One wafer instead of a rack of GPUs
Cerebras builds a single chip out of an entire silicon wafer rather than cutting the wafer into hundreds of separate dies the way every other chipmaker does. Its S-1 filing describes the approach as solving a problem that had sat open for the industry's whole history: the third-generation Wafer-Scale Engine, WSE-3, packs 900,000 cores onto one piece of silicon, which the filing says is 57 times the size and 52 times the core count of the leading commercial GPU. The CS-3 system that houses it is pitched on power efficiency as much as raw size, claiming three times the compute per watt of an eight-way GPU server, and systems can be networked together up to 2,048 at a time over standard Ethernet. [1]
Andrew Feldman, chief executive since founding, led four of the five co-founders over from SeaMicro, a server maker they sold to AMD in 2012. Cerebras incorporated in April 2016 and has shipped three chip generations since: WSE-1 with CS-1 in 2019, WSE-2 with CS-2 in 2021, and WSE-3 with CS-3 in 2024, followed by a CS-4 system built around three WSE-3 Turbo processors in August 2026. The pitch throughout has been an alternative to clustering thousands of Nvidia GPUs: fewer, larger chips instead of many small ones networked together. [1][2]
A business built on one customer
Cerebras sells hardware and a cloud inference and training service, and it competes with Nvidia on a specific claim: that for the models it supports, its systems answer faster than GPU clusters running the same workload. Its own benchmarks put inference on Llama models at ten to twenty times the token throughput of Nvidia's H100 and claim an edge over Blackwell on some workloads, figures that come from the company itself rather than an independent lab. [3]
That business has been unusually dependent on a single customer. The S-1 disclosed that G42, the Abu Dhabi AI and cloud company, accounted for 83 percent of Cerebras's revenue in 2023 and 87 percent in the first half of 2024, and the filing itself warns that losing G42 or a reduction in its orders would materially harm the business. G42 has also been an investor in Cerebras. The company describes actively working to diversify its customer base, and lists partnerships with Meta's Llama team, Perplexity, Mistral AI, Hugging Face and OpenRouter as part of that effort. [1]
The largest tech listing since Uber
Cerebras priced its initial public offering at $185 a share on May 13, 2026, above the range it had already raised, selling 30 million Class A shares on the Nasdaq Global Select Market under the ticker CBRS. The stock opened well above the offer price and closed its first day up roughly 68 percent, putting the company's market capitalization at about $95 billion. Reporting at the time called it the largest US tech listing since Uber's 2019 debut. [4][5]
The IPO followed a run of private rounds that tracked the AI infrastructure boom: a Series E in November 2019 valued the company at $2.4 billion, a Series F in late 2021 topped $4 billion, a Series G in September 2025 raised $1.1 billion at an $8.1 billion valuation, and a further $1 billion round in January 2026 priced higher still. Each step traces the same story the IPO ended: private investors and then public markets both pricing wafer-scale computing as a credible, if still customer-concentrated, alternative to GPU clusters. [2]
Sources
- Cerebras Systems Inc. Form S-1
U.S. Securities and Exchange Commission · 2024
- Company
Cerebras · Sep 9, 2026
- Cerebras Inference
Cerebras · Sep 9, 2026
- Cerebras pops 68% in Nasdaq debut, pushing the AI chipmaker's market cap to $95 billion
CNBC · May 14, 2026
- Cerebras Systems Announces Pricing of Initial Public Offering
Cerebras · May 13, 2026