$48.8 billion. That’s the valuation Cerebras Systems will carry when it starts trading on the Nasdaq tomorrow. The order book closed 20 times oversubscribed. Wall Street is calling it the hottest IPO of 2026. The “Nvidia killer” narrative is reaching fever pitch.
And buried on page 147 of the S-1 filing, there’s a line that tells you everything you need to know about the AI chip market: OpenAI holds warrants for 33.4 million shares of Cerebras stock at an exercise price of $0.00001 per share. That’s 10% of the company. For essentially nothing.
Let that number sit for a second. The company that’s supposed to break Nvidia’s monopoly on AI chips just gave away a tenth of itself to its largest customer. That’s not a partnership. That’s a terms sheet you sign when you have no other option.
The Numbers Behind the Narrative
On the surface, Cerebras looks like the anti-Nvidia success story the market has been waiting for. Revenue jumped from $290 million in 2024 to $510 million in 2025, a 76% year-over-year leap. The wafer-scale chips are real technology, not vaporware. AWS signed on to deploy CS-3 systems through Amazon Bedrock. The inference-first positioning is smart because that’s where the AI workload is shifting as models mature.
But peel one layer deeper and the picture gets uncomfortable.
In 2024, a single customer, Abu Dhabi’s G42, accounted for 85% of Cerebras’s revenue. Eighty-five percent. In 2025, that concentration shifted but didn’t disappear. Mohamed bin Zayed University of Artificial Intelligence represented 62% of revenue, and G42 still contributed 24%. That means two customers in the same country generated 86% of revenue in a year Cerebras is using to justify a $49 billion valuation.
And the company isn’t profitable. The GAAP net loss was $75.7 million in 2025. The $237 million “profit” number you might see floating around includes a $363 million non-cash gain from restructuring the G42 deal. Strip that out and the operating loss was $146 million.
The Warrants Problem
Here’s where the Cerebras story becomes a founder lesson.
OpenAI didn’t just become Cerebras’s biggest future customer. It became its financial lifeline. The $20 billion multi-year deal is massive, yes. But it came with strings that no truly independent company would accept: warrants for 10% of the company at a near-zero exercise price, plus a $1 billion working capital loan tied to accelerating OpenAI’s chip deliveries.
Those warrants are worth roughly $5 billion at the IPO midpoint. That’s approximately half the gross profit Cerebras expects to earn on the entire OpenAI deal. In other words, Cerebras is doing $20 billion of work to keep maybe $5 billion in profit, and then handing $5 billion of that right back to the customer in equity.
If you’re a founder and your biggest customer demanded half your gross margin back in equity before they’d sign, you’d call that a hostage negotiation. When it’s an IPO on the Nasdaq, Wall Street calls it a strategic partnership.
We saw the mirror image of this dynamic just last week, when Nvidia spent $40 billion investing in the same companies buying its chips. The AI chip market is becoming a circular economy where the money flows in a loop between buyer and seller, and nobody’s quite sure who’s subsidizing whom.
The “Nvidia Killer” Trap
The deeper pattern here is structural, and it affects every startup founder betting on AI chip alternatives.
Every serious Nvidia challenger, Cerebras, Groq, custom silicon from Google and Amazon, faces the same bootstrapping problem: you need a massive anchor customer to prove the chips work at scale. But the kind of customer who can absorb that much silicon is, by definition, one of the five or six companies big enough to dictate terms. So you end up giving away economics that would make your public shareholders uncomfortable, because the alternative is having no revenue at all.
Cerebras went from 85% dependence on G42 to building its future on OpenAI. The customer concentration didn’t decrease. It just got a more famous name. And the price of that famous name was 10% of the company.
For founders building on AI infrastructure, this matters because the “Nvidia alternatives are coming” narrative might be oversimplified. The alternatives aren’t arriving as independent competitors. They’re arriving as captive suppliers to the same handful of hyperscalers that already control the market. If your cloud provider offers you Cerebras chips through Bedrock, that’s not competition breaking out. That’s the same monopoly wearing a different jersey.
What This Actually Means for Founders
If you’re building an AI startup and hoping the chip market diversifies so your inference costs drop, here’s the honest read.
Short term, Cerebras going public is good news. More silicon in the market means more compute supply, and more supply means downward price pressure. OpenAI using Cerebras for code-generation inference means the wafer-scale approach works for at least one production workload.
But medium term, watch the ownership structure, not the press releases. When China is building its own chip supply chain because Nvidia dependency is a national security risk, and Cerebras is giving away 10% of itself to lock in a single American customer, the chip market isn’t diversifying. It’s fragmenting into captive ecosystems.
The question for your startup isn’t “will chips get cheaper?” They probably will. The question is: “who owns the chipmaker my cloud provider is offering me, and what does that mean for my switching costs three years from now?”
Because right now, the answer to “who’s the Nvidia alternative?” is increasingly: “your own customer, wearing a supplier hat.”
The Bet Wall Street Is Making
Tomorrow’s IPO will almost certainly pop. Twenty times oversubscribed, AI hype still running hot, and “the Nvidia alternative” is a narrative that practically sells itself. Cerebras might trade at $60 billion or higher by the end of the week.
But the S-1 tells you exactly what you’re buying: a company that does excellent engineering, has a real product, and has structured itself as a dependent of whichever single customer it can land. G42 yesterday, OpenAI today. The chips are real. The independence is not.
And in an industry where every founder is taught that customer concentration above 30% is a red flag, Cerebras is going public at $49 billion with numbers that would get a Series A deck thrown out of any VC’s office.
The AI chip market needs more players. It needs real competition to Nvidia. But if the only way to compete is to hand 10% of your company to your first big customer, then we’re not building competition. We’re building a different kind of dependency. And founders betting their infrastructure costs on that “competition” should know what they’re actually betting on.
