OCP CEO George Tchaparian described the organization’s increasingly broad mandate as “chip-to-grid.” Keynotes from AMD, NVIDIA, Meta, Arm, and Broadcom reinforced the point that scaling AI now requires coordination across processors, networking, power distribution, cooling, and the electrical grid.
Meanwhile, Agentic AI introduces new demands for CPU capacity and infrastructure efficiency, making that coordination more urgent.
We teased out a few key themes from the event that advance the state of AI infrastructure
- Agentic AI puts the CPU back on the roadmap
- Ethernet moves into AI Scale-Up Networking
- Power Management comes to OCP, now an intrinsic part of IT infrastructure
- Time to Capacity is the new competitive advantage
- Openness is now the mantra for AI infrastructure
Let’s take a look.
Agentic AI Puts the CPU Back on the Roadmap
Agentic AI relies on CPUs for orchestration, tool calls, code execution, and other tasks, alongside GPU-based inference. In her keynote, AMD CEO Lisa Su estimated that inference accounts for “about 60%” of the AI compute market and cited a 14-fold increase in agentic AI traffic this year.
These workloads require more deliberate balancing of CPU and GPU/accelerator resources.
Several announcements highlighted the growing focus on CPU infrastructure for agentic workloads:
- AMD Helion: An open, rack-scale reference design for agentic workloads that pairs EPYC CPUs with Instinct MI350P GPUs in configurable ratios. It scales to 32,768 sixth-generation EPYC cores in its CPU-only form and features an integrated token router. AMD will contribute Helion to OCP, and partner systems will ship starting in November.
- NVIDIA Vera CPU rack: A liquid-cooled MGX rack housing 256 Vera CPUs for tool calling, orchestration, and sandboxed agent execution. NVIDIA also said that Meta plans to deploy Vera-based CPU infrastructure.
- Astera Labs Orion: A compute-agnostic rack-scale reference design featuring pre-validated, interchangeable compute trays, enabling customers to choose their own CPU and accelerator. Systems will reach customers through OEM and ODM partners beginning in 1Q27.
- Arm agentic CPU benchmarking: Arm published a white paper, developed with input from Fujitsu, Meta, and OpenAI, on how to measure CPU performance for agentic applications, a workload that conventional benchmarks fail to capture.
AMD is backing up its predictions, raising its forecast for the data center CPU addressable market three times over nine months, putting its 2030 estimate above $200 billion.
For infrastructure planners, CPU-to-GPU ratios, workload placement, and resource utilization are increasingly important. AMD and NVIDIA’s designs suggest that CPU-dense systems will play a larger role alongside accelerator racks.
Ethernet Makes Its Move Into AI Scale-Up Networking
AI networking spans scale-up connections among tightly coupled accelerators, scale-out fabrics connecting larger clusters, and scale-across links among compute domains. Ethernet already dominates much of scale-out networking.
Its expansion into scale-up poses a greater challenge to proprietary fabrics.
Broadcom’s Asad Khamisy said in his keynote that “more than five” hyperscalers already run Ethernet scale-up networks in production.
The summit highlighted developments across these networking domains.
- Meta Tahan, MetaRoCE, and OSPA: Meta showcased Tahan, an Ethernet scale-up switch that delivers 6.4 Tbps per MTIA 300 accelerator and released open-source implementations of its MetaRoCE RDMA transport. In collaboration with Microsoft and Alibaba, Meta launched the Open Switch Platform Architecture initiative and submitted its v1.0 specification to OCP.
- Meta and NVIDIA Spectrum-6: Meta is adopting NVIDIA Spectrum-6 102.4 Tb/s Ethernet switches in its N-IceCube switch, which runs Meta’s open-source FBOSS in the liquid-cooled IcePack rack. Spectrum-X Multiplane spreads GPU traffic across up to eight network planes for clusters with more than 100,000 GPUs.
- NVIDIA and Cisco USFP: A new pluggable optics form factor with a path to 400 Gb/s per lane, delivering four times the bandwidth density of 200G-per-lane optics and 1.3x better power efficiency. TE Connectivity and Amphenol join a multi-source effort to develop cages, connectors, and modules.
- NVIDIA and Google Falcon: Future NVIDIA ConnectX-10 SuperNICs and BlueField DPUs will support Google‘s Falcon transport, enabling NVIDIA and Google NICs to communicate over a single, reliable Ethernet link.
While NVIDIA still relies on NVLink for Vera Rubin scale-up connectivity, Meta’s Tahan effort demonstrates Ethernet’s growing potential in tightly coupled accelerator systems, and Spectrum-6, USFP, and Falcon broaden the open networking ecosystem.
