AMD Instinct MI355X.

The next step in high-memory AMD inference capacity. MI355X brings liquid-cooled density and a larger memory envelope to full-scale, sovereign serving deployments.

Request capacity

At a glance

More memory per accelerator. More room to operate.

MI355X is designed for teams preparing to serve larger models, more concurrent users, and regional workloads without compromising on isolation or control.

HBM3E memory

288 GB

Memory bandwidth

8 TB/s

Accelerator power

1,400 W

Eight-GPU memory

2.3 TB

Platform profile

Density without a new operating burden.

The hardware advances, while the operating experience stays direct: a platform that turns capacity into isolated compute through the terminal UI, API, and CLI.

Liquid-cooled density

Direct liquid cooling allows the platform to sustain a larger power envelope while fitting an eight-accelerator UBB 2.0 system into a deployable rack design.

A full-scale serving unit

Eight OAM accelerators connect across an AMD Infinity Fabric mesh, creating a coherent, high-memory system for large inference deployments.

An operating platform, not just hardware

ROCm, hardened virtualization, and the automated layers beneath Kubernetes are part of the environment, not a follow-on integration project.

Deployment path

Built to expand deliberately.

Our approach is smaller, well-operated deployments in more data centers, not a single oversized buildout. MI355X capacity fits that plan.

A practical delivery path

Our relationships with AMD, Dell, Broadcom, and Advizex let us plan smaller regional deployments with realistic supply, networking, and service expectations.

A platform that already operates

We are applying three years of automation and operator feedback to MI355X capacity, rather than building a new control plane around a new accelerator.

MI355X capacity

Plan the deployment before capacity arrives.

Share the deployment you are planning. We will use it to align hardware, networking, isolation requirements, and a realistic delivery path.

Capacity requests start a planning conversation; they are not a commitment to purchase.

Available now

Need inference compute today?

MI300X capacity is available through the same platform, with isolated VMs and bare metal provisioned without a sales handoff.

Explore MI300X