Cerebras' Wafer-Scale Breakthrough to Nasdaq

In 2016, Andrew Feldman and Gary Lauterbach founded Cerebras Systems in Los Altos, California, with a plan to overturn the limits of traditional chipmaking for artificial intelligence. The company began with a bold bet: instead of linking together thousands of smaller chips, they would create the world’s first wafer-scale processor, building a single, massive chip from an entire silicon wafer. Feldman and Lauterbach’s vision was to design hardware that could handle the explosive growth in neural network size and complexity, a problem that standard graphics processing units struggled to keep up with.

The Cerebras founding team had previously worked together at SeaMicro, a company acquired by AMD. Their experience there shaped their understanding of why communication between chips was such a limiting factor in high-performance computing. The wafer-scale approach meant fabricating a chip more than 46,000 square millimeters in area—about 56 times larger than the largest standard graphics card chip. Standard chips are built by slicing silicon wafers into many small dies; Cerebras instead kept the entire wafer as a single, unified processor, the Wafer Scale Engine (WSE). The first version, the WSE-1, was unveiled in 2019, containing 400,000 AI-optimized cores and 1.2 trillion transistors.

The technical challenge of creating such a large chip was considered impossible by many in Silicon Valley. Traditional wisdom held that defects would make a wafer-scale chip unreliable, as a single flaw could render the entire processor useless. Cerebras engineered around this by building redundancy directly into the chip, allowing defective cores to be bypassed, so the processor as a whole could continue running even with some faults present. They also developed their own interconnects and memory systems to keep all the cores fed with data, overcoming latency that would cripple smaller chips trying to work together over circuit boards.

Cerebras’s first systems, the CS-1 and later the CS-2, were the size of a refrigerator and powered by the WSE. Each CS-2 used a single wafer-scale chip, drawing 15 kilowatts of power—roughly 10 times more than a typical server, but delivering AI training performance that would otherwise require dozens of racks and thousands of GPUs. Their customers included government research labs and cloud providers seeking to train very large models. In 2021, Argonne National Laboratory and Lawrence Livermore National Laboratory announced they were using Cerebras systems to accelerate research in computational science and AI. The reason these labs sought out Cerebras was the chip’s ability to process entire neural networks in memory at once, rather than splitting data across hundreds of chips.

To fund the development of their wafer-scale technology, Cerebras raised more than $720 million in venture capital over several rounds. Their lead investors included Benchmark, Altimeter Capital, and Coatue Management. The funding enabled them to create specialized manufacturing partnerships with TSMC, the world’s largest semiconductor foundry. TSMC agreed to fabricate the massive wafers using its most advanced process nodes, a logistical challenge since standard semiconductor packaging lines were not built to handle single chips of this scale.

In early 2024, Cerebras filed paperwork to list its shares on Nasdaq, aiming to capitalize on a wave of investor enthusiasm for artificial intelligence hardware. The initial public offering raised $5.55 billion, and the stock opened at $350 per share, 89% above the IPO price of $185. The initial market cap was nearly $40 billion. The IPO was the largest of the year, and it drew significant attention from institutional investors.

The IPO was timed amid record demand for AI chips, as competitors like Nvidia and AMD struggled to fill orders. Cerebras positioned itself as an alternative for customers unable to acquire sufficient GPU clusters, and the company began selling complete AI “supercomputers”—racks of CS-2 systems networked together—to major cloud providers and international research centers. Cerebras formed partnerships with technology groups in the United Arab Emirates, such as G42, to expand its presence in international markets and deliver large-scale AI compute infrastructure. The scale of these systems rivaled installations at Meta and Google, leveraging the wafer-scale architecture.

Cerebras’s rise was propelled by the software tools it developed alongside the hardware. The company launched the Cerebras Software Platform, which allowed AI researchers to port existing PyTorch and TensorFlow workloads directly onto the wafer-scale engine, minimizing the friction of switching from GPU-based systems. This capability was critical for commercial customers, who could not afford to rewrite entire codebases to leverage new hardware. The software also handled automatic mapping of neural networks onto the wafer’s hundreds of thousands of cores, abstracting away the complexity of the underlying silicon.

By mid-2025, Cerebras reported that its systems were being used to train large language models in hours rather than days. In one demonstration, a single CS-2 system trained a GPT-class model with 1.3 billion parameters in under 12 hours. Comparable training jobs on traditional GPU clusters took more than 30 hours, due to the communication overhead between chips and nodes. This performance advantage was made possible by the wafer-scale design, which eliminated the need to split up neural networks and kept all model data in close physical proximity on a single piece of silicon.

The company’s latest generation chip, the WSE-3, increased the number of AI-optimized cores to more than 800,000 and boosted on-chip memory to 40 gigabytes. This allowed models with over 100 billion parameters—larger than all but the most advanced commercial AI systems—to be trained or deployed on a single wafer. Early customers for the WSE-3 included pharmaceutical companies, weather modeling agencies, and financial institutions seeking to run real-time prediction engines. These organizations were drawn by the ability to experiment with much larger models and datasets without the scaling challenges and power costs of traditional supercomputers.

Cerebras’s bet on wafer-scale technology created a new class of AI infrastructure. In one contract, the company delivered a cluster of 16 CS-2 systems to a national research lab, collectively capable of processing more than 1.3 exaflops of AI compute—an order of magnitude greater than many of the top-ranked supercomputers in the world. The mechanism behind this leap was the integration of massive numbers of compute cores and memory resources on a single die, reducing both latency and inter-chip communication costs.

With the Nasdaq listing, Cerebras secured both funding and global attention, setting a precedent for other semiconductor startups targeting the AI hardware market. The company’s founders spent nearly a decade pursuing a vision most of the industry considered impossible, risking hundreds of millions of dollars and years of secretive development to produce the only commercial wafer-scale chip in existence. Their success in overcoming defect tolerance, power delivery, and software compatibility set new benchmarks for what could be achieved with silicon engineering. At launch, the WSE-3 measured more than 55,000 square millimeters—making it not only the largest chip ever sold, but the largest single piece of computing silicon ever manufactured for commercial use.