Open Compute Project summit puts AI infrastructure in focus
The OCP Global Summit is taking place this week in San Jose, California, drawing hardware vendors, cloud operators, chip designers, and data center engineers into a single forum for open infrastructure. The annual conference is organized around the Open Compute Project, an industry initiative that develops more open, efficient, and scalable hardware designs for data centers and other computing environments. The agenda spans the physical layer behind cloud computing and artificial intelligence, from AI servers and accelerators to standardized hardware designs, power and cooling systems, networking, and connectivity.
The timing reflects the pressure facing modern data centers. AI workloads are pushing rack densities higher, forcing operators to rethink how power is delivered, how heat is removed, and how quickly new capacity can be brought online. The announcements made at the summit show vendors responding with designs that are more integrated, more liquid-cooled, and more modular than the infrastructure built for earlier generations of cloud computing. They also show how open specifications are becoming a way for suppliers to align around common building blocks while still competing on implementation and performance.
Nvidia contributes Vera Rubin rack and compute tray designs
Nvidia submitted key rack and compute tray designs for its Vera Rubin platform to the Open Compute Project through its modular MGX architecture. The goal is to give server vendors and other infrastructure suppliers a shared design foundation for building AI systems. By contributing the designs, Nvidia is trying to make it easier for partners to develop systems around the platform without starting from a blank sheet for every rack and tray.
One feature that distinguishes Nvidia’s rack and tray designs is that there are no fans or cabling because everything is liquid cooled. That approach removes the need for air movers inside the rack and reduces the complexity of internal cabling. In practice, fewer moving parts and fewer cable connections can make systems easier to deploy and service. The design also points to a broader shift in high-density AI infrastructure, where liquid cooling is moving from an option for specialized deployments to a foundational element of the rack architecture.
The MGX contribution fits the Open Compute Project’s broader role as a place where companies can share mechanical, electrical, and thermal designs that become reference points for multiple vendors. For customers, that can mean more compatible options and less integration work. For suppliers, it can mean a faster path from a chip platform to a deployable rack-scale system. The emphasis on liquid cooling also signals that future AI racks will be designed around coolant distribution, service access, and power delivery rather than around airflow alone.
Arm proposes a new way to measure CPU performance for agentic AI
Arm released a white paper, developed with partners, that proposes a use-case-driven methodology for measuring the CPU’s contribution to agentic AI systems. The paper defines five representative workload types: software engineering, enterprise operations, financial and scientific research, personal assistants, and robotics. It pairs component-level benchmarking with end-to-end task measurement built on re-playable agent trajectories.
The methodology addresses a gap in how processors are often evaluated. Agentic AI systems do more than run a single model query. They plan, call tools, coordinate multiple steps, create isolated execution environments, and return results. The CPU plays a central role in that orchestration, even when accelerators handle the heaviest model computation. Measuring only conventional CPU benchmarks can miss the work that determines whether an agent completes a task efficiently in a real rack-level deployment.
Arm also highlighted independent testing by Signal65 that found its AGI CPU could deliver up to twice the rack-level performance of AMD’s Epyc Turin generation server processors across select agentic AI workloads. The comparison emphasizes completed work under rack-level constraints rather than conventional CPU benchmarks alone. In the testing, an Arm AGI CPU averaged 1.7x better performance for tooling, 2.0x for agentic workflows and orchestration, and 2.2x for reasoning and planning in modeled per-rack performance versus an AMD Epyc Turin processor. Arm said its AGI CPU reduces sandbox creation latency by up to 45% versus AMD Epyc 9965, helping agents spin up isolated execution environments faster.
Those figures are vendor-backed and workload-specific, but they illustrate the direction of travel in AI infrastructure. As agents become more common, the CPU is being judged by how well it supports tool use, scheduling, isolation, and task completion under power and space limits. The proposed methodology is an attempt to make those comparisons more realistic, tying processor performance to the kinds of work that AI systems actually perform.
Kioxia introduces high-density SSDs for AI storage
Kioxia unveiled its LD4 Series enterprise solid-state drives, designed to increase flash storage density in hyperscale data centers and high-density 1U servers. The drives target read-intensive workloads, including data storage applications supporting artificial intelligence infrastructure. That focus matters because AI pipelines generate and consume enormous volumes of data, from training sets and checkpoints to inference logs, metadata, and retrieval systems.
The new SSDs offer eighth-generation BiCS FLASH 3D quad-level cell NAND technology with capacities of 15.36TB and 30.72TB. They use a validated architecture designed to support capacities of up to 122.88TB. The drives support PCIe 5.0 and NVMe 2.0e, along with the Open Compute Project’s Datacenter NVMe SSD Specification 2.6. They feature a single-port design for hyperscale and data center-class servers. Kioxia said the LD4 Series is currently sampling to select customers.
QLC NAND is attractive for read-intensive storage because it can pack more bits into each cell and lower the cost per terabyte. The trade-off is that write endurance and performance characteristics must be matched to the workload. By aiming the LD4 Series at read-heavy AI infrastructure, Kioxia is targeting the tier of storage where capacity, density, and power efficiency can matter as much as raw write speed. The support for OCP’s NVMe SSD specification also indicates that the drives are being designed for standardized hyperscale deployments rather than only proprietary server designs.
Cooler Master brings gaming liquid cooling experience to the data center
Cooler Master, a company best known for providing liquid cooling to gaming systems, is bringing that experience to the data center. It is expanding its end-to-end liquid-cooling portfolio for AI data centers and highlighting a 2,500 kW liquid-to-liquid coolant distribution unit designed for high-density AI and high-performance computing infrastructure.
The system supports coolant flow rates of up to 990 gallons per minute and features dual-system architecture, intelligent load balancing, and centralized management for as many as 32 interconnected CDUs, according to the vendor. Mass production is targeted for the fourth quarter of 2026. The scale of the unit reflects the heat loads that come with rack-scale AI systems, where multiple high-power accelerators and CPUs are packed into a small physical footprint.
A coolant distribution unit sits between the facility water loop and the technical cooling loop that serves IT equipment. It transfers heat, manages flow, and helps isolate the data center’s primary water system from the coolant that runs through servers and racks. As AI racks grow denser, operators need CDUs that can handle higher capacities while still providing redundancy, balancing, and centralized control. Cooler Master’s expanded portfolio is aimed at data center operators building new AI facilities as well as those retrofitting existing sites to accommodate increasing rack densities and thermal demands.
Flex expands prefabricated modular data center portfolio
Flex announced an expansion of its prefabricated modular data center portfolio with integrated systems designed to accelerate deployment of AI infrastructure by as much as 50% compared with traditional construction methods. The new AI Modules combine power delivery, high-density computing, and advanced liquid cooling in factory-built systems. This approach shifts much of the integration and testing work from the data center construction site to Flex, the company says.
The systems include an 800-volt direct-current power rack for power conversion and distribution, cooling distribution units to manage heat from power-intensive AI workloads, and high-density IT racks incorporating AI servers, networking equipment, and high-speed interconnects. By assembling and testing these elements in a factory, the company aims to reduce the variability and schedule risk that come with building complex AI infrastructure on site.
The expanded portfolio is part of Flex’s cloud and power infrastructure business, which the company plans to spin off as an independent, publicly traded company called Axiom Solutions International in the first quarter of 2027. The modular approach reflects a larger trend at the summit: AI infrastructure is being treated as an integrated system, with power, cooling, compute, and networking designed together rather than added in separate phases.
Source:Network World News