These disclosures included new details for Intel’s upcoming data center GPU, new server racks packing more than 46,000 cores’ worth of Intel Xeon CPUs, a new flexible server rack reference design from AMD as well as open-source and Vera Rubin news from Nvidia.
Interest in the future of data center infrastructure is so high that the organizer of this year’s OCP Global Summit in Silicon Valley said it has reached “maximum capacity” and would refuse entry for anyone not registered, even those who already booked travel.
With a record 15,000 people reportedly attending the event organized by the Meta-founded Open Compute Project, major semiconductor companies Intel, AMD and Nvidia showed off how they will power future AI data centers with new disclosures, many of which are based around the standardized server specifications championed by OCP.
[Related: How Intel’s AI Chip Strategy Is Coming Into Shape After Years Of Struggles]
These disclosures included new details for Intel’s upcoming air-cooled data center GPU, new server racks packing more than 46,000 cores’ worth of Intel Xeon CPUs, a new flexible server rack reference design from AMD, the first AMD EPYC CPU to include an LPDDR memory interface as well as open-source and Vera Rubin news from Nvidia.
Some of these announcements were aimed at partners and customers building data centers for agentic AI workloads, especially when it came to the CPU-related news because of how important the processor type has become for enabling such applications. Others, on the other hand, are meant to help with the broader AI data center buildout.
While Nvidia is far ahead of Intel and AMD in the AI infrastructure market in terms of revenue, all three companies are needed to help channel partners fulfill the ongoing buildout of AI data centers, according to Bob Venero, CEO of Fort Lauderdale, Fla.-based solution provider Future Tech Enterprise, which ranked No. 90 on CRN’s 2026 Solution Provider 500.
“The channel needs AMD, Intel and Nvidia to continue to push the limits,” he said.
Though hyperscalers and neoclouds are building much of the data center capacity to support agentic AI right now, Venero sees a huge opportunity for the channel to help enterprises secure their own infrastructure due to security and cost concerns.
“Customers are definitely figuring this out because they spent a ton of money and blew through their token budgets because they let [AI] agents run unregulated within their organizations,” he said. “So you’re seeing customers now more looking at, ‘Hey, how do we control costs? How do we drive these agents internally that will use our internal models and do things that are not necessarily cloud-based but on-prem?’”
What follows are the most important details from these six big OCP 2026 announcements.
Intel Details 480-GB D-Series GPU For Air-Cooled Servers
Intel revealed that its upcoming 480-GB “Crescent Island” data center GPU for AI inference will be known as the Intel GPU D-series and feature up to 1.3 petaflops of 8-bit floating-point compute within a 350-watt power envelope.
The chipmaker said that the D-series GPU, which is sampling with customers now, will come in a standard, air-cooled PCIe form factor that will make it easy to slot into existing data centers, with up to eight PCIe cards per server.
Intel said it’s working with a “broad ecosystem” of OEMs, ODMs, neoclouds, ISVs, appliance providers and AI-native companies to bring the GPU D-series into production. These vendors include Asus, Cisco, Dell Technologies, HPE, Lenovo, Supermicro, NetApp, Nutanix, IBM-owned Red Hat and Broadcom-owned VMware.
The GPU D-series represents Intel’s latest attempt to gain a foothold in the AI accelerator chip market after previous efforts, including its Gaudi line of chips, failed to catch traction. The new product makes use of the company’s new Xe3P architecture, which stems from the Xe architecture it began developing for client and server GPUs several years ago.
The semiconductor giant said the GPU D-series is designed for large language models, long-context workloads and concurrent services, with a single card supporting models with up to 300 billion parameters and an 8-GPU system supporting up to 1 trillion parameters.
The new product line is set to go head-to-head with Nvidia’s RTX Pro 6000 Blackwell Server Edition GPU, which features 2 petaflops of 8-bit floating-point compute (FP8), 4 petaflops of 4-bit floating-point compute (FP4) and 96 GB of GDDR7 memory in a PCIe form factor that can be configured to consume up to 600 watts.
While Nvidia does not divulge the difference in performance between the air- and liquid-cooled versions of the RTX Pro 6000 on its website, a head-to-head comparison of public specifications to the Intel GPU D-series puts the former card ahead in terms of FP8 performance and offering FP4 support, which can help applications run faster.
The GPU D-series, however, packs far more memory, 480 GB, with roughly the same memory bandwidth at 1.5 TBps despite using LPDDR5X. And it seems to offer greater power efficiency by fitting 1.3 petaflops of FP8 compute into only 350 watts.
Intel said the GPU D-series falls in line with its strategy to offer open systems and software architecture that allows developers to easily deploy workloads on the right hardware.
New Intel Xeon Server Racks Pack 46,000 CPU Cores For Agentic AI
Intel showed off new server rack designs that can pack more 46,000 cores’ worth of the company’s Xeon 6+ CPUs for agentic AI workloads.
These rack designs come from server vendors such as Asus, Gigabyte and Supermicro. Many of them are based on the Open Compute Project’s Open Rack v3 standard to, in Intel’s words, “simplify deployment and integration within open data center infrastructure.”
The semiconductor giant made the display as it experiences a surge in demand for CPUs due to the meteoric rise of agentic AI workloads in data centers.
The server racks at the OCP Global Summit were designed with high densities of CPUs to “orchestrate workloads, process data, manage I/O and execute tasks that keep” accelerator chips such as GPUs running at full speed,” the chipmaker said.
