AMD delivered the opening keynote at IFA 2026 in Berlin under the theme "The Era of Personal AI." The headline hardware was the Threadripper Halo Station, a liquid-cooled workstation pairing a 96-core Threadripper PRO with AMD Instinct MI350P accelerators. The whole design points at a single goal: running trillion-parameter models locally, with no cloud connection.
A 96-Core CPU and up to Four Liquid-Cooled Instinct Cards
At the center of the Threadripper Halo Station sits the Zen 5 generation Threadripper PRO 9995WX, codenamed Shimada Peak. It offers 96 cores and 192 threads, a boost clock of up to 5.4 GHz, and support for up to 2 TB of DDR5 system memory. That places it at the very top of the current workstation CPU stack.
Paired with it are AMD Instinct MI350P accelerators. Each card carries 144 GB of HBM3E with 4 TB/s of memory bandwidth. The configuration shown at IFA used two PCIe MI350P cards for a combined 288 GB of accelerator memory, and users can expand the system to four cards for 576 GB.
Each MI350P is rated at up to 600 W TBP and is liquid-cooled on a per-card basis, as is the CPU. Even from the keynote alone, it is clear this is a machine that solves power and thermals by brute force rather than finesse.
Why Accelerator Memory Capacity Is the Headline Number
What AMD emphasized was capacity rather than raw compute. The 576 GB figure was reverse-engineered from the requirement to hold a trillion-parameter model entirely on the device. If the weights do not fit, something has to spill out to system memory or storage, and inference latency is decided right there. For local execution, capacity is speed.
Jack Huynh, SVP and GM of the Computing and Graphics Group, framed the system as a new class of workstation that brings supercomputer-class compute to individual developers and users. Complete system specifications, availability and pricing were all left unannounced. For now this is a statement of intent rather than a product on a shelf.
Desktops and Laptops Get the Ryzen AI Max PRO 400 Series
The other pillar of the keynote was the Ryzen AI Max PRO 400 Series. These are not workstation chassis but processors intended for commercial AI PCs, mobile workstations and small form-factor desktops.
They combine the Zen 5 architecture with RDNA 3.5 graphics and an XDNA 2 generation NPU, offering 192 GB of unified memory of which 160 GB can be allocated as VRAM. The lineup consists of three models.
| Model | Cores / Threads | Max Boost | Total Cache | Graphics | NPU |
|---|---|---|---|---|---|
| Ryzen AI Max+ PRO 495 | 16C / 32T | 5.2 GHz | 80 MB | Radeon 8065S (40 CU) | Up to 55 TOPS |
| Ryzen AI Max PRO 490 | 12C / 24T | 5.0 GHz | 76 MB | Radeon 8050S (32 CU) | Up to 50 TOPS |
| Ryzen AI Max PRO 485 | 8C / 16T | 5.0 GHz | 40 MB | Radeon 8050S (32 CU) | Up to 50 TOPS |
All three models have a configurable TDP range of 45 W to 120 W. AMD describes the series as the first x86 client processors able to run 300-billion-parameter models locally at 4-bit quantization. Availability is set for the third quarter of 2026, primarily through OEM partners including HP and Lenovo.
AMD's own compact developer platform, Ryzen AI Halo, also moves to the PRO 400 Series in its next generation. The first version pairs a Ryzen AI Max+ 395 with up to 128 GB of unified memory and can handle models of up to 200 billion parameters. Its selling point is that familiar tooling such as PyTorch, vLLM, llama.cpp, Ollama and LM Studio runs as-is, optimized for ROCm.
Will Keeping Data Off the Cloud Land at a Realistic Price?
Line the announcements up and AMD's argument is consistent. Send every inference request to the cloud and costs accumulate with usage while your data leaves the building. Move the execution layer onto the desk instead.
Whether buyers agree will come down to price and power. Fill a Threadripper Halo Station with four MI350P cards and the accelerators alone draw up to 2,400 W, which collides head-on with ordinary home and office electrical limits. It is best understood as fixed business equipment. The Ryzen AI Max PRO 400 Series looks like the more practical answer, precisely because of that 192 GB of unified memory. Assembling the same capacity from discrete GPUs would change both the price and the footprint by an order of magnitude.
With no pricing announced, any cost-benefit judgment has to wait. The comparison becomes worth revisiting once HP and Lenovo systems arrive in the third quarter.
Summary
At the IFA 2026 opening keynote, AMD unveiled the Threadripper Halo Station, a liquid-cooled workstation combining a 96-core Threadripper PRO 9995WX with up to four AMD Instinct MI350P accelerators. A four-card configuration provides 576 GB of HBM3E, enough to hold a trillion-parameter model locally. AMD also detailed the Ryzen AI Max PRO 400 Series with 192 GB of unified memory, arriving from HP and Lenovo in the third quarter of 2026. Pricing and availability for the Halo Station remain unannounced.
