AMD announces we’re “entering the era of personal AI” at IFA 2026 – and the Halo Threadripper Station

AMD kicked off IFA 2026 with a bang today, with the General Manager of its Computing and Graphics Group Jack Huynh declaring that the world of personal AI was here. And introducing what he called a “supercomputer on your desk”, if you have tens of thousands of dollars to spare.

“We are entering the era or personal AI,” said Jack Huynh on-stage. “Personal AI understands your context, it stays close to your data, and expands what you can imagine and achieve,” he said. “It becomes part of your computing experience that travels with you.”

To be more specific, Hyunh laid out three categories that define personal AI:

• Local compute (“more intelligence running directly on your device”)
• Privacy and control (“your most personal data stays yours, and you decide what gets shared and where”)
• Personal context (“AI that understands your goals, preferences and the way you work”)

Hyunh expanded on the idea of local compute by pointing to the GPT-OSS model, which was released in August 2025. So almost exactly a year ago. This had 120 billion parameters and scored 80.1 in the GPQA benchmarks. That’s roughly graduate level.

“That’s impressive, but just months later something remarkable happened. Then came Qwen 3.5.” Released in March 2026, this included only 9 billion parameters, but outscored GPT-OSS with 81.7 in GPQA. “Better performance, 13 times fewer parameters, that is the trend thath matters,” said Hyunh.

Local AI: performance and cost

The point being that AI is getting smaller and smarter, with AMD’s latest mobile chips capable of running models locally with up to 24 billion parameters. And this means that AI which previously required an AI data centre to run due to its memory demands can now run on a PC sitting on your desk.

And there are also cost benefits. “No cloud. No token costs. Frontier cloud AI is moving onto the PC.” He compared the cost of ten million output tokens on Claude Sonnet 5 at 90 euros compared to zero for running on a Ryzen AI Max system.

“This is where it starts to feel different,” said Hyunh. “On Strix Halo, you can run ten AI agents locally running in parallel. Your laptop is no longer just a tool you operate. It’s a team that works with you.”

A data centre on your desk: AMD Halo Threadripper Station

“This is the most powerful workstation in the world.” Those were Hyunh’s not-so-humble words when introducing the AMD Halo Threadripper Station on stage at IFA. “It’s capable of running AI models with nearly a trillion parameters.”

It uses a 96-core Threadripper Pro processor, a pair of AMD Instinct MI350P accelerators and up to 2TB of system memory. “And all liquid cooled, as we want all that super power to stay super quiet,” said Hyunh. “This is about as close as you can get to a personal supercomputer.”

While AMD hasn’t yet said anything about cost, that full spec would be around the £50,000/$50,000 mark. Possibly even higher with the current cost of memory.

More from IFA 2026

• First look at Lenovo Yoga Pro 9n and what it tells us about the RTX Spark platform
• What we learned about Windows 11 K2 at IFA 2026: “Unmetered intelligence on every desk”
• First impressions of the colourful Dell 14S from IFA 2026