Arm recently announced that it, along with AMD and Nvidia, has been appointed to the Open Compute Project (OCP) board. These companies will collaborate with major players such as Meta, Google, Intel, and Microsoft to drive innovation in open and interoperable designs for AI data centers.
Arm stated that the AI economy is reshaping computing infrastructure from cloud to edge, creating unprecedented demands for performance, efficiency, and scale. Data centers are undergoing a significant transformation, shifting from general-purpose servers toward rack-scale systems and large clusters specifically designed for AI workloads, while also facing power consumption challenges.
By 2025, the computing power of a single AI rack is expected to match that of the top supercomputers in 2020, consuming electricity equivalent to about 100 US households combined.
To address this challenge, advancing infrastructure to a new stage is imperative, with open collaboration playing a key role in the rapidly evolving ecosystem.
Arm views converged AI data centers as the next phase in infrastructure evolution, aiming to maximize AI compute density per unit area to reduce overall power consumption and costs associated with AI execution.
Achieving this requires co-design across computing, acceleration, memory, and networking layers. For example, Arm Neoverse contributes precision in converting data into tokens, which drive advanced AI models and agents, delivering real value in AI science, healthcare, and business applications.
Arm noted that developing converged AI data centers cannot rely solely on a single general-purpose chip. High-end chips tailored for specific applications will be critical to improving system integration density. Chiplet technology, through advanced packaging and 2.5D/3D techniques, offers viable paths for higher-density system design and opens opportunities for cross-vendor collaborative design.
Recently, Arm also announced its contribution of the Foundation Chiplet System Architecture (FCSA) specification to OCP, deepening industry cooperation in the chiplet domain. FCSA extends Arm's existing Chiplet System Architecture (CSA) by creating a vendor-neutral framework independent of CPU architecture. This unified standard for chiplet systems and interface definitions accelerates chiplet design and integration while promoting large-scale reuse and interoperability.
The origins and expansion of OCP
The global cloud data center industry historically lacked common standards until Facebook initiated OCP in 2011, publicly sharing hardware specifications and build standards for its data center servers and storage systems. This move sparked a worldwide trend toward open specifications for cloud data centers.
OCP quickly attracted participation from Google, Microsoft, AMD, Intel, Nvidia, Amazon, Meta, Dell, HPE, and Taiwanese firms, including Quanta, Wiwynn, and Foxconn. Today, OCP boasts over 500 member companies, with OCP Ready facilities and experience centers expanding continuously, and multiple hyperscalers adopting jointly developed open standards in actual deployments.
OCP's main goals include establishing open standardized specifications for servers, storage, networking equipment, and racks; enhancing energy efficiency through streamlined power and cooling designs to reduce data center energy use; lowering costs by avoiding redundant development and enabling shared hardware platforms among cloud providers and supply chains; and promoting interoperability so components from different vendors can operate compatibly.
Key highlights of the 2025 OCP Summit
The October 2025 OCP Summit focused on four key topics: AMD's Helios ORW, 800V DC power architecture, Google's Deschutes 2MW cooling distribution unit (CDU), and ESUN networking solutions.
The Helios rack features a dual-wide (2T) design scheduled for shipment starting in the second half of 2026. Meta and AMD demonstrated modular rack designs supporting both standalone and horizontal scaling configurations. The 800V DC solution is planned to launch alongside Nvidia's Rubin Ultra platform in the second half of 2027.
However, AWS remains absent from the 2025 OCP event. Industry rumors suggest that AWS prohibits employees from participating in OCP activities due to its internally customized specifications, aiming to avoid external design influences.
Article edited by Jack Wu
