On September 4, 2026, Microsoft announced a new developer PC concept called "Project Zenith." Qualifying machines feature 64GB or more of shared memory and memory bandwidth of at least 250GB/s, with availability starting on AMD Ryzen AI Halo-equipped devices. The selling point is that development tools and Windows settings come pre-configured, enabling AI models with more than 30 billion parameters to run locally.
Microsoft does not describe Project Zenith as a new Windows edition. Rather, it's an attempt to ship developer configurations—made generally available at Build 2026 in June—preinstalled on PCs equipped with high-bandwidth shared memory. While many of the settings already work on existing Windows 11 machines, questions remain about local AI speed and the safety features that isolate agents. The purchasing value becomes clear once you separate these three elements.
Drawing the Line at 64GB and 250GB/s: What Counts as a "Developer PC"
The new line Project Zenith draws is a hardware floor: at least 64GB of shared memory and memory bandwidth of at least 250GB/s. Microsoft chose AMD Ryzen AI Halo as its first supported platform, and says it will add products from other OEMs and chipmakers in the coming months. However, it has not disclosed which additional companies, model names, or sales regions will be involved.
With shared memory, the CPU and integrated GPU draw on the same physical memory pool. This makes running large AI models less constrained by dedicated GPU VRAM capacity alone. The initially supported Ryzen AI Halo platform pairs a Ryzen AI Max+ 395 with 128GB of LPDDR5X and supports both Windows and Linux. AMD also cites up to 50 TOPS of NPU performance and 60 FP16 TFLOPS of GPU performance.
Zenith's 250GB/s floor sits just below the 256GB/s that AMD cites for a 64GB Ryzen AI Max+ configuration. Microsoft has not explained this proximity as a design rationale, but it's clear the requirement goes beyond a generic PC with more memory capacity—it also demands enough bandwidth for the GPU to repeatedly read large models. Zenith is thus both a software configuration—"Windows with development tools baked in"—and a hardware criterion for selecting PCs suited to local AI.
The Preinstalled Setup Can Already Be Replicated on Existing PCs
On Zenith machines, Windows Terminal and Visual Studio Code are pinned to the taskbar. File Explorer is configured to show file extensions, hidden files, full paths in the title bar, and the details pane, with long path support enabled. Recent files and folders, sync provider hints, Start menu tips, and account notifications are disabled, while Command Palette is turned on.
Most of these settings are already available through Microsoft's published "Windows Developer Configurations." Announced as generally available at Build 2026, the Windows Developer Config on GitHub uses WinGet configuration files to turn a standard Windows 11 machine into a development workstation. It's an idempotent configuration that can be re-run on existing PCs, requiring only a single restart if WSL is enabled during setup.
The published configuration installs PowerShell 7 and Visual Studio Code, along with Git and the GitHub CLI. Runtimes are set up with .NET SDK 10, Python 3.14 with uv, and Node.js, and Coreutils for Windows is added. Oh My Posh and PowerToys are also included. It configures WSL and Ubuntu, and sets PowerShell 7 as the default profile in Windows Terminal. Developer mode and long path support are enabled, and Explorer and Start menu defaults are changed. Search and Edge settings are adjusted as well.
There's no confirmation that Project Zenith's factory image is identical to this published configuration. Even so, the boundary is clear between features that require buying a new PC to try and settings that can be applied to your current machine right now. What Zenith machines eliminate is the work of configuration—not access to the configuration options themselves.
Reading "30B+" Carefully: "It Runs" vs. "It Runs Fast"
Microsoft states that Zenith machines can run models with more than 30 billion parameters locally, without incurring cloud API usage fees. What this claim directly indicates is the capacity to hold a model and execute inference on it. Microsoft has not disclosed which specific models or quantization schemes it's referring to. Context length, tokens generated per second, and time to first response also remain unknown.
The theoretical storage capacity required for a 30B model's weights varies with numerical precision, as shown below.
| Weight precision | Theoretical capacity for 30B parameters |
|---|---|
| 4-bit | 15GB |
| 8-bit | 30GB |
| 16-bit | 60GB |
The calculation is "30 billion parameters × bits per parameter ÷ 8." In actual inference, the KV cache that holds conversation history and the runtime environment consume additional memory. Because the same shared memory must also be allocated to Windows and development tools, this table does not represent the full memory requirement. Nor does it guarantee that 64GB is enough for a 30B model to run comfortably.
