IMPLICATOR
AMD showed a liquid-cooled Threadripper Halo Station prototype built to run very large AI models locally at its opening IFA keynote in Berlin on Sept. 4, 2026. The processor, memory and two accelerators in the demonstrated configuration would cost more than $100,000 at street prices that day. The tower puts AMD opposite Nvidia's DGX Station, listed by partners for order at a stated price, while AMD offered no price or order date.
The component estimate starts with a Ryzen Threadripper PRO 9995WX that sold for roughly $11,000 to $12,000 as of the keynote. Each of the two Instinct MI350P accelerators was estimated at about $20,000 on Sept. 4 because AMD did not sell them through consumer channels. Filling the host to its supported maximum of 2 TB of DDR5 cost about $50,000 that day. AMD published no price for the machine; these figures are street-price estimates for separately purchased parts, not a quoted system price. Storage, power equipment and cooling could take a finished configuration past $150,000 at early-September prices.
What Changed
- AMD showed the Threadripper Halo Station, a liquid-cooled prototype workstation, at its opening IFA keynote in Berlin on Sept. 4, 2026, pairing a 96-core Ryzen Threadripper PRO 9995WX with two Instinct MI350P accelerators.
- The processor, 2 TB of DDR5 and the two accelerators would have cost more than $100,000 at street prices that day, and storage, power equipment and cooling could take a finished configuration past $150,000. AMD published no price for the machine.
- Each MI350P carries 144 GB of HBM3E at 4 TB per second, giving the demonstrated system 288 GB of accelerator memory. AMD said there was 'a path to four' cards, which would reach 576 GB, but the tower shown in Berlin had room for only two.
- AMD has given a 2027 availability window without an exact date, named no OEM partner, and disclosed no software stack for the box. Its own site calls the machine a prototype.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
The Sept. 4 machine paired the 96-core, 192-thread Zen 5 processor with two MI350P cards. Each accelerator carried 144 GB of HBM3E memory with peak bandwidth of 4 TB per second, giving the demo 288 GB of accelerator memory. The processor supported eight DDR5 channels and 128 PCIe Gen5 lanes. AMD said there was "a path to four" accelerators, which would lift HBM3E capacity to 576 GB.
FREE WEEKDAY MORNING BRIEFING
Follow the deskside AI hardware race.
The Implicator Morning Briefing filters the AI news cycle to the stories worth your attention and explains their consequences. From San Francisco, every weekday at 4:45 a.m. Pacific, 7:45 a.m. Eastern.
About five minutes. No hype. No spam.
The tower shown in Berlin did not have room for four cards. It appeared to use an ASUS Pro WS WRX90E-SAGE SE motherboard, plus an unidentified SFP networking card, and was assembled from commercial off-the-shelf AMD parts. Each processor had a separate closed-loop radiator instead of sharing a custom cooling loop. The MI350P cards were the first liquid-cooled versions seen after the cards launched with air cooling for standard PCIe servers.
Nvidia's DGX Station offered a different memory design as of the IFA presentation. Its 72-core Grace CPU and Blackwell Ultra GB300 superchip combined 252 GB of HBM3e with a 748 GB coherent CPU-GPU pool connected over NVLink-C2C at 900 GB per second. AMD split as much as 2 TB of host memory from the accelerators' HBM and connected the cards over PCIe Gen5 x16. The DGX Station was already listed near $100,000 and supported models with as many as one trillion parameters.
That last figure is also AMD's central claim. "This is the most powerful workstation in the world," computing and graphics chief Jack Huynh said. "Designed and engineered for a complete new era of computing, capable of running AI models that are a trillion parameters." The trillion-parameter claim rests on AMD's own statement, and no independent benchmark of it has been published.
Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.
Real-world tests cited on Sept. 4 covered AMD's smaller Ryzen AI Halo box, not the Threadripper tower. AMD said the smaller machine's 128 GB unified-memory configuration could handle models with as many as 200 billion parameters. Tests put the practical ceiling nearer 70 billion parameters for uncompressed models while preserving enough KV cache for operation. The tested ceiling was less than half the advertised one.
AMD's site calls the Threadripper Halo Station a prototype. AMD has given a 2027 availability window without an exact date, named no OEM partner or price, and disclosed no software stack for the box. "This is about as close as you can get to a personal supercomputer," Huynh said.
Frequently Asked Questions
What is the AMD Threadripper Halo Station?
A liquid-cooled deskside workstation built to run very large AI models locally, shown at AMD's opening IFA keynote in Berlin on Sept. 4, 2026. AMD's own site calls it a prototype.
How much does it cost?
AMD published no price. The processor, 2 TB of DDR5 and the two accelerators in the demonstrated configuration would have cost more than $100,000 at street prices on Sept. 4, 2026, and storage, power equipment and cooling could take a finished configuration past $150,000. Those are street-price estimates for separately purchased parts, not a quoted system price.
What hardware is inside it?
A 96-core, 192-thread Ryzen Threadripper PRO 9995WX supporting eight DDR5 channels and 128 PCIe Gen5 lanes, paired with two Instinct MI350P accelerators carrying 144 GB of HBM3E each at 4 TB per second, for 288 GB of accelerator memory.
How does it compare with Nvidia's DGX Station?
Nvidia's DGX Station pairs a 72-core Grace CPU with a Blackwell Ultra GB300 superchip, combining 252 GB of HBM3e with a 748 GB coherent CPU-GPU pool connected over NVLink-C2C at 900 GB per second. AMD splits as much as 2 TB of host memory from the accelerators' HBM and connects the cards over PCIe Gen5 x16. The DGX Station was already listed near $100,000 and supported models with as many as one trillion parameters.
When can buyers get one?
AMD has given a 2027 availability window without an exact date. It has named no OEM partner and disclosed no software stack for the machine.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
