sovereign AI university in Abu Dhabi.

The timing of this IPO is not incidental. Cerebras is a company with a genuine engineering lead, a landmark customer, and a manufactured profit, racing to prove it can operate at hyperscale before Nvidia closes the performance gap, before TSMC allocates more wafer capacity to its single 5nm process node, and before the next geopolitical tremor forces another CFIUS review. The $23 billion valuation prices in flawless execution across manufacturing, power infrastructure, and deployment timelines — none of which the company has demonstrated at anything approaching this scale.

Cerebras is a hardware story. It is also a geopolitics story, a manufacturing story, and a capital allocation story. The interesting question is whether those things can coexist long enough for the thesis to matter.