Cerebras accelerates the expansion of six data centers, increasing inference speed by ten times

Cerebras Systems recently announced that it will build six new data centers in North America and Europe to enhance its artificial intelligence (AI) reasoning capabilities. This move will greatly enhance the company's computing power and support the development of various AI applications.

Under the plan, 85% of computing power will be concentrated in the United States, with three facilities already operating in Santa Clara, California, Stockton, and Dallas, Texas. Next, the new center will open in Minneapolis (expected in the second quarter of 2025), Oklahoma City and Montreal (expected in the third quarter of 2025), and Atlanta and France (expected in the fourth quarter of 2025).

The core of the new data center will be the "wafer Scale Engine" developed by Cerebras, a special chip architecture optimized for AI applications. The company said its CS-3 system can process 40 million token data of the Llama-70B model per second, greatly improving the processing speed of inference tasks. More than 300 CS-3 systems are expected to be equipped in the Oklahoma City facility. The center is built in accordance with the third-level plus three standards, has the ability to resist tornadoes and earthquakes, and is equipped with triple redundant power supplies. It is expected to start operations in June 2025.

At present, several well-known AI companies have signed cooperation agreements with Cerebras, including French startup Mistral and its Le Chat assistant, as well as the AI ​​Q&A engine Perplexity. HuggingFace and AlphaSense have also joined the ranks of using the Cerebras platform. This technology is particularly suitable for inference models that require long-term calculations and generation of large numbers of tokens, such as Deepseek-R1 and OpenAI o3.

The expansion is part of Cerebras' overall expansion strategy for 2025, with some facilities operating in partnership with UAE company G42. In Montreal, a new center managed by Bit Digital subsidiary Enovum is expected to go live in July 2025, with inference speeds that will be ten times faster than the current GPU.

Cerebras Systems is an American company focusing on the development of AI chips. Its unique design philosophy is to use the entire wafer as a single chip. It has now launched the third-generation wafer-scale engine WSE-3. The system has been used in institutions such as Argonne National Laboratory, Pittsburgh Supercomputing Center and GlaxoSmithKline. Despite its technological advantages, there are limitations, such as not supporting native CUDA (Nvidia's standard) and not as good as Nvidia solutions in terms of server compatibility.