AMD Threadripper Halo Station: A Desktop Supercomputer for On-Prem AI Builders

AMD has unveiled the Threadripper Halo Station, a liquid-cooled desktop workstation designed to run trillion-parameter AI models locally. For AI builders weighing on-prem deployment against rising cloud costs, the Halo Station offers a high-memory, high-bandwidth alternative to Nvidia's DGX Station, but final pricing and real-world performance remain unconfirmed.

What the Threadripper Halo Station brings to a desk

The system packs a 96-core Threadripper Pro 9995WX CPU, up to two terabytes of system memory, and 576 GB of HBM3E high-bandwidth memory. Two MI350P data center accelerators handle the compute load, all kept cool by liquid cooling. AMD SVP Jack Huynh demonstrated the system at IFA by generating a 3D world and flight simulator from a single prompt, claiming it is the “most powerful workstation in the world.”

The memory advantage over Nvidia’s DGX Station

The Halo Station’s direct competitor is the Nvidia DGX Station. AMD claims its workstation offers more than three times the memory of Nvidia’s offering, which allows it to hold larger models entirely in memory for local inference. However, Nvidia’s system is already shipping at around €100,000 and provides smoother integration when moving workloads to the cloud. The trade-off is clear: AMD bets on raw local capacity, while Nvidia leans on ecosystem lock-in.

On-prem AI for enterprises: the practical shift

As cloud AI subscription costs climb, more organizations are looking to run large models on their own infrastructure. The Threadripper Halo Station is explicitly aimed at corporate and institutional buyers, not individual developers. If the specs hold, it could enable teams to deploy trillion-parameter models locally without per-token API costs. That matters for workflows that require high throughput, data privacy, or low latency where cloud round-trips are unacceptable.

What’s still unknown: price, benchmarks, and timeline

AMD has not disclosed the final price. Industry estimates suggest it could exceed €100,000. The release window is “early next year,” and independent benchmarks have not been published. All performance claims, including the trillion-parameter capability and the 3D demo, come from AMD’s own presentation. Until third-party testing validates real-world inference speed, power draw, and software compatibility, builders should treat the promises as directional.

For teams already spending six figures on cloud AI compute, the Halo Station is worth watching. For everyone else, waiting for confirmed pricing and community benchmarks before planning a purchase is the sensible move.