The Threadripper Halo Station packs up to 576 GB of HBM3E memory and aims to run trillion-parameter AI models without touching the cloud
AMD just threw a very expensive gauntlet at Nvidia’s feet. At IFA 2026 in Berlin, the chipmaker unveiled the Threadripper Halo Station, a desktop workstation it’s calling a “personal supercomputer,” built to run massive AI models entirely on local hardware. The target audience: developers and small teams who’d rather not ship their data to a cloud provider every time they want to run inference on a trillion-parameter model.
The machine is a direct shot at Nvidia’s DGX Station, the reigning heavyweight in high-end AI workstations. AMD SVP Jack Huynh introduced the system on September 4, framing it as the start of a “completely new era of computing” driven by artificial intelligence.
What’s under the hood #
At its core sits a 96-core Threadripper Pro 9995WX CPU, AMD’s most powerful desktop processor to date.
On the GPU side, the system ships with dual Instinct MI350P accelerators, expandable to four units, all liquid-cooled. Those four GPUs collectively deliver up to 576 GB of HBM3E memory. System memory tops out at 2 TB of DDR5.
AMD claims the configuration can handle trillion-parameter AI models without relying on cloud infrastructure. That’s a meaningful distinction. Running models of that scale typically requires renting time on clusters of GPUs in data centers operated by the likes of AWS, Google Cloud, or Microsoft Azure.
Why local AI hardware matters now #
The pitch for on-premises AI computing has gotten louder over the past two years, and for good reason. Cloud-based AI inference comes with three persistent headaches: latency, cost, and data sovereignty.
Latency is straightforward. Sending data to a remote server and waiting for a response adds time that compounds across millions of queries. For real-time applications, like AI agents making sequential decisions, that round-trip delay can be a dealbreaker.
Then there’s data sovereignty, which matters most in regulated industries like finance, healthcare, and defense. Keeping sensitive data on a local machine eliminates an entire category of compliance risk.
AMD is betting that enough buyers care about all three of these problems to justify a workstation that will likely cost well into six figures when fully configured. Industry estimates peg the Halo Station’s price in that range, though AMD hasn’t disclosed official pricing.
The Nvidia problem #
Nvidia’s DGX Station has essentially owned the high-end AI workstation category since its introduction. The company’s CUDA software ecosystem, which developers have built on for over a decade, creates a moat that’s hard to cross with hardware specs alone.
AMD has been investing heavily in its ROCm software stack, the open-source alternative to CUDA, to make it easier for developers to port their AI workloads to AMD hardware.
On raw specs, AMD is making a credible case. The company claims the Halo Station offers higher system memory than Nvidia’s competing offering, which could matter for workloads that need to keep entire large models in memory rather than swapping data in and out.
The Halo Station builds on AMD’s earlier AI push with the Ryzen AI Halo, a consumer-facing chip designed for AI-accelerated laptops and desktops.
Market implications and what to watch #
The expected availability window of 2027 means investors won’t see revenue impact anytime soon. Analysts will be watching two things closely: whether AMD can deliver the system on schedule, and whether the ROCm ecosystem matures enough to make the hardware accessible to developers who’ve spent years building on CUDA.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our