en.Wedoany.com Reported - US artificial intelligence chip company Cerebras plans to expand its data center and AI computing infrastructure in Europe, with an estimated investment of several billion dollars, aiming to increase local processing capacity to 200 megawatts by 2027. Cerebras CEO Andrew Feldman announced at the Raise Summit AI conference in Paris, France, that this expansion is primarily to meet Europe's growing demand for AI inference computing power, while also accommodating European customers' requirements for data control, local deployment, and computing resource management.
Cerebras currently operates three data centers deploying its chips in France, Finland, and Norway. The next phase of construction will build on existing facilities, continuing to expand server capacity, wafer-scale processors, power supply and distribution, cooling, and network connectivity, forming larger-scale AI computing nodes across different regions of Europe. The 200 MW figure does not refer to the power consumption of a single chip or server, but rather the overall carrying capacity of computing equipment and related supporting facilities in the data center. Actual construction will require simultaneous increases in server room space, power access, cooling systems, fiber optic networks, and operational management platforms.
The development of AI infrastructure in Europe is being driven by both computing power demands and data management requirements. As generative AI models enter enterprise applications, inference tasks must continuously process user requests after model training, with operational frequency typically higher than centralized training. With the increase in AI agents capable of autonomously invoking tools, querying information, and executing tasks, a single user request may trigger multiple rounds of model computation, data retrieval, and system interactions. Inference infrastructure must therefore handle higher concurrency while maintaining low response latency.
Feldman stated that the growth rate of demand for AI computing power in the European market has already exceeded the pace of existing facility expansion. By deploying data centers in France, Finland, Norway, and other European regions, Cerebras can enable some models and data to be processed within Europe, reducing enterprises' reliance on long-distance cross-border computing resources. For clients in pharmaceuticals, software development, and high-performance computing, local computing nodes can also shorten data transmission paths and provide more flexibility in data control and infrastructure choices.
Cerebras' core technology is the wafer-scale processor. Traditional chips are typically cut from a whole wafer into multiple smaller processors, which are then connected into a computing cluster via high-speed networks. Cerebras, however, uses a processor nearly the size of an entire wafer as a single computing system. A larger chip can accommodate more computing cores and storage resources within the same processor, reducing the latency caused by repeated data transmission between multiple independent chips.
This architecture is primarily designed for large model training and AI inference scenarios. Multi-chip systems rely on interconnections within servers and data exchange across server networks. As the scale of connections expands, communication latency and data synchronization may limit overall computing efficiency. Wafer-scale systems attempt to concentrate more computation and data exchange within a single large processor, enabling models to complete inference faster and return results. However, large processors impose higher requirements on chip manufacturing, power supply, cooling, system packaging, and fault management, and the associated data centers must be configured around the characteristics of these devices.
Cerebras' European clients include British pharmaceutical company GlaxoSmithKline, as well as high-performance data centers and software developers in Scotland and Germany. The company has also signed a computing capacity supply contract with OpenAI, reportedly valued at over $20 billion, to provide computing resources for ChatGPT at least until 2028, and has recently established a partnership with Amazon Web Services. These collaborations mean that Cerebras' new computing capacity will not only serve its own data center clients but may also be offered through AI platforms and cloud service channels.
In terms of construction pace, Cerebras needs to increase its European processing capacity to 200 MW within a relatively short timeframe. Whether the project can be completed on schedule will depend on data center site selection, power resources, equipment delivery, chip production capacity, and the progress of supporting infrastructure. Existing nodes in France, Finland, and Norway will undertake the initial expansion tasks. Whether the company enters more European countries will depend on local power conditions, data management requirements, and client distribution.
This expansion reflects that competition in AI infrastructure is extending from single-chip performance to the delivery of computing services. Companies such as Nvidia, AMD, and Cerebras not only need to provide processors but also address issues related to server integration, network interconnection, energy supply, and long-term operation after chips enter data centers. The 200 MW European computing facility planned by Cerebras will serve as a key milestone in determining whether its wafer-scale system can expand its application in large-scale commercial inference scenarios.
