AI chip supplier Cerebras Systems Inc., one of the formidable competitors to NVIDIA, is in discussions to raise approximately $10 billion in a new financing round to support its long-term rivalry with NVIDIA. This move aims to further solidify the advantages of its AI computing clusters over NVIDIA’s AI GPU clusters in terms of cost-effectiveness and energy efficiency. According to informed sources, the company’s pre-money valuation for this round will reach $22 billion, marking a substantial increase of 170% compared to the approximately $8.1 billion valuation in September last year. Meanwhile, Cerebras Systems continues to actively advance its plans for an initial public offering (IPO) in the United States.
Under the leadership of CEO Andrew Feldman, the company is actively seeking to challenge NVIDIA’s approximately 90% market share in the AI chip sector. Cerebras Systems is renowned for its unique “wafer-scale engine” architecture, which places an entire AI model on a single, exceptionally large chip, designed to significantly enhance inference performance and memory bandwidth. The company not only provides physical computing clusters but also offers remote AI computing services to major clients such as Meta Platforms Inc., IBM, and Mistral AI.
Unlike NVIDIA, Broadcom, AMD, and others, which employ smaller chips integrated through advanced packaging, Cerebras Systems manufactures exceptionally large chips that cover the entire silicon wafer. Analysis from semiconductor research institutions like Semianalysis points out that this wafer-scale architecture can achieve higher performance density and energy efficiency compared to traditional AI GPUs or AI ASICs when handling large language model inference tasks. Consequently, Cerebras Systems is regarded as one of NVIDIA’s strongest competitors in the rapidly growing AI inference market.
The company’s latest CS-3 system, equipped with the WSE-3 chip, has demonstrated in multiple public comparisons that its performance in running large model inference tasks, such as Llama 3 70B, is reportedly about 21 times faster than NVIDIA’s latest Blackwell architecture B200 AI GPU system, while also offering lower overall costs and energy consumption. This gives it a significant advantage in terms of cost-effectiveness and energy efficiency for large-scale inference tasks, particularly when handling large language models. However, these advantages are more focused on specific inference scenarios, with NVIDIA still maintaining leadership in general-purpose computing, AI training, and CUDA ecosystem compatibility.
The demand for AI inference is growing rapidly, showing a trend of approximately doubling every six months. At the same time, competitive pressures are increasing. For instance, Google positions its new generation of TPUs, such as TPU v7/Ironwood, as specialized accelerators “born for the era of AI inference,” and they demonstrate advantages in cost-effectiveness compared to NVIDIA’s Blackwell architecture in specific scenarios. Reports indicate that OpenAI’s large-scale leasing of TPUs through the Google Cloud Platform is partly motivated by the desire to reduce inference costs.
As the wave of AI inference fully hits, Cerebras Systems is strengthening its position through financing and advancing its IPO, aiming to challenge the existing industry landscape in an expanding market. Whether its technological approach and business model can continue to attract capital and clients will largely determine its ability to secure a foothold amid competitive pressures from giants like NVIDIA and Google.
