We are faster at creating tokens. We are only at the beginning.
Cerebras raised $5.5B at $185/share (well above $115-125 range) with 25x oversubscription, closing at $311 after peaking at $385 on opening day. The company's WSE-3 chip is 60x larger than conventional AI chips, 15x faster than competitors, and can replace entire Nvidia H200 clusters while being more power-efficient.
- Cerebras raised $5.5B at $185/share, 48% above the original $115-125 range, with 25x oversubscription
- WSE-3 chip contains 4 trillion transistors, is 60x larger than conventional AI chips, and 15x faster than competitors
- Company generated $510M in revenue in 2025 (76% growth), with $88M profit, up from $482M loss in 2024
- Market cap reached nearly $70B; company depends on two major clients: OpenAI ($20B contract) and G42
Cerebras Systems, an AI chip manufacturer positioning itself as a Nvidia competitor, debuted on Nasdaq with a 68% gain, raising $5.5B and achieving a $70B market cap on strong investor demand for AI exposure.
Cerebras Systems opened for trading on the Nasdaq this morning with a 68% surge, marking the largest initial public offering of the year and signaling sustained investor appetite for artificial intelligence infrastructure plays. The chip manufacturer, which has positioned itself as a potential challenger to Nvidia's dominance, priced its shares at $185 each—a sharp jump from the originally targeted range of $115 to $125. Demand was so intense that the offering was oversubscribed 25 times over. The stock more than doubled at the opening bell, reaching $385 before settling into a close at $311, giving the company a market capitalization of nearly $70 billion, roughly equivalent to General Motors or Vale.
The company's core product is the Wafer-Scale Engine, or WSE-3, a processor roughly the size of a small notebook that contains 4 trillion transistors and 44 gigabytes of onboard memory. What distinguishes it from conventional AI chips is both its scale and its efficiency. The device is 60 times larger than standard artificial intelligence processors and performs inference—the computational work of delivering results—and model training at speeds that are dozens of times faster. According to the company's leadership, a single WSE-3 can accomplish tasks that would otherwise require an entire cluster of Nvidia's H200 chips, each of which is roughly the size of a 3-by-4 photograph and contains 80 billion transistors. CEO and cofounder Andre Feldman told Bloomberg that Cerebras is 15 times faster than its nearest competitor, and he framed the company's advantage in terms of the expanding demands of artificial intelligence itself. As AI systems become more capable and more widely deployed, they require processing of exponentially more data units, or tokens. "We are only at the beginning of AI utilization," Feldman said. "The more useful AI becomes, the more tokens are needed. And we are faster at creating tokens."
The efficiency gains matter beyond raw speed. Computational capacity and electrical supply have emerged as two of the most significant bottlenecks limiting the global expansion of artificial intelligence and the training of new systems. Chips that consume less power while delivering more throughput could help overcome both constraints. Cerebras was founded in 2015 in Silicon Valley and has grown substantially. Last year the company generated $510 million in revenue, a 76 percent increase from 2024, and reported a profit of $88 million after losing $482 million the year before. In February, just months before the public offering, Cerebras raised $1 billion in a Series F funding round led by Tiger Global Management, which valued the company at $23 billion on a post-money basis.
The investor roster reflects both established venture capital and prominent figures from the artificial intelligence world. Fidelity is now the largest shareholder with 11.3 percent of Class B shares. Benchmark, the venture capital firm that was among the earliest backers, has seen its investment multiply by 12 times at the IPO price. Angel investors include Sam Altman and Greg Brockman from OpenAI. Nvidia itself made a defensive move late last year, spending $20 billion to license technology from Groq, a startup that designs chips similar to Cerebras's approach.
Yet beneath the enthusiasm lies a structural vulnerability that analysts and investors are beginning to scrutinize. Cerebras currently relies on a very small number of major customers. The company has two primary clients: G42, an artificial intelligence company created by Abu Dhabi in partnership with Microsoft, and OpenAI, which recently signed a contract worth up to $20 billion with Cerebras. The company also has partnerships with Amazon Web Services, Meta, and Mistral. Vikram Sekar, a semiconductor engineer and founder of SemiExponent, framed the central question for investors this way: "The issue for investors is not whether Cerebras affects Nvidia, but whether the fully customized architecture can generate revenue beyond strategic agreements with OpenAI." The Wall Street Journal has identified this customer concentration as one of the primary uncertainties hanging over the company's future.
The successful debut nonetheless reflects the intensity of investor demand for exposure to new artificial intelligence opportunities. Several major public offerings remain on the calendar for later this year, including those of OpenAI and Anthropic, both expected in the second half, and SpaceX, anticipated in June. Cerebras has demonstrated that the market will reward companies that can credibly claim to solve the infrastructure constraints that are currently limiting artificial intelligence's expansion.
Notable Quotes
We are only at the beginning of AI utilization. The more useful AI becomes, the more tokens are needed. And we are faster at creating tokens.— Andre Feldman, CEO and cofounder of Cerebras
The issue for investors is not whether Cerebras affects Nvidia, but whether the fully customized architecture can generate revenue beyond strategic agreements with OpenAI.— Vikram Sekar, semiconductor engineer and founder of SemiExponent
The Hearth Conversation Another angle on the story
What makes this chip actually different from what Nvidia is already selling?
The WSE-3 is built as a single, unified processor rather than a collection of smaller chips that have to talk to each other. That means less time wasted moving data around. It's 60 times larger in terms of transistor count, and because everything is on one piece of silicon, the efficiency gains compound. Nvidia's approach requires clustering multiple chips together, which introduces latency and power overhead.
So why hasn't Nvidia just copied this approach?
They're trying. The $20 billion licensing deal with Groq suggests Nvidia sees the threat and is hedging. But there's a difference between understanding a design and being able to manufacture it at scale. Cerebras has spent a decade perfecting the engineering. Nvidia's strength has always been in software ecosystem and market dominance, not necessarily in pushing the physical limits of chip design.
The stock jumped to $385 and then fell to $311. What does that tell us?
It tells us there's real enthusiasm but also real uncertainty. The opening surge is typical of oversubscribed IPOs—people buying on momentum. The pullback suggests some investors took profits, or perhaps some realized the customer concentration issue is more serious than the opening-day euphoria acknowledged.
Two customers. That's the real story, isn't it?
It is. OpenAI and G42 represent the bulk of revenue. That's not a business model yet—that's a proof of concept. If Cerebras can't diversify its customer base, it becomes a supplier to a handful of players rather than an infrastructure company. The market is betting they can, but that bet is unproven.
Why would other companies hesitate to buy from them if they're genuinely faster and more efficient?
Switching costs are real. Companies have built their entire AI operations around Nvidia's ecosystem. There's software, training, institutional knowledge. Cerebras requires rewriting code, retraining teams. And there's the risk calculus: Nvidia is a $3 trillion company. Cerebras is new. If something goes wrong, who do you blame? That's a harder sell than it sounds.