Threadripper Halo Station: Specs, 576GB HBM3E & Price (2026)

On September 4, 2026, AMD unveiled the Threadripper Halo Station, a liquid-cooled desk-side AI workstation, during its IFA 2026 keynote in Berlin. The system pairs a Threadripper PRO 9995WX CPU (96 cores/192 threads) with up to four AMD Instinct MI350P accelerators, delivering up to 576GB of HBM3E GPU memory. AMD calls it 'the most powerful workstation in the world' and says a 4-bit quantized trillion-parameter-class model can fit within GPU memory. Pricing, availability, and detailed real-world performance have not been announced as of this writing; this article separates confirmed facts from press speculation.

If you want to check whether a specific model will run on your own GPU or Mac, our free VRAM calculator estimates required VRAM just by picking a model, quantization level, and context length. For background on sizing local hardware, see our local AI hardware requirements guide.

AMD claims '3.4x system memory and more than 2x memory bandwidth' versus the DGX Station, but these figures describe hardware capacity and bandwidth, not guaranteed inference throughput. A larger memory pool doesn't automatically translate into faster real-world performance — the maturity gap between ROCm and the CUDA ecosystem, along with overhead from model sharding and offloading, can significantly affect actual results. This tension is also discussed in our comparison of local LLM inference engines.

Let's sanity-check AMD's claim that a trillion-parameter-class model can fit in GPU memory at 4-bit quantization. At 4 bits (0.5 bytes per parameter), a model's weight footprint is roughly parameter count x 0.5 bytes. For a trillion-parameter model, that's approximately 500GB just for the weights — consuming most of the 576GB of GPU memory. On top of that, inference requires a KV cache, whose size grows with context length and batch size, leaving limited headroom for long contexts or concurrent requests. How much actually fits depends heavily on the model architecture (dense vs. MoE) and context length settings, so the figures below are rough estimates only. For more on the relationship between KV cache and context length, see our KV cache and VRAM guide.

Research institutions and universities that need on-premises infrastructure to evaluate large-scale models
IT departments at companies that can't send sensitive data to external clouds and need local fine-tuning or inference infrastructure
Large enterprises building shared internal AI infrastructure that must serve multiple large models simultaneously

Individual developers or small teams doing experimentation — on-demand cloud GPUs are typically more cost-effective
General business use cases where models in the tens or low hundreds of billions of parameters suffice — a Mac Studio or consumer RTX-class GPU is often enough (see our piece on running huge MoE models with modest memory)
Organizations that need to finalize a budget now — pricing and availability aren't public yet, so planning is premature

Official pricing (the $100K-$150K figures are speculative, not confirmed)
Whether and when it will be sold in Japan
How well major frameworks (PyTorch, vLLM, etc.) run on ROCm, and real-world throughput
Detailed power consumption, electrical requirements, and footprint
Actual benchmark results with a full 4-card MI350P configuration
Whether the rumored 2027 shipping timeline holds