Cerebras Sets IPO Price at $185, OpenAI $20 Billion Deal as Trump Card

According to Reuters citing sources with knowledge of the matter, AI chip company Cerebras Systems set its IPO price at $185 per share, above the upper limit of the $150-$160推介 range. A total of 30 million shares were issued, raising $5.55 billion and valuing the company at approximately $56.4 billion on a fully diluted basis, making it the largest stock IPO so far in 2026. The company will officially be listed on Nasdaq today under the ticker symbol CBRS.

The pricing underwent three adjustments. Cerebras initially started the roadshow with a price range of $115-$125 per share. After receiving more than 20 times the subscribed amount, the price range was adjusted to $150-$160, and the number of shares issued was expanded from 28 million to 30 million. The final pricing once again broke through the upper limit.

The core business logic disclosed during the roadshow is a "take-or-pay" contract signed with OpenAI. The contract involves 750 megawatts of electricity, with a total value of over $20 billion, a prepaid payment of $1 billion, and deployment of 250 megawatts in three installments for the years 2026, 2027, and 2028. The 2026 portion will be delivered through the Cerebras cloud platform, with OpenAI having the option to renew the lease for the fourth and fifth years; for the last two installments, OpenAI can choose to buy out the equipment or continue using the cloud services. Additionally, OpenAI has an option for an additional 1.25 gigawatts of capacity. Cerebras stated that the current financial model assumes the most conservative scenario, where all three phases of capacity are delivered through the cloud platform, leading to the longest income spread and the highest capital expenditure.

Cerebras' core competitiveness lies in speed. The company's roadshow materials show that its wafer-scale processor's inference speed is approximately 15 times faster than the fastest GPU, covering models of all sizes and various mainstream models in China and the United States. The technical principle is that for each word generated during inference, all model weights must be transferred from memory to the computing unit, with memory bandwidth determining the upper limit of speed. The HBM memory transfer speed of a GPU is about 8TB per second, while Cerebras' wafer-scale processor reaches 21,000TB/s, which is about 2600 times faster. Its processor area is equivalent to the size of a dinner plate, 58 times larger than the largest GPU, and its memory bandwidth exceeds that of a GPU by about 2500 times.

In terms of financial data, the company's revenue in 2025 was about $500 million, with a gross margin of about 50%. The gross margin for 2026 is expected to drop temporarily due to the in-house data center's capacity not being fully operational, requiring the rental of external capacity to support the fulfillment of OpenAI's contract, which lowers the overall profit margin. The company expects that as its in-house capacity expands and gradually exits the rental model, the gross margin will rise back to a long-term target of over 60%. It is worth noting that over the past four months, Cerebras has increased its pricing by 30%, and the management believes there is still significant room to explore the pricing ceiling.

In terms of supply chain, Cerebras has bypassed the currently most scarce HBM memory and CoWoS packaging, opting for TSMC's 5-nanometer process instead of the scarce 3-nanometer capacity. TSMC is a strategic investor in Cerebras and has secured ample capacity for 2026 to 2027. The company expects its manufacturing capacity to increase by about tenfold in the coming years. Last year, Cerebras was not among TSMC's top 150 customers, but this year it will rank among the top 15.

The main competitor, Groq, had a revenue of about $50 million in 2025 with a negative gross margin, which is significantly different from Cerebras' revenue of $500 million and a 40% gross margin. Against the backdrop of AI inference demand evolving from single Q&A to "agent + inference" chain calls, a single query triggers multiple inference streams, and the amount of computation grows exponentially, the speed advantage of Cerebras will be further amplified in this trend.