$3,999. That's the price AMD is asking for the Ryzen AI Halo PC, a compact desktop machine shipping with 128GB of unified memory - the single spec that determines whether you can run serious AI models locally without a cloud subscription.
Unified memory means the CPU and GPU share one large pool of RAM rather than having separate allocations. For running LLMs (large language models) locally, this matters enormously. A 70 billion parameter model like Llama 3.3 70B requires roughly 35-40GB of memory in its compressed 4-bit form, and closer to 140GB at full precision. With 128GB available, you can run 70B models in higher quality formats, or push toward 100B+ models that are simply out of reach on most consumer hardware.
What $3,999 Actually Gets You
The Ryzen AI Max+ chip (codenamed Strix Halo) at the heart of this machine has been AMD's answer to Apple's M-series unified memory architecture. Apple's M4 Max tops out at 128GB too, and Mac Studio configurations at that tier run $2,000-$2,500 - so AMD is pricing above its most direct competitor on the local LLM use case.
The premium may be justified if you're running Windows-native workflows, need AMD's ROCm GPU compute stack for specific toolchains, or simply won't run macOS. For pure local inference value, the price gap is hard to ignore.
Who This Is For
Anyone running local models for privacy reasons - processing sensitive documents, client data, or proprietary code without sending it to a third-party API - has limited hardware options. This machine joins a short list that includes Apple Silicon and high-end NVIDIA GPU setups costing significantly more once you add adequate VRAM.
Developers building RAG pipelines (where your own documents get loaded into a model's working memory) will find 128GB allows keeping much larger document sets in context during inference. Content creators running image generation or video models locally also benefit from the unified pool.
At $3,999, it's not impulse-purchase territory. But compared to a monthly bill for GPT-4o or Claude API calls at volume, the math closes faster than it looks.