AMD Threadripper Halo Station: Desktop AI Workstation for Trillion-Parameter Models

AMD unveiled the Threadripper Halo Station at IFA Berlin, a desktop workstation that packs data-center AI specs into a form factor that could sit under a desk. With 2 TB of system memory, 576 GB of HBM3E, a 96-core Threadripper Pro CPU, and two MI350P accelerators, the system is designed to run AI models with over a trillion parameters locally. Pricing has not been announced, but analysts expect it to exceed EUR 100,000, putting it in enterprise and institutional territory rather than consumer PCs. Release is planned for early next year.

A desktop workstation built for trillion-parameter models

The Threadripper Halo Station is AMD's answer to the Nvidia DGX Station, but with a memory advantage. The system delivers up to 576 GB of HBM3E memory, more than three times what Nvidia's competing workstation offers, according to AMD. That memory capacity is critical for loading and running large language models that exceed what even high-end cloud instances can cache locally. The workstation also includes a 96-core Threadripper Pro 9995WX CPU and two MI350P GPUs, with the option to expand to four GPUs via a liquid-cooled chassis specs from ITPro. AMD demonstrated the machine generating an entire 3D world and flight simulator from a single prompt at the show.

How the Threadripper Halo Station stacks up

The direct competitor is Nvidia's DGX Station, which retails around EUR 100,000 and is already available. The Halo Station offers significantly more local memory and bandwidth, which could matter for teams running very large models or multi-agent systems that need to keep many layers in memory simultaneously. However, Nvidia's ecosystem provides more seamless code portability between local workstations and its cloud platform. AMD's advantage is raw capacity; Nvidia's advantage is integration. For teams that prioritize data sovereignty or want to avoid escalating cloud inference costs, the Halo Station's on-premise approach may be appealing. Cloud AI subscription costs add up quickly for heavy users, and a workstation like this can amortize over time if utilization is high.

What this means for AI builders

For AI builders, the Threadripper Halo Station signals a growing shift toward on-premise AI infrastructure. A desktop system that can handle trillion-parameter models locally reduces dependence on cloud GPU availability and data egress fees. This matters for teams working with sensitive data, requiring low-latency inference, or running long agentic workflows where cloud costs become unpredictable. The workstation is not for everyone -- the expected price tag of over EUR 100,000 limits it to well-funded enterprise R&D labs, academic institutions, and defense contractors. For smaller teams, cloud remains the more flexible and capital-efficient option.

Limitations to keep in mind

Several caveats apply. Pricing and final specifications have not been confirmed by AMD beyond what was presented at IFA. The system relies on liquid cooling and will require significant desk space and power. Code portability from cloud to local is less seamless than Nvidia's offering, especially for teams using Nvidia-optimized frameworks. And the workstation is only as useful as the models that can run on it -- most open-weight models under ~400 billion parameters can already run on high-end consumer GPUs. The trillion-parameter use case is still niche. Finally, the actual release date and volume availability remain unannounced. Early adopters should expect a wait and a premium price.

For AI teams already running heavy on-premise inference or training workloads, the Threadripper Halo Station is worth watching. For everyone else, the announcement is a signal that the local AI compute landscape is becoming more competitive, which may drive down costs across the board over time.