AMD's AI desktop: what the Ryzen AI Halo costs, and what it can really do

I build software that runs AI on people's own machines, so a post about AMD's CEO showing off a tiny AI computer was always going to get my attention. This LinkedIn post by Silicon Valley's Journal shows Lisa Su with a compact system that can run models of up to 200 billion parameters without the cloud. Its argument: the next big shift in AI may happen on your desk.

The machine is the AMD Ryzen AI Halo. The post doesn't mention prices, and it describes the 128GB version, which is no longer the newest. So here are the current models, what they cost, and how far that “on your desk” promise really goes.

The video

▶ Watch the video on LinkedIn Click to load it here. LinkedIn may set cookies once it loads.

What it is

The Ryzen AI Halo is, in AMD's words, its first compact developer platform for local AI. It's about 15 × 15 cm and weighs under 1.2 kg. The trick is unified memory: the CPU, graphics and NPU share one large memory pool, so a big model can fit where a normal PC would run out of room. It works with tools developers already use, like PyTorch, vLLM, llama.cpp, Ollama, ComfyUI and LM Studio.

The prices

Machine	Memory	Price
Ryzen AI Halo (Max+ 395)	128GB	$3,999
Ryzen AI Halo (Max+ PRO 495), newest	192GB	from $7,299
For comparison: NVIDIA DGX Spark Founders Edition	128GB	$4,699

Both Halo models come with a 2TB SSD and, according to AMD and HotHardware, are sold exclusively at Micro Center, a US retailer. These are US prices without sales tax, as reported in early November 2026. The 192GB model launched in October, and it nearly doubles the price for 50% more memory.

In Europe I couldn't find an official Halo price or availability. The closest guide is other manufacturers' mini PCs with the same 192GB chip: at IFA 2026, Igor's Lab reported about €5,000–6,000 for a Minisforum and around €8,000 for an Acemagic. Those were trade-fair figures, not final list prices.

Specs side by side

	Max+ 395	Max+ PRO 495
CPU	16 cores, 32 threads	16 cores, 32 threads
Graphics	Radeon 8060S, 40 CUs	Radeon 8065S, 40 CUs
NPU	50 TOPS	55 TOPS
Memory	128GB LPDDR5X-8000	192GB LPDDR5X-8533, up to 160GB for graphics
Largest models (AMD's claim)	up to 200 billion parameters	300 billion and up, at 4-bit quantization
Operating system	Windows 11 or Linux	Linux now, Windows reported to follow

Fits in memory is not the same as fast

This is the part I'd want a friend to know before spending the money. The 200-billion figure says a model fits, not how quickly it answers. AMD's claims are also AMD's: StorageReview's review of the 128GB model reports the claim but didn't test a 200-billion model.

What it did test is telling. Serving models to many requests at once, the Halo trailed NVIDIA's DGX Spark by about 2 to 4 times in most scenarios. The Spark also has much faster networking for linking machines. The Halo won elsewhere: faster CPU work in several tests, and it runs Windows or Linux, while the Spark is Linux only. StorageReview's conclusion is roughly that the Spark suits maximum local speed, and the Halo suits people who want one full PC for both daily work and local AI.

Could it replace your AI subscriptions?

The source post asks exactly this, and its comment section pushes back hard, for good reason. My own take, as someone who ships local AI:

• Privacy is the strongest reason. Contracts, patient notes and customer data never leave the building. For European businesses that's a real advantage, as I wrote in my Mistral Large 4 post. It's also why I built LOKIDocs.
• Cost is the weaker reason. $3,999 is 20 months of a $200 subscription, and that subscription gives you the very largest models, which a desktop won't run. If you only want to save money, a cheaper plan or the API will usually win.
• No usage limits matters for agents. Long-running agents and batch jobs can run all night on hardware you own, with no meter running.
• You can start much cheaper. Try your real workload on smaller local models, or on NVIDIA's free endpoints, before buying anything.

For most people, I think the source post gets it right in the end: cloud AI isn't going anywhere, and the future looks like a split between big datacenters and small computers on our desks. The interesting news is that the small ones have become serious.

Prices and availability change quickly, so check with the retailer before buying. I couldn't open Micro Center's page directly, so the prices come from the reports linked here.

Sources

• Silicon Valley's Journal on LinkedIn: post and video
• AMD: Ryzen AI Halo and Ryzen AI Max PRO 400 Series (20 May 2026)
• HotHardware: the 192GB Ryzen AI Halo (13 October 2026)
• StorageReview: AMD Ryzen AI Halo review (6 July 2026)
• Tom's Hardware: AMD challenges the DGX Spark at $3,999
• VideoCardz: launch at $7,299
• Igor's Lab: 192GB mini PCs at IFA 2026 (7 September 2026)
• Micro Center: AMD Ryzen AI Halo