AMD and Microsoft launched Project Zenith on November 9 — a pre-configured Windows 11 environment built specifically for local AI development, shipping on the Ryzen AI Halo platform. The hardware supports up to 192 GB of unified memory and can run 300-billion-parameter models without a cloud dependency. It is available now from Micro Center at $3,999 (128 GB) or $7,299 (192 GB), in-store pickup only.
Not Just Hardware — A Pre-Validated Stack
The important part of Project Zenith is not the hardware. It is the software. AMD and Microsoft co-engineered a specific combination of drivers, runtimes, and developer tools that have been validated to work together on this hardware, out of the box. The AMD Ryzen AI Developer Center gives developers a central hub to stay on that validated configuration as software updates land.
What comes preinstalled: Visual Studio Code, Git, GitHub CLI, GitHub Copilot CLI, Azure CLI, Python, Node.js, .NET, PowerShell 7, and WSL. AMD also ships five AI Playbooks — guided, end-to-end workflows for running tasks on the hardware:
- Running LLMs with PyTorch and AMD ROCm
- Image generation with ComfyUI
- Local coding with VS Code and Qwen3-Coder
- Serving models through LM Studio
- Workflow automation with n8n and local LLMs
This matters because getting ROCm, PyTorch, and local model serving to work together on AMD hardware has historically required significant debugging. Project Zenith removes that obstacle — which is the entire point of the initiative.
The Hardware Numbers That Matter
The chip underneath is the Ryzen AI Max+ PRO 495. For AI workloads, the relevant figures are:
- 192 GB unified memory, with 160 GB allocatable to the GPU as VRAM
- 273 GB/s memory bandwidth
- 131 TOPS total AI platform performance (CPU + GPU + 55 TOPS NPU)
- Support for models up to 300 billion parameters, running fully locally
The 192 GB unified memory pool is the story. It is what allows a single desktop box to run a model that would otherwise require a multi-GPU cloud setup. Memory bandwidth at 273 GB/s keeps that memory accessible fast enough to be useful for inference at scale.
Does the Price Make Sense?
$3,999 for 128 GB. $7,299 for 192 GB. That is real money. The honest framing: a cloud A100 instance costs $3–4 per hour. A local workstation at this price range breaks even against cloud costs in around 2–4 months at 8 hours of daily use. After year one, running cost is electricity. If you are spending more than $150 a month on cloud GPU time, local pays for itself fast.
The main constraint is availability — Micro Center in-store pickup only at US locations. No online orders, no international shipping.
Already Own a Ryzen AI Halo Machine?
You do not need new hardware. Microsoft open-sourced the Windows Developer Config project, which reproduces the full Project Zenith setup on any existing Ryzen AI Halo system running Windows. AMD also confirmed Project Zenith is coming to other Ryzen AI Max-powered devices.
The Caveat Worth Naming
Project Zenith solves the setup problem. It does not solve the ecosystem problem. AMD ROCm delivers roughly 75–85% of NVIDIA inference performance on comparable hardware, and the developer ecosystem remains heavily skewed toward CUDA — roughly 18:1 by most proxy measures. For LLM inference via llama.cpp, vLLM, or LM Studio, ROCm is genuinely capable. For custom training pipelines, specialized kernels, or production ML beyond standard inference, CUDA still wins.
AMD’s bet is that if setup friction disappears and price-performance is right, developers will build on ROCm. Project Zenith is the clearest execution of that bet yet. Whether the ecosystem follows is a separate question that a preinstalled toolchain cannot answer on its own.
For teams that need local inference for privacy, compliance, or cost reasons and are building around standard model serving rather than custom training, the Ryzen AI Halo with Project Zenith is a serious option. The full hardware configuration details are on AMD’s blog.
