AMD Instinct in 2026: Helios racks, MI355X servers and the MI350P card

At its Advancing AI event in San Francisco on 23 July 2026 AMD launched Helios, its first rack-scale AI system. According to Techwire Asia’s report, each rack holds 72 Instinct MI455X GPUs and 18 sixth-generation EPYC Venice processors, with AMD claiming up to 2.9 exaflops of peak FP4, 31 TB of HBM4 and 1.7 PB/s of memory bandwidth. Scale-up runs over UALink on Ethernet, scale-out uses Ultra Ethernet-aligned networking with Pensando Vulcano 800 NICs, and the rack follows the Open Rack Wide specification.

The named customers are large. OpenAI expects to begin bringing Helios online in the fourth quarter of 2026 as the first phase of a six-gigawatt agreement, Meta is testing Helios racks, and Anthropic is preparing to install up to two gigawatts of MI455X. Hardware partners listed for the launch were Bull, HPE, Lenovo, Supermicro, Sanmina and Wiwynn. On AMD’s second-quarter earnings call on 4 August, Lisa Su said Helios was in production with initial shipments on track to begin later that quarter, ramping through the fourth quarter and into 2027. AMD also set out MI500 for 2027 and MI600 for 2028.

What enterprise buyers can order

For organisations below hyperscale, the relevant products are the ones that fit existing data centres. The MI355X is offered in OEM eight-GPU servers; Dell’s PowerEdge XE9785L, for example, is a 3 OU liquid-cooled node for the IR7000 rack with eight MI355X 288 GB OAM accelerators at 1,400 W each and two EPYC 9005 processors, and the same chassis can instead be configured with NVIDIA HGX B300. On 7 May 2026 AMD also introduced the Instinct MI350P, a dual-slot PCIe card with 144 GB of HBM3E for standard air-cooled servers, up to eight per system, with AMD claiming up to 4,600 TFLOPS at MXFP4. SDxCentral reported Lenovo, Cisco, Dell and HPE among the companies lining up for it.

The MI350P has no high-speed card-to-card fabric, so multi-card traffic runs over PCIe. It suits inference and retrieval workloads that fit on one or a few cards, not large-scale training.

What it means for buyers

Treat Helios as a 2027 option for most projects outside the named customers, and plan near-term deployments on MI355X servers or MI350P cards. Specify cooling first: a 1,400 W OAM accelerator in the XE9785L requires direct liquid cooling, while the MI350P is designed for air-cooled racks. Confirm that your software stack is validated on ROCm at the version the OEM ships, and ask for the OEM’s support term and firmware update policy in writing. Lead times for OEM Instinct servers depend on the specific chassis and cooling option, so request them per configuration. Instinct accelerators are subject to the same US export licensing framework as comparable NVIDIA parts, so end-user documentation is needed at quotation stage.