AMD's own published materials likewise always pair model size with quantization conditions. For a 64GB Ryzen AI Max+ configuration, AMD gives an example of allocating up to 48GB to dedicated graphics memory to run a 4-bit Gemma 3 27B QAT model. For a 128GB configuration, it allocates up to 96GB, citing examples of up to 128B parameters at 4-bit or up to 32B at FP16. These are AMD's own reference examples, not measurements taken on actual Zenith hardware under the same conditions.
Avoiding per-token cloud API charges doesn't mean computing resources become free. There are still costs for the PC itself, electricity, obtaining and updating models, and maintaining the local environment. The value of moving from cloud to local can only be judged by measuring speed, accuracy, and power consumption with the same model and quantization settings, and comparing that against the ongoing effort required to run it continuously.
What's Ready to Use Now, and What Safety Features Are Still Pre-Production
Microsoft presents Project Zenith's developer environment, local AI, and agent safety features as a single unified experience. But these three components are not at the same stage of public readiness.
| Layer | Status confirmed as of September 5, 2026 | Remaining conditions |
|---|---|---|
| Developer configuration | Windows Developer Configurations is generally available. Can be applied to existing PCs via the public repository | Full identity with the Zenith shipping image is unconfirmed |
| Local AI hardware | Floor of 64GB+ memory and 250GB/s+ bandwidth announced. Rolling out starting with AMD Ryzen AI Halo | Official Zenith model names, pricing, regions, and real-world 30B+ model benchmarks are unpublished |
| Agent isolation | Microsoft Execution Containers (MXC) code is published as an early preview | Current policies grant excessive permissions and cannot yet be treated as a production security boundary |
While the developer configuration that shapes Zenith can be applied to standard Windows 11, the real-world performance of 30B+ models and whether MXC is ready for production use cannot be confirmed from Microsoft's announcement alone.
MXC is a mechanism for isolating untrusted processes—such as AI-generated code and plugins—using multiple methods ranging from OS-level process sandboxes to full virtual machines. Microsoft says Project Zenith machines will use OS-enforced agent identity, MXC-based isolation, and enterprise management features from day one.
However, MXC's official GitHub repository positions the current code as an early preview meant for integration and gathering feedback. The underlying sandbox is expected to change going forward, and there are known cases where policies generated by the SDK grant overly broad permissions. The README explicitly states that current MXC profiles should not be treated as a security boundary. The safety design Zenith is aiming for and the maturity of the publicly available implementation today should be evaluated separately.
Shipping Details to Verify Before Worrying About Specific Models
The shipping information needed to decide on Zenith isn't complete yet. Microsoft has not named the first supported machines by product name, nor has it disclosed pricing. When and in which countries it will be sold also remain unclear. AMD Ryzen AI Halo already exists as a separate developer-focused product, but its pricing and sales terms cannot simply be applied to Zenith machines. Which OEMs and chipmakers will be added in the coming months has also not been revealed.
There are parts you can try before buying. Applying the published Windows Developer Config to an existing PC lets you check whether the preconfigured environment suits your daily workflow. If you find you genuinely need to run 30B-class models locally on an ongoing basis, hardware meeting the 64GB/250GB/s threshold becomes a candidate. For developers who have already automated their own setup, what will matter for a purchase decision is real-world performance and the maturity of enterprise management—not the preinstalled software.
The remaining decision-making information splits into hardware and software. Pricing and sales regions need to be finalized for each supported machine, and generation speed needs to be shown with matched model names and quantization settings. For MXC, what matters is how the shipping implementation evolves from the current early preview. It will also be necessary to confirm whether the excessive permissions in current policies have been resolved and whether it can be treated as a production-ready security boundary. Once this information is disclosed, it will be possible to judge whether Zenith establishes local AI as a lasting fixture of Windows development PCs, or whether it remains just a convenient set of default settings.