Project Zenith: run 30B models off your token bill

Local inference got a first-party Windows story on September 4. Microsoft announced Project Zenith, a developer-configured Windows 11 experience aimed at a hardware class that can hold a real model in memory — and the headline number is not a benchmark. It is the word unmetered.

What actually happened

Per the Windows Developer Blog, Zenith targets machines with 64 GB or more of unified memory and 250+ GB/s of memory bandwidth. On that profile, Microsoft says developers can "run 30B+ parameter models locally and unmetered," which cuts cloud token consumption during experimentation.

The rest is housekeeping, and it is the good kind. The image ships with languages, runtimes, source control, and productivity tools preinstalled. Windows Terminal and VS Code are pinned by default. WSL is integrated with container support. File Explorer, Search, Start, and the Taskbar are preconfigured to strip out the consumer clutter that every developer removes by hand on day one anyway.

Zenith arrives first on AMD's Ryzen AI Halo, with more OEM and silicon partners "in the coming months." Microsoft has not published a ship date or a price. TechRepublic frames it as a follow-on to the developer work Microsoft showed at Build 2026.

Why local inference matters for your business

Most teams treat their AI bill as one number. It is two. There is production inference — customer-facing, latency-sensitive, worth paying for. And there is development inference: the agent re-reading your repo for the ninth time, the prompt you tuned twelve ways, the test suite that calls a model on every run. The second bucket is where budgets quietly bleed, because nobody instruments it.

A 30B model on a desk does not replace a frontier model. It does replace the boring calls. Codegen scaffolding, commit summaries, test fixtures, log triage, first-pass classification — that work does not need the best model on earth, and it is the work that runs thousands of times a week.

So do the split before you buy anything. Tag every API call in your dev environment separately from production for two weeks and look at the ratio. If development is a meaningful slice of your spend, a one-time hardware cost with no per-token meter starts to pencil out fast — and it comes with a second benefit worth more than the money in regulated work: the code and data never leave the machine.

The trap is treating this as an either/or. Build the routing layer first, then decide what runs where. A hardcoded local model is the same lock-in as a hardcoded vendor.

Sources: Windows Developer Blog, TechRepublic.