Cerebras IPO Pops 68% on Day One, Then Drops 10%. What the Wild First Week Actually Means
Cerebras Systems priced at $185, opened at $350, closed its first day at $311, then slid 10% on day two. The AI chip maker's debut was the biggest U.S. tech IPO since Uber. Here's what the numbers say about the AI hardware bet.
Cerebras Systems went public on May 14 and put on a show. The AI chip maker priced at $185 per share, opened at $350, hit $385 before a trading halt, and closed at $311.07, a 68% gain. By day two, it had slipped to $279.72, down 10%. Welcome to the public markets.
The $5.55 billion raise makes it the largest U.S. tech IPO since Uber in 2019. With the greenshoe option fully exercised, that number climbs past $6.3 billion. At the day-one close, Cerebras was worth about $68 billion.
Those are the numbers. The question is what they mean.
What Cerebras actually makes
Cerebras builds the WSE-3 (Wafer Scale Engine), a single chip about 58 times larger than NVIDIA's B200. Where NVIDIA strings thousands of smaller chips together, Cerebras packs an entire cluster's compute onto one piece of silicon. The WSE-3 has about 4 trillion transistors on a single piece of silicon, compared to roughly 208 billion on an NVIDIA B200, eliminating the inter-GPU communication lag that bottlenecks inference on very large models. For inference, Cerebras claims 15x speedups over comparable GPU setups. For training, the advantage is less clear since CUDA frameworks have been optimized for distributed training over a decade.
OpenAI signed a multiyear deal with Cerebras worth over $10 billion for compute capacity, plus a $20 billion chip purchase agreement. An endorsement like that turns a speculative hardware bet into a real business. The downside: every AI framework is built for NVIDIA first, and Cerebras means rewriting software to its platform.
The actual financials
Cerebras is one of the rare AI hardware companies already turning a profit. Revenue for 2025 hit $510 million, up 76% year over year, with net income around $238 million and a 47% net margin. Three years ago, it booked $24 million.
The question for public investors is whether the 133x trailing earnings multiple leaves room for upside. NVIDIA trades at roughly 35x. The most important risk in the S-1: OpenAI accounted for 83% of 2025 revenue. That is a dependency, not a customer relationship. If OpenAI builds more internal chip capability or renegotiates, Cerebras loses the majority of its business overnight.
The NVIDIA question, post-IPO
NVIDIA's market cap sits above $5 trillion. Every major AI lab builds on its hardware. Cerebras isn't trying to dethrone NVIDIA. It's going after a specific slice: running the largest models at inference time, where communication overhead between GPUs becomes the bottleneck. The OpenAI deal shows the advantage is real.
The bull case: demand for AI compute is growing so fast that even a distant second place becomes a massive business. The bear case: the CUDA software moat has killed every challenger before, and "just buy more NVIDIA" is the safe choice. The middle case: Cerebras needs to convert enough inference-heavy workloads to grow revenue from $510 million toward $3-5 billion, which means diversifying beyond the 83% OpenAI revenue concentration to at least two more hyperscale customers. The technology makes this plausible. Whether the sales team can execute is a separate question.
What the 20x oversubscription actually signals
Cerebras priced at $185 after originally marketing the deal at $150 to $160. Demand was 20x oversubscribed. They upsized multiple times and still popped 68% on the open.
That level of demand tells you where institutional money thinks AI returns will come from next. The first wave went into foundation model companies. The second wave, accelerating now, is going into infrastructure: chips, data centers, power generation. Foundation model companies may or may not build sustainable businesses, but they all need chips. Betting on the suppliers is the picks-and-shovels play.
The 20x oversubscription also signals that institutional investors want NVIDIA alternatives for diversification. NVIDIA at $5 trillion with 75% margins has limited room for multiple expansion. Cerebras at $68 billion with 76% revenue growth has a different risk-reward profile. The demand is less "Cerebras will beat NVIDIA" and more "the AI chip market is big enough for more than one winner."
What this means for other AI chip startups
The Cerebras IPO opens a window for Groq, d-Matrix, SambaNova, and Graphcore. The stock held well above the $185 offering price even after the day-two pullback. If Cerebras maintains a market cap above $50 billion through its first earnings report, expect at least two more AI chip S-1 filings before the end of 2026. Groq is the most likely next candidate, with an inference-focused chip and a cloud API that lets developers try the hardware without buying it.
What I think after watching the first week of trading
I follow AI hardware companies closely, and the Cerebras IPO is the most revealing data point I have seen about public market appetite for AI chip investments. The 20x oversubscription, 68% pop, and controlled pullback rather than a crash is about as healthy a debut as a hardware company can hope for. The market wants an NVIDIA alternative and is willing to pay a premium without treating CBRS like a speculative frenzy. That is genuine institutional conviction.
What the first week does not tell me is whether Cerebras can execute. Going from $510 million to a $20 billion agreement is an entirely different category of challenge. Manufacturing wafer-scale silicon, managing the supply chain, building customer support. None of these scale linearly. If Cerebras executes well, the valuation holds. If it stumbles, the market reprices fast.
What retail investors should watch
I am not giving investment advice, but here is what the prospectus and first-week action suggest you should track. Revenue concentration is the number one risk: anything that affects OpenAI's business affects Cerebras disproportionately. Watch for announcements about new enterprise customers that diversify the revenue base. Execution risk is number two: scaling from $510 million in revenue to fulfilling a $20 billion chip agreement while maintaining 47% margins is a genuinely hard operational challenge. Competitive risk is number three: if inference workloads shift toward smaller, specialized models rather than larger general ones, the wafer-scale advantage diminishes.
The first two quarters as a public company will tell the story. If growth stalls or the OpenAI dependency deepens, the 133x multiple becomes impossible to justify. The next earnings call is the one that matters.
Beyond the immediate quarters, the long-term thesis depends on a structural trend. AI inference workloads are shifting from single-turn chatbot queries to multi-step agentic workflows that consume far more compute per user interaction. If that holds, the inference market grows faster than training, making wafer-scale architectures proportionally more valuable. The bet on CBRS is ultimately a bet on the inference market growing faster than NVIDIA's ability to supply it. Based on the Blackwell Ultra backlog, that is not an unreasonable assumption.