Cerebras Systems: A Decade-Long Rise of the Wafer-Scale Chip
The Startup Story: From "Impossible" to Reality
Cerebras Systems was born from a bold 2015 idea: build a wafer-scale chip dozens of times larger than a traditional GPU to break through AI compute bottlenecks. At the time, this was considered impossible—wafer-scale integration (WSI) had never been successfully commercialized in 75 years. CEO Andrew Feldman said the founding goal was to "do what others had tried and failed at."
After four years of stealth development, Cerebras unveiled the Wafer Scale Engine 1 (WSE-1) in 2019: 1.2 trillion transistors across 46,225 mm², the largest chip ever made, making Cerebras the first company to commercialize wafer-scale processors. Iterations followed:
- WSE-2 (2021): 7nm process, 2.6 trillion transistors
- WSE-3 (2024): 5nm process, 4 trillion transistors, 125 PFLOPS of AI compute
- Hardware and cloud: Direct sales of CS-series systems, plus Cerebras Cloud (on-demand or dedicated capacity), with availability through AWS and Azure marketplaces. In 2025, Cerebras signed a supercomputing deal with OpenAI potentially worth up to $30 billion: 750 MW of inference capacity with an option for an additional 1.25 GW—one of the largest AI infrastructure deals ever.
- Products and services: The flagship CS-3 delivers 125 PFLOPS peak; multiple racks interconnect into AI supercomputers. Cerebras also offers end-to-end services (data prep, model design, training, inference optimization). Its top ten customers increased spending by an average of 80% within 12 months.
- Financials: In 2025, revenue reached $510 million (+76% YoY) with a net profit of $87.9 million—a rare profitability milestone for a chip startup. The February 2026 Series H valued the company at $23 billion; the IPO aims to raise an additional $200 million.
- Ecosystem lock-in: CUDA's developer moat is deep; Cerebras mitigates this with PyTorch compatibility.
- Capacity and cost: WSE-3 can currently only be produced on TSMC's 5nm process; yield and cost will determine price competitiveness.
- Competition: NVIDIA, AMD, Intel, and startups like Groq and SambaNova are all pursuing AI accelerators.
The road was not smooth. Cerebras' first IPO filing in 2024 was withdrawn due to a CFIUS review tied to investment from the UAE's G42 Group, but the company refiled its S-1 in April 2026, targeting a Nasdaq listing. Customers include the Mayo Clinic and AstraZeneca.
Technical Barriers: The Wafer-Scale Breakthrough
The core innovation is treating an entire wafer as a single chip. The WSE-3 measures 46,225 mm²—58× the area of NVIDIA's flagship B200—with 900,000 AI cores, 44GB of on-chip SRAM, and 21 PB/s memory bandwidth (2,625× the B200's packaged bandwidth).
Two key breakthroughs:
1. Extreme memory bandwidth and capacity: By co-locating compute and storage on one wafer, WSE largely sidesteps the "memory wall," eliminating frequent external memory reads. Cerebras claims CS-3 delivers up to 15× the inference speed of conventional GPUs.
2. Fault tolerance and yield: Cerebras divides the wafer into many small dies interconnected at high speed. If a die is defective, traffic is routed around it—so 100% yield is not required. Dozens of faulty dies can be tolerated on a single WSE-3 wafer, an impossibility in conventional chip design.
A companion software stack (CSoft compiler, inference serving, cluster management) lets developers map models onto the WSE without writing CUDA or managing distributed clusters.
Business Model: Hardware Plus Services
Reshaping the AI Chip Landscape
Cerebras is challenging NVIDIA's dominance in AI inference: CS-3 is over 2× faster than Blackwell GPUs on Llama 4 inference and up to 5× faster on GPT-OSS-class models, prompting OpenAI to shift its fast-inference business from NVIDIA to Cerebras.
Challenges remain: