AMD’s Threadripper Halo Station Brings Trillion-Parameter AI to a Deskside Workstation

AMD used IFA Berlin 2026 to introduce the Threadripper Halo Station, a prototype liquid-cooled deskside workstation designed to put data-center-class AI hardware in a single system for developers, researchers, and organizations. The central pitch is local AI: AMD says the platform can run models with more than one trillion parameters without a cloud connection. 47

That statement is AMD’s positioning and capacity claim for the highest-memory configuration. Whether a particular trillion-parameter model can be used effectively will still depend on its architecture, precision or quantization, the task being performed, and software support.

What AMD announced

AMD describes Halo Station as a new class of workstation that brings “supercomputer-class” compute to individual users and developers—hence its personal supercomputer framing. It is not a conventional consumer desktop: the system is intended for high-throughput local AI inference and development workloads. 412

The workstation’s local-first approach is aimed at teams that want to keep AI workloads and associated data in their own environment rather than relying entirely on hosted cloud infrastructure. AMD’s IFA materials specifically position it for training and running massive models locally. 16

The hardware: 96 CPU cores and up to four Instinct accelerators

At the center of the design is the AMD Ryzen Threadripper PRO 9995WX, with 96 cores and 192 threads. The platform supports up to 2TB of DDR5 RDIMM system memory. 411

For AI acceleration, Halo Station is designed to support up to four AMD Instinct MI350P accelerators. Each MI350P carries 144GB of HBM3e memory, so a four-accelerator configuration would provide up to 576GB of accelerator memory. 424

AMD quotes maximum platform figures of:

• Up to 2.6TB of combined memory
• Up to 16.4TB/s of total system memory bandwidth
• Up to four Instinct-class HBM accelerators 7

Those memory figures are a major part of the product’s argument: very large AI models require not only compute but enough fast memory to keep their weights and working data available locally.

What was actually shown at IFA

The system displayed at IFA was a liquid-cooled prototype with two MI350P accelerators, for 288GB of HBM3e in that demonstration configuration. That should not be confused with the proposed four-accelerator maximum of 576GB. 314

AMD calls Halo Station a prototype, rather than a finalized off-the-shelf workstation. Its official product page describes it as a system first shown at IFA 2026 that brings data-center-class AI compute into a single deskside machine. 7

Why it is being compared with Nvidia DGX Station

Halo Station targets the same broad category as Nvidia’s DGX Station: workstation-shaped systems designed to give organizations substantial AI compute without deploying a full rack server. The difference AMD is emphasizing is the use of Instinct data-center accelerators in a deskside form factor. 712

AMD also compares the maximum Halo Station configuration favorably on total memory capacity against Nvidia’s DGX Station reference platform. That is an AMD comparison based on the respective configurations, not an independent performance test. 25

For buyers, raw memory capacity will be only one consideration. The software environment, framework compatibility, workload tuning, support model, power and cooling requirements, and real-world performance will all matter alongside specifications.

Software and the local-AI case

The system is intended to use AMD’s ROCm software ecosystem for AI and high-performance computing workloads. ROCm is AMD’s alternative to Nvidia’s CUDA-oriented development stack, and prospective buyers will need to assess whether their specific tools and models are well supported on it. 27

The appeal of running models locally is control: teams can keep sensitive prompts, datasets, and model workflows inside their own infrastructure. But Halo Station is designed for institutions and specialized professionals with unusually demanding AI workloads—not for typical desktop users.

Availability and price remain open questions

AMD had not disclosed a final price, final configuration lineup, or named OEM partners at the time of the announcement. 2628 AMD-associated reporting indicates the platform is expected in 2027, but the route to market and final systems have yet to be detailed. 25

Media estimates have placed a suitably equipped system around $100,000 to $150,000, but those figures are not AMD-confirmed pricing and should be treated as estimates rather than a purchase price. 23

The significance of Halo Station is therefore less about a product anyone can order today and more about AMD’s attempt to create a local-AI workstation tier: a deskside machine with enough CPU cores, accelerator memory, and bandwidth to bring some frontier-scale model workloads out of the cloud and into an organization’s own environment.