576GB of HBM3E under a desk: AMD's Halo Station is aimed squarely at Nvidia's DGX

AMD used its IFA 2026 keynote to announce Threadripper Halo Station, a workstation built around a claim it is happy to make loudly: that it is the most powerful workstation in the world.

The specification is the argument. A 96-core, 192-thread Threadripper PRO 9995WX on Zen 5, boosting to 5.4GHz, supporting up to 2TB of DDR5 system memory — and up to four Instinct MI350P accelerators, each carrying 144GB of HBM3E at 4TB/s.

At four cards that is 576GB of high-bandwidth GPU memory in a single desk-side chassis. AMD's presented configuration uses two cards for 288GB, with the upgrade path to four.

Everything significant in the system is liquid cooled, which AMD frames as a noise decision as much as a thermal one. A machine intended to sit next to a person has different constraints from one in a rack.

The One Use Case

AMD has been unusually direct about what this is for. The system is designed around running AI models with more than one trillion parameters locally, without a cloud connection.

The memory figure is what makes that plausible. Model weights have to fit somewhere, and at 4-bit precision a trillion-parameter model lands within reach of 576GB. The MI350P's bandwidth — which AMD puts at fourteen times that of any LPDDR5X variant — is the difference between a model that technically loads and one that responds at usable speed.

This is the specification that determines whether local inference at frontier scale is real or theatrical, and AMD has chosen to compete on exactly that number.

The Target

The comparison AMD wants is with Nvidia's DGX Station, and the positioning is transparent.

Nvidia has spent years making the desk-side AI workstation its own category, and AMD's pitch is that supercomputer-class compute should be available to an individual developer rather than rationed through a cloud queue. Whether that lands depends almost entirely on software — Nvidia's advantage has rarely been silicon alone, and CUDA remains the reason many teams do not seriously evaluate alternatives.

Hardware parity does not resolve that. It only makes the question worth asking again.

Who It Is Actually For

The honest audience for this machine is narrow, and worth naming.

It is for teams working on data they are not permitted to send anywhere — medical records, defence work, unreleased financials, anything under a jurisdictional constraint. It is for researchers whose iteration speed is bottlenecked by cloud scheduling rather than raw compute. And it is for organisations that have run the arithmetic on sustained inference and found that renting indefinitely costs more than owning.

It is not for most developers, who will continue to rent because renting is cheaper until utilisation is high and constant.

The Missing Number

AMD has not announced pricing, availability, or complete system specifications.

That absence matters more than usual here, because the entire economic case rests on it. A local machine competes with a cloud subscription, and without a price the comparison cannot be made. Four MI350P cards represent serious silicon, and the plausible range for a fully configured system starts well above what most individual developers will fund themselves.

Until AMD publishes a number, 'available to individual developers' is a positioning statement rather than a claim that can be tested.

What To Watch

The price, first and above all. Whether the four-card configuration ships at launch or arrives later. And whether the software stack has closed enough of the gap that a team can move a working pipeline across without rewriting it — because that, not memory capacity, is what has kept this category Nvidia's.