Cerebras Systems was founded in 2015 by Andrew Feldman, Gary Lauterbach, Michael James, Sean Lie, and Jean-Philippe Fricker, five engineers who previously worked together at SeaMicro. They launched the company with a specific goal: to build an artificial intelligence processor at wafer scale instead of cutting silicon into separate dies. This architectural choice meant designing around an enormous interconnected processor spanning most of a wafer, a move that challenged the industry standard of using graphics processing units for deep learning tasks. The team believed that GPUs were not optimal semiconductors for high-level processes because they were originally designed for graphics rather than neural network training.
Engineering the Impossible Chip
The path to a functional product was expensive and technically grueling. The company spent $200 million trying to solve integrated circuit packaging problems, including adhering silicon to a motherboard and managing heating and cooling pipes. It burned through $8 million per month during this phase, which required unique software to route around microscopic manufacturing defects and a machine capable of drilling 40 screws into the wafer simultaneously without cracking it. In July 2019, after exhaustive trial and error, the team finally produced a working product. They announced the first-generation Wafer Scale Engine (WSE) and the CS-1 supercomputing system in August 2019. The CS-1 is a 19-inch rack-mounted appliance containing a single WSE primary processor with 400,000 processing cores and 1.2 trillion transistors. Early customers included GlaxoSmithKline, AstraZeneca, and the National Energy Technology Laboratory, who used the systems for drug discovery and computational fluid dynamics.
I find the transition from hardware novelty to practical utility particularly compelling here. The founders faced systemic challenges that no one had solved before, such as maintaining electrical continuity across tens of thousands of interconnect points on a chip 58 times larger than conventional designs. Their first prototype wafer smoked upon initial power-up, an event the team dubbed a "thermal event." They brought in failure-analysis engineers to confirm their power density calculations were audacious but physically possible. By April 2021, the company released the CS-2 system based on the WSE-2, which has 850,000 cores and 2.6 trillion transistors manufactured by TSMC’s 7 nm process. The WSE-2 expanded on-chip SRAM to 40 gigabytes and memory bandwidth to 20 petabytes per second.
Scaling to Market Dominance
Cerebras expanded its footprint significantly after proving the technology works. In November 2022, the company unveiled the Andromeda supercomputer, which combines 16 WSE-2 chips into one cluster with 13.5 million AI-optimized cores. This system delivers up to 1 exaflop of computing horsepower while consuming only 500 kW, a drastically lower amount than comparable GPU-accelerated supercomputers. The company raised $250 million in a Series F round in November 2021, valuing it at over $4 billion. By March 2024, Cerebras introduced the WSE-3 architecture, a 5nm-based chip hosting 4 trillion transistors and 900,000 AI-optimized cores. This generation provides twice the performance of the CS-2.
| System | Cores | Transistors | Memory/Specs |
|---|---|---|---|
| WSE-1 | 400,000 | 1.2 trillion | 18 GB memory |
| WSE-2 | 850,000 | 2.6 trillion | 40 GB SRAM |
| WSE-3 | 900,000 | 4 trillion | 5nm process |
In May 2024, researchers simulated 800,000 atoms interacting on the WSE-2, calculating femtosecond increments in microseconds. The company beat NVIDIA’s Blackwell in Llama 4 inference testing in May 2025, achieving more than 2,500 tokens per second compared to 1,000 for Blackwell. A major financial milestone arrived in January 2026 when Cerebras signed a deal with OpenAI to deliver 750 megawatts of computing power through 2028 for $10 billion. As part of this agreement, the company is temporarily prohibited from selling products to Anthropic. On May 14, 2026, the company became public via the largest U.S. technology initial public offering since 2019, raising $5.55 billion. The stock closed its first trading day approximately 68% above the IPO price. Does the shift to hosted inference services fully replace the need for on-premise supercomputers?
