AMD Unveils a $100,000-Plus Desktop Workstation Built to Run Trillion-Parameter AI Models

AMD used the opening keynote of IFA Berlin 2026 on Friday to unveil the Threadripper Halo Station, a liquid-cooled desktop workstation the company says can run artificial intelligence models exceeding one trillion parameters without a connection to the cloud. The machine, shown publicly for the first time at the trade show, pairs a 96-core Ryzen Threadripper PRO processor with as many as four AMD Instinct MI350P accelerators, the same silicon AMD sells into data centers, repackaged into a tower a single researcher could keep beside a desk.

The announcement came during a keynote titled "The Era of Personal AI: The Future of Imagination," delivered by Jack Huynh, AMD's senior vice president and general manager of its Computing and Graphics Group. AMD is not yet selling the system. According to its own product page, the Threadripper Halo Station is "a prototype system AMD is showing for the first time at IFA 2026," with pricing, full specifications and OEM partners still to be announced.

The numbers

At the center of the machine is the Ryzen Threadripper PRO 9995WX, a 96-core, 192-thread Zen 5 chip that boosts up to 5.4 GHz, carries 384 MB of L3 cache and draws 350 watts on its own. Sold as a standalone part, that processor already retails for roughly $11,000 to $12,000. AMD pairs it with up to 2 terabytes of DDR5 system memory across eight channels and 128 lanes of PCIe Gen5 connectivity.

The system's headline feature is memory attached to its GPUs. Each Instinct MI350P card carries 144 GB of HBM3e running at 4 terabytes per second of bandwidth; the unit AMD displayed at IFA used two cards, for 288 GB of accelerator memory, but the design scales to four cards and 576 GB of HBM3e. Combined with system RAM, AMD says the workstation can reach up to 2.6 terabytes of total memory and 16.4 terabytes per second of aggregate bandwidth. That is enough, the company says, to hold a model with more than a trillion parameters entirely in GPU memory at 4-bit precision — large enough, for instance, to run open-weight models on the scale of Moonshot AI's 2.8-trillion-parameter Kimi K3 by offloading part of the model to system RAM. As configured with two accelerators, memory, CPU and cooling, analysts covering the launch estimated a street price north of $100,000, with a four-accelerator build potentially exceeding $150,000.

How AMD got here

The Threadripper Halo Station extends a strategy AMD has been building for most of 2026 around what it calls "Personal AI" — the idea that as AI shifts from single-prompt chatbots to continuously running autonomous agents, running models locally becomes cheaper and more private than renting cloud GPU time by the hour. AMD laid out that thesis in more detail at its Advancing AI conference in July, where it framed local compute, data privacy and personal context as the three pillars of the approach, according to an AMD Newsroom post from that event.

The company's first product built on that pitch was the smaller Ryzen AI Halo, a roughly $4,000-class developer box built around AMD's "Strix Halo" mobile-class chip with 128 GB of unified memory, which The Register reviewed earlier this year. The Threadripper Halo Station is aimed, in AMD's words, at "the other end of the market" — individuals and small teams whose AI work has outgrown that entry-level hardware but who cannot justify a rack of data-center GPUs. It marks the first time AMD has put its data-center Instinct accelerators into a workstation chassis at all, effectively repurposing a server design — swapping an EPYC server processor for a Threadripper Pro chip — into something that plugs into a wall outlet in an office rather than a data-center power distribution unit.

Who is affected

AMD's own FAQ for the product describes the intended buyer as "individuals and small teams doing serious AI work who are currently constrained by cloud access or shared infrastructure — AI researchers, model developers, engineering and simulation users, and creators working with large generative models," as well as organizations that need to keep sensitive data and intellectual property on premises rather than in a shared cloud tenancy.

The system is also a direct shot at Nvidia, whose DGX Station — a workstation built around a 72-core Grace CPU and a 252 GB Blackwell Ultra GPU that also sells for around $100,000 — has effectively had the high-end local-AI-workstation category to itself since it was announced last year. AMD contends the Threadripper Halo Station offers up to 3.4 times the total system memory and more than double the memory bandwidth of Nvidia's machine, though independent benchmarks do not yet exist because no review units have shipped. The launch also lands amid a global shortage of high-bandwidth memory chips, as AI data-center demand for HBM has driven up prices and squeezed supply for consumer and workstation-grade memory alike — a backdrop that could complicate how quickly AMD and its partners can actually build the machine at volume.

AMD itself called the machine "the most powerful workstation in the world," a claim the company made in unveiling the system at IFA 2026, Tom's Hardware reported.

Reaction and what happens next

Hardware press covering the keynote in Berlin were struck by the gap between AMD's ambitions for the machine and what it was willing to confirm. ServeTheHome and StorageReview both noted that AMD showed a working, liquid-cooled unit on stage — built on an off-the-shelf Asus workstation motherboard rather than a bespoke design — but declined to name a single OEM partner, a shipping date more specific than "next year," or a price. The Register's analysis suggested a four-accelerator configuration, drawing potentially 2,400 watts or more from its GPUs alone, may need a dedicated 20-amp circuit to run in a typical North American office, which could explain why AMD limited the demonstration unit to two cards.

AMD said it expects to ship the Threadripper Halo Station starting next year, positioning it for delivery through the same system-integrator and OEM channel — companies such as Asus, HP and Lenovo — that AMD has used to bring its smaller Ryzen AI Halo box to market. Until a partner confirms pricing and a firm ship date, the machine remains, by AMD's own description, a concept demonstrating what the company's engineers can build rather than a product on any near-term order sheet. Whether that gap closes before Nvidia counters with an updated DGX Station, and whether the ongoing memory shortage keeps a five- or six-figure workstation out of reach for the individual researchers AMD says it is targeting, are likely to shape how the category develops over the next year.