Threadripper Halo Station has no price, but the parts add up to 150,000 dollars
Key takeaways
- Threadripper Halo Station pairs a 96-core Threadripper Pro 9995WX with dual liquid-cooled Instinct MI350P accelerators and a path to four.
- Memory is the largest single cost, with 2TB of DDR5 running near 50,000 dollars at current prices.
- AMD gave no price and no retail timeline, so this is a prototype statement rather than something you can order.
AMD showed Threadripper Halo Station at IFA 2026 and called it the most powerful workstation in the world. It did not give a price.
The specification explains the omission. A 96-core Threadripper Pro 9995WX, two liquid-cooled Instinct MI350P accelerators with a path to four, 2TB of DDR5 system memory, and 288GB of HBM3E scaling to 576GB. AMD says it can run trillion-parameter models locally, which is the entire pitch.
Working out the Threadripper Halo Station price from parts
The MI350P accelerators land somewhere near 20,000 dollars each. The 2TB of DDR5 runs roughly 50,000 dollars at current memory prices. Core components alone clear 100,000 dollars, and a configured system with storage, power delivery and cooling gets close to 150,000.
Note which line is biggest. The memory costs more than either accelerator, and around a third of the estimated bill is RAM rather than compute. That is not how workstation pricing used to read.
The memory shortage is the actual story
DDR5 pricing has been climbing all year, and the case for buying RAM now rather than waiting has only sharpened since. Halo Station is what that trend looks like at the top of the market: a machine whose whole reason to exist is keeping a model on your own desk, priced mostly by the memory needed to hold it.
Supply is moving, slowly. CXMT reaching G5 DRAM mass production adds capacity that did not exist eighteen months ago, though not at the HBM3E end where these accelerators live.
What you can actually afford
Halo Station is a prototype with no retail timeline, so the practical question is what the same idea costs one tier down. A workstation board, a consumer Threadripper and a 128GB DDR5 kit will run quantized models in the tens of billions of parameters without leaving the room, which covers most of what people actually do locally.
The wider question is where inference ends up running at all. The split between GPUs, TPUs, Trainium and custom silicon is being settled in data centres, not on desks, and Halo Station is AMD arguing that some of it should come back.
The part worth sitting with is the price of control. Putting a frontier-scale model under your own roof is now a legible number, and most of that number is memory.
Some links in this article are affiliate links. We may earn a small commission at no extra cost to you.
