The artificial intelligence boom has produced no shortage of ambitious chip startups, but few have managed to challenge the dominance of Nvidia in any meaningful way. One exception is Cerebras Systems, which this week completed one of the biggest technology IPOs of 2026, achieving a market valuation of more than $66 billion on its first day of trading.
The California-based company raised $5.55 billion after listing on the NASDAQ under the ticker CBRS, capping off more than a decade of high-risk engineering decisions centred around its unconventional wafer-scale AI processors.
Cerebras built an AI chip unlike anything else
Founded in 2015 by former SeaMicro executive Andrew Feldman, Cerebras took a radically different approach to semiconductor design at a time when most AI hardware firms were focused on increasingly powerful GPUs.
Instead of manufacturing multiple small chips from a silicon wafer and linking them together later using high-speed interconnects, Cerebras decided to use the entire wafer itself as a single processor.
The result was the Wafer-Scale Engine, or WSE — a processor measuring 46,225 square millimetres, roughly comparable in size to a dinner plate.
A major gamble on wafer-scale computing
At the time, conventional high-end GPUs measured roughly 800 square millimetres each and relied heavily on technologies such as NVLink to connect multiple processors into larger AI systems.
Cerebras argued that reconnecting smaller chips after fabrication introduced unnecessary complexity and bottlenecks. By building a single enormous processor instead, the company aimed to deliver vastly greater bandwidth and lower latency for AI workloads.
Its first-generation WSE processors were designed specifically for deep learning and machine learning training. A major innovation was their use of “sparsity” — a technique that takes advantage of the fact many neural network parameters effectively contain zero-value calculations.
This allowed Cerebras to significantly boost effective computing performance beyond the chip’s raw dense compute figures.
Competing with Nvidia in the AI race
While Nvidia later introduced its own sparsity support in the Ampere GPU generation, Cerebras maintained an advantage in memory bandwidth and specialised AI architecture.
However, the company’s approach came with trade-offs.
Unlike GPUs that relied on large amounts of HBM or GDDR memory, Cerebras processors used enormous quantities of ultra-fast SRAM directly on the chip. SRAM offers exceptional speed but is considerably less space-efficient, limiting memory capacity.
Even so, the architecture proved powerful enough for Cerebras to continue scaling the platform.
WSE-2 and WSE-3 expanded the company’s reach
The second-generation WSE-2 processor arrived in 2021 using TSMC’s 7nm manufacturing process, more than doubling transistor density, bandwidth and memory capacity compared with the original design.
Large clusters containing up to 192 wafer-scale systems became possible, although most commercial deployments typically used between 16 and 32 systems per site.
It was during this period that Abu Dhabi-based cloud and AI group G42 emerged as Cerebras’ largest financial backer. By 2023, the startup had secured approximately $900 million in orders tied to nine supercomputing installations.
A year later, Cerebras launched the WSE-3 platform using a 5nm process node. Compute performance again doubled, reaching more than 125 petaFLOPS of sparse AI performance at 16-bit precision.
The systems now power much of the Condor Galaxy infrastructure built for G42, alongside several deployments across North America and Europe.
Cerebras finds unexpected success in AI inference
For much of its history, Cerebras focused heavily on AI model training. That changed in 2024 when the company entered the rapidly expanding AI inference market.
Its hardware turned out to be particularly effective for running large language models due to its exceptional memory bandwidth. The WSE-3 reportedly offers memory speeds approaching 21 petabytes per second — dramatically faster than conventional GPU systems.
Combined with speculative decoding techniques, the platform enabled extremely high token generation speeds for AI applications.
According to benchmarking firm Artificial Analysis, Cerebras systems can generate more than 2,200 tokens per second when running GPT-OSS 120B High, placing the company among the fastest inference providers globally.
IPO delayed amid revenue concerns
Cerebras initially filed to go public in 2024 but delayed its IPO after concerns emerged over its financial dependence on G42, which reportedly accounted for 87 per cent of company revenue at the time.
Over the following year, the business diversified its customer base substantially, adding clients including Amazon Web Services, Meta, Mistral AI, OpenAI and Perplexity AI.
Investors responded enthusiastically when the company finally floated this week, sending shares soaring nearly 70 per cent on debut trading.
What comes next for Cerebras?
Despite its successful IPO, Cerebras faces increasing pressure to maintain its technological lead.
Rivals are rapidly developing competing inference hardware, while Nvidia’s acquisition of Groq has strengthened its own position in ultra-fast AI inference systems.
Industry analysts now expect Cerebras to unveil a next-generation WSE-4 platform in the near future, potentially focused on lower-precision formats such as FP8 and FP4, which are becoming increasingly important for large language models.
The company is also widely expected to expand SRAM capacity using advanced 3D chip-stacking techniques from TSMC to support larger AI models more efficiently.
Partnerships likely to shape the next phase
Further partnerships could prove equally important.
Earlier this year, Amazon Web Services confirmed plans to combine its Trainium3 accelerators with Cerebras systems to improve inference performance, mirroring broader trends across the AI hardware sector.
Similar collaborations with firms such as AMD could follow as demand for specialised AI infrastructure continues to rise.
For Cerebras, the IPO marks a major milestone — but also the beginning of a far more demanding phase. Investors who backed the company’s wafer-scale gamble will now expect rapid growth, stronger revenues and another leap forward in AI silicon technology.
“Entrepreneur. Freelance introvert. Creator. Passionate reader. Certified beer ninja. Food nerd.”
More Stories
Fish and Chip Shops Turn to Alternative Species as Cod Prices Surge
Surge in ‘Battery Cow’ Dairy Farming Across the UK as Farmers Battle Rising Costs
Manchester Airport named worst in UK for flight delays