Project Zenith: Microsoft's 64 GB, 250 GB/s answer to the Mac as a local AI dev box
At IFA on September 4 Microsoft announced Project Zenith: a Windows 11 configuration for what it calls developer-class machines, defined by at least 64 GB of unified memory and at least 250 GB/s of memory bandwidth, with the explicit goal of running 30B-plus models locally without metered cloud tokens. The first machines are built on AMD Ryzen AI Halo; Lenovo ships one in November 2026 from $3,699. The box comes with a preconfigured toolchain: Terminal and VS Code pinned to the taskbar, Copilot, PowerToys and WinAppCLI. The Windows Developer Blog and Engadget had the announcement, and the target is unmistakably the Mac as a local AI dev box.
The spec that matters is the second one
64 GB decides what fits. 250 GB/s decides how fast it runs, and it is the number Microsoft would rather you did not stare at. Decode is memory-bandwidth-bound: each generated token pulls the active weights through memory once, so tokens per second is roughly bandwidth divided by bytes per token. A 30B model at 4-bit is about 17 GB of weights, which on 250 GB/s gives a ceiling near 15 tokens per second, and real figures land below the ceiling. Apple's M-series Max and Ultra parts sit at roughly 400 to over 1,000 GB/s, so the same model on the same memory footprint runs two to four times faster on the Mac before either vendor's software stack gets a vote.
This is the trade-off I laid out in unified memory Mac sizing: capacity and bandwidth are separate purchases. Zenith buys capacity generously and bandwidth minimally. That is a deliberate product choice, and it defines which tier the machine is for.
Zenith is not a Mac killer; it is a bet that a good-enough 30B model at 15 tokens per second is what most developers actually need, and the bet might be right.
What good enough looks like
For a coding assistant, an agent that runs tests, a local RAG over a repository, 15 tokens per second on a 30B model is usable. It is not pleasant for long generations and it is not what you want behind a team server, but for one developer it beats a cloud round trip on time to first token and beats it on privacy by definition. That is the customer Microsoft is going after: the developer whose company will not pay metered tokens or will not allow code to leave the laptop. TOPS do not enter into it; the NPU accelerates prefill and small models, and Microsoft is smart enough not to lead with TOPS this time.
Tokens per second per dollar
$3,699 buys 64 GB at 250 GB/s. A Mac with 64 GB at Max-class bandwidth costs more and decodes two to four times faster. A discrete-GPU box, per the Nvidia vs AMD local inference comparison, wins bandwidth by a wide margin but loses capacity: a 30B model has to fit in 24 or 32 GB of VRAM, which means aggressive quantisation or offload. Three different points on the same curve. Price per token per second is the metric to compute, and it depends on which model size you will actually run every day, not the largest one you could load.
- Decide the model size first. If it is 30B and you rarely go bigger, Zenith is priced for you.
- If you want 70B locally, 64 GB at 250 GB/s is a 5 tokens per second experience; buy bandwidth instead.
- Count KV cache: at long context on a 30B model the cache takes a real slice of that 64 GB. The KV cache math applies unchanged on Windows.
- Wait for third-party decode numbers on Ryzen AI Halo; 250 GB/s is a floor in the spec, not a measurement.
The honest limitation
The preconfigured toolchain is a real advantage I have not priced: a named configuration IT can approve is worth something in procurement, and it has nothing to do with bandwidth. Beyond that, no Zenith machine exists to benchmark until November, and Microsoft has published a minimum spec, not a throughput. Ryzen AI Halo's inference software on Windows is also younger than MLX on macOS, which may cost more than the bandwidth gap does in the first months. The idea, a standard unified-memory dev box for local models, is right. Whether the first hardware delivers it is a measurement I cannot make yet.