Specifications alone will not guarantee interoperability. Practical multivendor operation will require consistent congestion handling, telemetry, implementation behavior, and validation.
Power Management Comes to OCP
Power availability increasingly determines where and how quickly companies can deploy AI capacity. At gigawatt scale, data centers must also manage their impact on the electrical grid.
Meta’s Prometheus cluster in Ohio, for example, targets nearly 1 GW in 2026, while its Hyperion campus in Louisiana is designed to scale toward 5 GW.
These challenges are pushing OCP beyond IT equipment into facility design.
Summit announcements included several specifications to standardize how AI factories consume, distribute, and manage power:
- Open Data Center Specification v0.7: NVIDIA, AMD, Google, and Meta published a common blueprint for AI factory power architecture, cooling infrastructure, and physical layout. Equipment makers and colocation providers now have a shared target, replacing one-off customer specifications.
- OCP Grid Resilience Specification: Google, Meta, Microsoft and NVIDIA defined how large AI sites ride through utility disturbances, staying connected, responding automatically and restoring near-full load within two seconds.
- NVIDIA MGX 1.2 and DSX MaxLPS: MGX 1.2 adds a liquid-cooled 5,000A busbar, rack-scale blind-mate liquid cooling, and support for racks up to 250 kW. In a test with Nscale, MaxLPS dynamic power management fit 192 GPUs into a 264.4 kW budget that previously held 140, increasing inference throughput by about 49%.
- Visionbay K-1: An 800 VDC AI factory platform built on NVIDIA’s reference architecture, with Vertiv power and thermal technology, phasing in DC distribution where it delivers the most benefit. NVIDIA’s Ian Buck said that 800 VDC reduces copper requirements by about 45%.
NVIDIA’s Ian Buck argued that AI factories must act as both grid contributors and consumers.
While shared facility and resilience specifications can reduce duplicated engineering and provide equipment suppliers and colocation providers with more predictable design targets, their practical value will depend on adoption, testing, and implementation.
Time to Capacity Becomes a Competitive Advantage
At gigawatt scale, deployment speed has direct economic consequences. Every month spent constructing, integrating, or commissioning a site delays the use of already-committed capital. Summit presentations treated time to capacity as an engineering target.
Three examples from the keynotes stood out:
- Meta fabric structures: Meta houses parts of Prometheus within engineered fabric structures from Sprung Structures, achieving operational readiness in about three to five months. Meta trained a model on the cluster while it was still being built.
- Meta AI-designed concrete: An AI-assisted mix reduced embodied carbon by 19% and cured 43% faster, with no additional material cost. Meta released both the tool and the formulation.
- NVIDIA Vera Rubin tray assembly. Simplified mechanical design reduced compute tray assembly time from about 90 minutes on Grace Blackwell to roughly one minute on Vera Rubin.
At thousands of racks per campus, minutes saved per tray and months saved per building can significantly accelerate the delivery of usable compute. Construction methods, modular systems, and simpler assembly now matter as much as chip performance.
For infrastructure operators, time to capacity is emerging as a competitive differentiator.
Openness Becomes an Infrastructure Strategy
The summit also showed that open infrastructure is becoming central to vendors’ strategies. AMD’s Helion, NVIDIA’s MGX contributions, Meta’s networking initiatives, and common facility specifications are expanding the ecosystem capable of manufacturing, integrating, and deploying AI systems.
Vendors still differentiate on silicon, software, and proprietary interconnects, but increasingly standardize the surrounding infrastructure to reduce integration friction and accelerate adoption.
Open specifications are only the starting point. Effective multivendor infrastructure requires compatible implementations, repeatable qualification, and operational tooling.
NVIDIA’s open-source DSX management components, including rack management and automated bare-metal provisioning, align with the growing emphasis on deployment and operations.
Final Thoughts
The most important takeaway from OCP 2026 is that AI infrastructure competition now extends well beyond processor performance:
- Agentic workloads are driving more heterogeneous compute designs
- Ethernet is expanding architectural options
- Power, cooling, grid behavior, and deployment schedules increasingly determine how much capacity can come online.
These developments also reveal a pragmatic approach to openness, in which hyperscalers and suppliers continue to compete at the component and software layers while collaborating on the surrounding infrastructure needed to deploy those innovations at scale.
For buyers, that creates opportunity but also underscores the importance of validating interoperability and operational maturity.
As OCP prepares to move its 2027 summit to Moscone Center in San Francisco, expect grid integration, CPU-intensive agentic infrastructure, and deployment efficiency to remain central themes over the coming year.
OCP 2026 made clear that scaling AI is now a chip-to-grid engineering challenge.
