AMD's Threadripper Halo Station packs 96 cores and 576GB of HBM3E for local trillion-parameter AI

At IFA 2026, AMD introduced the Threadripper Halo Station, a liquid-cooled workstation aimed at running trillion-parameter AI models entirely on-premises. For builders managing sensitive data or high inference costs, this is a serious alternative to cloud GPU clusters. But with no pricing, availability, or final system specs, it's still a future consideration.AMD unveiled the Threadripper Halo Station at IFA 2026, a liquid-cooled workstation designed for running AI models with more than one trillion parameters locally. The system pairs the 96-core Ryzen Threadripper PRO 9995WX processor (Zen 5, up to 5.4 GHz) with up to four Instinct MI350P accelerators. Each MI350P card includes 144 GB of HBM3E memory and delivers 4 TB/s of memory bandwidth. A fully configured four-card version would provide 576 GB of HBM3E memory, enough to hold a trillion-parameter model in memory.AMD said the MI350P's 4 TB/s bandwidth is 14 times higher than any LPDDR5X variant. The CPU supports up to 2 TB of DDR5 system memory, and the system uses independent liquid cooling for both the accelerators and the processor.

Why this matters for on-premises AI deployment

For teams building large models, the Halo Station offers a path to keep data and computation entirely on-premises. This avoids cloud egress costs, data residency concerns, and the unpredictable latency of shared GPU instances. The combination of a 96-core CPU and massive GPU memory is suited for model fine-tuning, inference over large context windows, or running agent workflows that require persistent in-memory state. AMD described it as bringing "supercomputer-class compute" to individual users and developers.

The Threadripper Halo Station directly competes with NVIDIA's DGX Station, offering a similar form factor for local AI workloads. If AMD delivers on the promised specifications, it could be a strong option for research labs, defense contractors, and enterprises that need to process sensitive models without sending data to the cloud.

Power, cooling, and the open questions

The hardware is not plug-and-play for a standard desk. Each MI350P accelerator can draw up to 600 watts of total board power, so a four-card configuration alone pulls 2.4 kW plus the CPU. AMD said the cards are independently liquid-cooled, and the Threadripper processor has its own liquid-cooling setup. That means the Halo Station requires a dedicated space with adequate cooling and electrical capacity.

AMD has not announced pricing, availability, or full system specifications. It also did not detail storage, networking, or chassis options, or whether it plans to sell the system directly or through workstation manufacturers and system integrators. The demo at IFA showed a two-card version, leaving questions about the four-card configuration's physical design and thermal performance. Until pricing is known, builders cannot assess whether this is cost-competitive with renting cloud GPUs.

Given the unknowns, the Halo Station is a compelling reference design rather than a current purchasing option. If your workload demands on-premises control and can tolerate high up-front cost and power requirements, this system is worth watching. Otherwise, cloud instances remain the pragmatic choice for most teams until AMD or its partners ship concrete products with clear pricing.