Microsoft spent a decade telling developers that Windows was where they should build. On Thursday in Berlin it finally attached a number to the claim: 64GB of unified memory and 250GB/s of memory bandwidth. Machines that clear that bar get a name — Project Zenith — and a version of Windows 11 that arrives already configured to run large language models on the desk instead of renting them from a data center.
Logan Iyer, Microsoft's corporate vice president for Windows platform and developer, announced the program on September 4 on the Windows Developer Blog, framing it as the shipping form of the developer-optimized Windows shown at Build in June. "Today, we're taking the next step with Project Zenith, a ready-to-code [and] distraction-free Windows experience on developer-class [PCs] with 64 GB or more of unified memory and 250 GB/s memory bandwidth or more, so developers can jump right in," Iyer wrote. The payoff he claims is specific: "On these devices, developers can run 30B+ parameter models locally and unmetered."
What ships, and on what
Zenith is not a new SKU of Windows. It is a hardware class plus an opinionated default configuration, published in the open at github.com/microsoft/WindowsDeveloperConfig and applicable to any existing Windows 11 machine through winget.
The bundle installs Visual Studio Code, GitHub Copilot, PowerToys, WinAppCLI and Windows Dev Skills, with a September 2026 language baseline of Python 3.14 or newer, uv, nvm, Node 24 or newer, WSL 2 with Ubuntu, and .NET 10. Windows Terminal and VS Code are pinned to the taskbar. File Explorer shows file extensions, hidden files, the full path in the title bar and the details pane, with long-path support enabled. PowerToys Command Palette is on; Start menu tips and account notifications are off.
The first silicon is AMD's Ryzen AI Halo, shown at IFA 2026. AMD's Jack Huynh, senior vice president and general manager of the Computing and Graphics Group, used the Berlin keynote to pitch "personal AI" as a cost and privacy argument, citing an internal estimate that monthly AI token processing climbed from roughly 0.7 quadrillion to 1.7 quadrillion in a year and could reach 120 quadrillion by 2030. "The future of AI isn't just about more compute, it's about better compute," Huynh said. He put Ryzen AI Halo at 192GB of unified memory and claimed local capacity for models up to 300 billion parameters.
Microsoft's Pavan Davuluri, executive vice president of Windows and Devices, joined Huynh on stage and confirmed Zenith will ship out of the box on Ryzen AI Halo systems. He also pitched the agent workloads Zenith machines are meant to host, gated by Microsoft Execution Containers: "With your permission, agents can work across your apps, your services, your files to help you get things done." Other silicon follows in the coming months; Nvidia's RTX Spark line is the obvious next entrant.
Pricing exists, but not from AMD. AMD showed a Ryzen AI Halo mini PC with no price attached. Lenovo's ThinkCentre X Ultra, announced September 3, starts at $3,699 and ships in November 2026 — a 1.6-liter box with a Ryzen AI Max+ PRO 495, up to 128GB of unified memory, an NPU rated around 55 TOPS, and the ability to cluster up to four units, though Lenovo has not disclosed what that clustering actually delivers in bandwidth or throughput.
The economics, and the asterisk
The unmetered pitch is real arithmetic. A $3,699 desktop amortized over three years is roughly $103 a month before power. Any team burning more than that on coding tokens has a case for moving inner-loop work — refactors, test generation, agent scaffolding, the throwaway completions nobody wants billed — onto local silicon and reserving frontier cloud models for the hard calls.
The asterisk is that "can run" and "is usable" are different sentences, and the difference is architectural. Bandwidth, not capacity, sets token speed, because a dense model must be read out of memory in full for every token generated. Community benchmarks on Ryzen AI Max+ 395 hardware with roughly 256GB/s of theoretical bandwidth put Qwen3-30B-A3B — a mixture-of-experts model with about 3 billion active parameters per token — at roughly 70 to 100 tokens per second. A dense 70-billion-parameter model at 4-bit lands near 5. ServeTheHome measured GPT-OSS 120B, also mixture-of-experts, at about 31. Microsoft's announcement carried no tokens-per-second figures and named no model behind the 30B+ claim. Zenith is a bet that the open-weights ecosystem keeps shipping sparse models. On dense 30B weights at 250GB/s, the experience is closer to reading than chatting.
The comparison Microsoft is implicitly making is with Apple, and it is not flattering. A Mac Studio with M4 Max offers up to 546GB/s; M3 Ultra reaches 819GB/s and configures to 512GB of unified memory. Nvidia's DGX Spark carries 128GB at 273GB/s — barely above Microsoft's floor — and went from $3,999 at its October 2025 launch to $4,699 in February 2026 on memory-supply pressure. Microsoft's 250GB/s threshold is a floor set roughly at the entry level of the category, not a competitive high-water mark.
Which leaves the real question: does a spec badge win back developers who left for macOS and Linux? Paul Thurrott, who installed the public configuration, called Zenith "a curious miscalculation" and argued developers each have setups they prefer and will spend hours undoing defaults. His test ended badly: "I had to wipe the PC I tried this on, it was maddening." Windows Central's Sean Endicott put the tension plainly: "What would be considered a useful preinstalled app for developers would be viewed as bloat for general users."
What to watch
Three things. Whether Microsoft or its OEMs ever publish tokens-per-second on a named model, which would convert the 30B+ claim from capacity into performance. Whether Nvidia RTX Spark systems clear the 250GB/s bar by enough to make the badge meaningful rather than nominal. And whether the second wave of Zenith machines lands under $2,500 — because at $3,699, the unmetered argument only works for developers whose cloud bills are already ugly.
“On these devices, developers can run 30B+ parameter models locally and unmetered.”— Logan Iyer, Corporate VP, Windows Platform and Developer, Microsoft