OCP Summit 2026: Intel's D-Series, AMD's Helion, Nvidia's 250kW Rack

The reference rack coming out of OCP Global Summit 2026 is measured in kilowatts. Over the week of 12 October, CRN collected the summit announcements into a single roundup, and the pattern in them is that power delivery and CPU density now set the limit on what an operator can order. Nvidia contributed a third-generation MGX reference design rated to 250kW per rack. Intel showed server racks carrying more than 46,000 Xeon 6+ cores inside a 133.45kW envelope. AMD put dates on open rack blueprints that hardware vendors will start shipping in November.

Two of those three are facility projects. The third is a card aimed at the racks that already exist.

Intel's D-series fits the slot you already have

Intel's 480GB "Crescent Island" inference accelerator will ship as the Intel GPU D-series. The company puts it at up to 1.3 petaflops of FP8 throughput in a 350W envelope, as a standard air-cooled PCIe card, with up to eight cards in a server. It uses the Xe3P architecture, carries roughly 1.5TB/s of memory bandwidth on LPDDR5X, and holds models up to 300B parameters on one card or up to 1T parameters across an eight-card system. Intel says the part is sampling with customers now.

The comparison Intel is inviting is with Nvidia's RTX Pro 6000 Blackwell Server Edition, which CRN lists at 2 petaflops FP8, 4 petaflops FP4, 96GB of GDDR7 and up to 600W. On paper the Nvidia card wins the slot on throughput. The Intel card's argument is the 350W figure. That is a card you can add to a rack with no liquid loop and no spare 600W of headroom per slot, which describes most of the installed base.

Intel named Asus, Cisco, Dell, HPE, Lenovo, Supermicro, NetApp, Nutanix, Red Hat and VMware as ecosystem partners. Partner lists are statements of intent rather than shipping products, and the D-series itself is still sampling. Nothing in the summit material puts a general availability date on it.

The rack is now the unit of specification

Intel's own rack designs, built by Asus, Gigabyte and Supermicro, pack more than 46,000 Xeon 6+ cores into a single rack at a 133.45kW power envelope, many of them on Open Rack v3. Xeon 6+ (Clearwater Forest) carries up to 288 efficiency cores at 450W. That is a large number of cores per rack for workloads that spend their time waiting on agent orchestration and tool calls rather than on matrix multiplication, and it is the reason Intel is talking about density rather than about FLOPs.

AMD's Helion is an open reference rack blueprint rather than a product. It takes EPYC CPUs and Instinct GPUs, or CPUs alone. Cisco, Dell, HPE, Lenovo and Supermicro plan Helion configurations starting November 2026. A CPU-only rack built on 6th-Gen EPYC can hold up to 32,768 cores. The all-in-one option mixes EPYC CPU nodes, Instinct MI350P GPU nodes and general-purpose compute, in liquid-cooled and air-cooled variants.

AMD also introduced Verano, a 6th-Gen EPYC part with up to 192 cores and the first in the family to carry an LPDDR memory interface. AMD positions it for the lowest system power in the range and as a host node for GPUs. Chief executive Lisa Su put production in the first half of 2027.

Read the Helion and Xeon numbers together and the buying decision changes shape. A rack with 32,768 cores is a power and cooling problem before it is a compute problem. The accelerator count is no longer the figure that decides whether a design fits a hall.

Nvidia's 250kW reference and a 5,000-amp busbar

Nvidia contributed MGX 1.2, the third generation of its MGX reference design, to OCP. It supports Vera Rubin NVL72 systems and Vera CPU racks, and adds rack-scale blind-mate liquid cooling, an interoperable rack-scale interconnect, a liquid-cooled 5,000-amp busbar, configurable I/O bays and fanless liquid-cooled compute trays. Nvidia states support for rack deployments up to 250kW.

The 5,000-amp busbar is the detail worth pausing on. A rack drawing that current is not something an operator retrofits into a populated hall over a weekend. The reference design assumes a facility built for it, or a rebuild of the power distribution in front of it, which puts MGX 1.2 in a different procurement category from an air-cooled PCIe card.

What to diarise

For anyone holding a hardware budget, the summit produced three dates and a set of envelopes to plan against. Intel's D-series is sampling, AMD's Helion configurations arrive from November, and Verano production sits in the first half of 2027. The power figures are the ones that determine whether a design is a drop-in or a construction project.

Diarise
  now            Intel GPU D-series sampling with customers
  November 2026  AMD Helion rack configurations from Cisco, Dell, HPE, Lenovo, Supermicro
  H1 2027        AMD Verano (6th-Gen EPYC, up to 192 cores) production

Power and density
  Intel D-series        up to 1.3 PFLOPS FP8, 350W, air-cooled PCIe, up to 8 per server
  Intel Xeon 6+ rack    more than 46,000 cores, 133.45kW
  AMD Helion (CPU only) up to 32,768 6th-Gen EPYC cores
  Nvidia MGX 1.2        rack deployments up to 250kW, 5,000A liquid-cooled busbar
  Nvidia RTX Pro 6000   2 PFLOPS FP8, 4 PFLOPS FP4, 96GB GDDR7, up to 600W

Attendance at the summit was reportedly 15,000. The organiser said the event reached maximum capacity and would refuse entry to unregistered attendees, which says as much about demand for the reference designs as any of the specification sheets do.

Treat every number here as vendor-supplied

The figures above come from the vendors, either directly or through CRN's roundup. Intel's petaflop and parameter counts are Intel's. AMD's core counts and dates are AMD's. Nvidia's 250kW ceiling is Nvidia's. None of it was independently benchmarked in the reporting available from the summit, and the numbers that decide a purchase, sustained throughput on your own workload and actual draw under load, are the two a specification sheet does not provide.

What the week does establish is the direction of the reference designs. Power per rack and cores per rack have moved to the front of the document, and the retrofit market now has at least one accelerator aimed squarely at it. An operator with an air-cooled hall and a tight power budget has a decision to make in the next two quarters. An operator planning a liquid-cooled build has a longer list of reference designs to choose from, and a harder conversation with the facilities team.