Agentic AI Is Redrawing the Data Center: What OCP 2026 Means for CIOs

For three years, the AI data center story has been about GPUs. At the Open Compute Project Global Summit in San Jose this week, the biggest infrastructure vendors started describing a different machine. Nvidia and AMD both introduced open rack designs built around CPUs for agentic AI, and TSMC's CEO named agentic AI CPUs as a new source of chip demand. The shift matters to enterprise leaders because it changes what an AI-ready data center needs, what it costs and how much choice buyers will have.

What was announced at OCP 2026?

Nvidia contributed its Vera Rubin rack and compute-tray designs to the Open Compute Project through its MGX architecture, according to a StorageNewsletter summary of the company's October 12 announcement. The updated MGX 1.2 specification supports racks of up to 250 kilowatts with liquid cooling. Alongside it, Nvidia released a Vera CPU rack design aimed at agentic workloads such as tool calling, orchestration and sandboxed code execution, and open-sourced four components of its DSX OS software for running large GPU clusters.

AMD went further on the CPU side. Its Helion blueprint, reported by Techzine on October 13, is an open rack design that mixes EPYC CPUs and Instinct GPUs in whatever ratio a deployment needs. A CPU-only version holds up to 32,768 cores. AMD described an all-in-one configuration for enterprises that can sit inside existing IT, and said it plans to contribute the specification to OCP, with first configurations expected in November. "Agentic AI is not a single workload," said Dan McNamara, AMD's senior vice president and general manager of Compute and Enterprise AI.

The industry is also writing shared facility rules. Nvidia, AMD, Google and Meta published version 0.7 of an Open Data Center Specification that covers power, cooling and physical layout, and Google, Meta, Microsoft and Nvidia are backing a grid resilience specification meant to restore near-full load in under two seconds after a grid disturbance, StorageNewsletter reported.

Why do AI agents need more CPUs?

The reasoning is straightforward. A chatbot spends most of its time on model inference, which runs on GPUs. An agent does much more around the model: it calls APIs, queries databases, runs code in sandboxes, waits on other systems and coordinates multi-step workflows. Techzine's account of AMD's pitch notes that much of that work falls to CPUs, while GPUs mainly handle inference.

TSMC's numbers point the same way. When TSMC reported results on October 15, CEO C.C. Wei said CPUs and networking chips for agentic AI are adding to demand that already outstrips the company's supply, CRN Asia reported.

AI Connect speakers have been describing this split for months. In a AI Connect panel on AI infrastructure at scale, a Rhino Federated Computing executive explained that his team runs preprocessing and postprocessing in separate CPU containers and reserves GPU machines for inference alone, a design choice made to control cost. What vendors announced this week is that same idea applied at rack scale.

Will open designs actually give enterprises more choice?

This is where the views diverge. The case for optimism is scale. IDC, in a study commissioned by the OCP Foundation and released October 13, now forecasts about $800 billion in annual spending on OCP-recognized equipment by 2029, roughly 2.7 times its previous forecast. For 2030, it expects about 70% of spending on such equipment to be AI-related. Open designs at that volume tend to bring more suppliers, faster qualification and lower prices, which eventually reaches enterprise buyers through server makers and colocation providers. Those figures are projections, not spending that has happened.

The skeptical view is that "open" has limits. An unsigned comments section appended to StorageNewsletter's coverage argued that the core of Nvidia's stack, including its GPUs, NVLink interconnect and CUDA software, remains proprietary. Opening the rack does not open the chip. Even Meta's adoption of Nvidia's Spectrum-6 switches, announced the same day, came with Meta stressing that it can run its own FBOSS network software across those fabrics. The largest buyers clearly care about keeping their own software layer, whoever supplies the hardware.

Both points can be true. Standard racks, power and cooling designs reduce the cost of building and running AI capacity, while the most valuable components remain concentrated among a few suppliers. Dave Mariani, co-founder and CTO of AtScale, made a related argument about software in the same AI Connect panel, saying infrastructure has to stay open because "we don't know where this stuff is going."

The facility is becoming the bottleneck

The other theme from the summit was power. A 250-kilowatt rack is far beyond what most enterprise data centers were built for. Steve Heibein, federal AI chief technologist at Hewlett Packard Enterprise, explained the physics in a AI Connect webinar on AI and HPC convergence. "A thousand watt GPU needs a thousand watts of cooling," he said, noting that older facilities may allocate as little as 10 kilowatts per rack. The new specifications assume liquid cooling and, in some cases, 800-volt DC power distribution.

What this means for enterprise leaders

• Plan for a mixed fleet. If agents become a large share of AI work, CPU capacity, memory and networking will matter as much as GPU count. Size agent platforms with that in mind, and measure where agent time actually goes.
• Ask vendors about OCP alignment. Servers and racks built to open specifications should be easier to source from more than one supplier and to place in colocation facilities.
• Check the facility before the hardware. Power density and cooling limits will decide whether on-premises AI is practical. Many enterprises will find that colocation or a packaged AI factory is faster than retrofitting.
• Watch the first configurations. AMD's first Helion builds, expected in November, and partner products based on Nvidia's Vera CPU rack design will show whether the agentic data center is something enterprises can buy or still mainly a hyperscaler design.

The headline from San Jose was openness. The more lasting change may be that the industry has started designing for agents, not just models.

Frequently asked questions

Why do AI agents need more CPUs than chatbots?

Agents call APIs, query databases, run code and coordinate workflows around the model. That work runs mostly on CPUs, while GPUs handle model inference.

What is AMD Helion?

An open rack blueprint that combines EPYC CPUs and Instinct GPUs for agentic AI. A CPU-only rack holds up to 32,768 cores, and AMD plans to contribute the specification to OCP.

Do open OCP designs reduce enterprise AI infrastructure costs?

They can lower the cost of building and running capacity by widening supplier choice, but key chips, networking and software remain concentrated among a few vendors.

Filed under: Data & AI Infrastructure, Agentic AI