Even if there’s been a lot of excitement around AI in the last two years, there’s been a very limited number of pure play IPOs given that the market is dominated by a handful of players in computing (Microsoft, Amazon…), models/software (OpenAI, Anthropic…) and computing hardware (Nvidia, AMD). News that Cerebras, a US-based AI chip maker, filed for its IPO then made some noise, notably as it competes directly with Nvidia and boasts impressive chip performances and strong revenue growth.
Cerebras’ solution to solve the memory and communication bottlenecks in High Performance Computing is unique in the industry. Instead of linking together hundreds of GPUs sitting in servers, Cerebras’ solution lies in the manufacturing of a massive chip, the size of a full silicon wafer, holding thousands of computing cores, all interconnected at the wafer level and all having a direct access to a huge pool of on-chip memory. This onboard memory, 3,000x times bigger than competing products, means that the 900,000 cores can handle very large language models (LLM) without having to rely on a networking infrastructure to exchange and store data, hence materially optimizing the latency and power consumption of the whole system.
Measured by the number of processed tokens by seconds, Cerebras’ massively parallel wafer-scale computing engine clearly outperforms the competition both in the training and inference modes of LLMs. Most importantly, on a Total Cost of Ownership (including operating costs, software recoding…), Cerebras’ solution is clearly competitive with Nvidia’s latest top-notch solution.
While Cerebras’ approach is interesting (massive single-chip approach vs. packaging of large numbers of GPUs for Nvidia), it faces two major challenges. First, manufacturing (in partnership with TSMC) will take time to scale as Cerebras’ chips are unique in the industry as we said above and require very specific processes, while Nvidia ships millions of chips each quarter.
Second, we believe that a migration out of Nvidia, if any, will take time as developers have a long history working on Nvidia’s CUDA software stack, the company’s proprietary programming language used by millions of coders around the world to write programs that run on GPUs and hence fully exploit their superior vector and parallel computing capabilities. CUDA has become the de-facto standard for GPU acceleration in deep learning and AI applications and the programs and libraries relying on CUDA are countless and can be found in every piece of AI/Machine learning software.
Confirming these challenges, business traction Cerebras has had so far is pretty limited. True, company revenue jumped to $136 million in the first half alone from $78 million in 2023 and losses shrank to $67 million. But these revenue jumped off a very low base and were essentially made with a single customer, that is not a US hyperscaler but a UAE-based AI firm…
A recent partnership with AI server vendor Dell could be a turning point for Cerebras and potentially help it strike deals with hyperscalers that will be critical for the company’s future. But for now, that’s a show me story.