In some configurations, these server racks can fit more than 46,000 cores from Intel’s Xeon 6+ processors in a rack power envelope of 133.45 kilowatts, according to the company.
Launched over the summer, the Xeon 6+ processors, code-named “Clearwater Forest,” feature up to 288 efficiency cores in a 450-watt power envelope.
AMD Reveals Flexible ‘Helion’ Server Rack Design For Agentic AI
AMD revealed a flexible server rack reference design called “Helion” that enables partners to customize compute needs for customers running agentic AI workloads.
Calling Helion an “open blueprint,” the chip designer said it’s based on an Open Compute Project spec-design that can use EPYC CPUs and Instinct GPUs—or just the CPUs.
AMD said server vendors plan to release Helion rack configurations starting this November. Supporting vendors include Cisco, Dell Technologies, HPE, Lenovo and Supermicro.
It’s yet another move by AMD to challenge Nvidia’s dominance of the AI infrastructure market and keep the pressure on other rivals such as Intel, particularly as agentic AI workloads creates unprecedented demand for CPUs.
The idea with Helion is to let customers “match CPU performance, core density and power with GPU acceleration for the work their AI agents perform,” AMD said.
This means customers can create “virtually any configuration” to meet their workload requirements, and this includes options for liquid- and air-cooled versions of Helion.
For instance, AMD said a CPU-only rack with 6th-Gen EPYC processors can pack up to 32,768 cores for hyperscalers that need “dedicated capacity for agent execution and connections to separate model inference infrastructure.”
Helion can also enable an all-in-one rack that brings together “AMD EPYC CPU nodes, AMD Instinct MI350P GPU nodes and general-purpose compute on one platform” for enterprise customers running agentic AI workloads, according to the company.
AMD Reveals 192-Core EPYC ‘Verano’ CPU With LPDDR Memory
AMD CEO Lisa Su revealed a new EPYC CPU with up to 192 cores that comes with an LPDDR memory interface for the first time in the processor family.
Code-named “Verano,” the CPU is part of AMD’s 6th-Gen EPYC “Venice” lineup, and it’s been “optimized for the lowest system power” with an updated CPU-to-GPU bandwidth that makes it “exceptionally good” as the host node in GPU servers, she said.
Verano will enter production in the first half of 2027, according to Su.
The CPU is part of AMD’s expanding lineup for the 6th-Gen EPYC family, which launched over the summer and includes variants for technical computing and high-performance computing as well as workloads that require high core counts with low power and high frequencies.
Nvidia Open-Sources Run:ai, Other AI Data Center Software
Nvidia said it has made software components for various aspects of running AI data centers, including Run:ai, available under open-source licenses.
The vendor said it’s open-sourcing these software components as part of its DSX AI Factory platform, which brings together software libraries, workflow guides and reference designs to help partners and customers maximize tokens per watt in AI data centers.
The push is part of Nvidia’s strategy to provide a “vertically integrated” platform that is “horizontally open,” which has become a common line by the company this year.
The most notable move is with Run:ai, the Kubernetes-based workload management and orchestration software that Nvidia acquired for a reported $700 million in late 2024.
When Nvidia announced the acquisition earlier that year, the company said it planned to use Run:ai to boost its DGX Cloud, which launched in 2023 as a full-stack commercial service that was meant to simplify AI training in public cloud environments.
However, the company switched gears in late 2025 when it reportedly restructured the DGX Cloud business and shifted its focus to an “internal cloud environment for building and operating AI,” as Nvidia’s web page for DGX Cloud now says.
With the open-source move, Nvidia said, “it is now possible for every NVIDIA DSX ecosystem partner and AI cloud to increase workload throughput, improve GPU utilization, and operate shared AI infrastructure more efficiently and effectively.”
This involves making available the full Run:ai platform as well as individual components: KAI Scheduler, GPU Fractioning Operator, KAI Resource Management and Workload Map. Partners and customers can determine whether to adopt the full platform, use it to build their own products or integrate individual components, according to Nvidia.
Outside of Run:ai, the company said it open-sourced software components for “configuring networks, validating clusters, managing racks and provisioning bare metal.”
These components, originally designed for DGX Cloud, consist of Nvidia Config Manager, Nvidia Cluster Readiness Engine, Nvidia Rack Management Service and Cluster API Provider for Nvidia Infrastructure Controller.
With the software components, Nvidia said, cloud providers and software partners can “integrate AI factory capabilities directly into their multi-tenant platforms.”
Nvidia Contributes Vera Rubin MGX Rack Design To OCP
Nvidia said it has contributed the third generation of its MGX AI data center reference design to help manufacturers build interoperable platforms around the AI infrastructure giant’s Vera Rubin NVL72 and Vera CPU server racks.
“Combined with open liquid-cooling technologies such as MQD and MQDB, the effort helps turn proven AI factory designs into industry standards that others can adopt, manufacture and scale,” the company said.
The MGX 1.2 specification, as it is called, features “rack-scale blind-mate liquid cooling, interoperable rack-scale interconnect infrastructure and liquid-cooled 5,000-amp busbar designed for higher-power AI systems,” according to Nvidia.
The vendor said the spec also expands the modularity of MGX with “configurable I/O bays, fanless liquid-cooled compute trays, standardized power and cooling interfaces.
“It improves assembly, test and service times to streamline manufacturing and field service to enable faster and more efficient infrastructure buildout,” the AI infrastructure giant wrote. “The architecture supports rack deployments up to 250kW while simplifying serviceability, deployment and infrastructure reuse.”
