Microsoft (MSFT) wants developers to run AI models on their own PCs instead of paying for cloud tokens
Key points
- Windows 11 preconfigured for 64 GB AI PCs
- Microsoft: 30B+ models run "locally and unmetered"
- AMD's Ryzen AI Halo developer PCs get it first
- One review measured a dense 70B model at about 5 tokens a second
Microsoft (MSFT) wants to help developers cut cloud AI costs by running models on their own PCs. On Friday, September 4, it announced Project Zenith, a Windows 11 setup with developer tools and settings already configured. It targets powerful PCs with at least 64 gigabytes of memory shared by the processor and graphics chip, plus memory bandwidth of at least 250 gigabytes per second. Microsoft says these machines can run AI models with more than 30 billion parameters without per-token cloud charges. Parameters are the numerical values a model learns during training.
The first computers to get it will be AMD's (AMD) Ryzen AI Halo, the chipmaker's own line of small developer PCs. The first model, powered by the Ryzen AI Max+ 395 with 128 gigabytes of memory, has been available at Micro Center since June for $3,999, in Windows and Linux versions. Microsoft gave no date for when Zenith reaches it. The open question is speed, and independent tests of the same AMD chip show a wide gap between the models that run quickly and the ones that don't. The announcement came in a post on the Windows Developer Blog by Logan Iyer, Microsoft's corporate vice president for Windows platform and developer, timed to AMD's presentation at the IFA show in Berlin.
Advertisement
What is Project Zenith?
Microsoft calls it "a ready-to-code distraction-free Windows experience on developer-class devices," not a new operating system. Windows Terminal and Visual Studio Code are pinned to the taskbar, the Linux subsystem is installed, file extensions and hidden files are shown, and the promotional clutter is turned off. Microsoft says it "will first become available with AMD's Ryzen AI Halo, with more devices from our OEM and silicon partners available in the coming months." The post does not say whether owners of existing PCs can apply the same setup themselves.
The hardware requirements are substantial. A typical laptop has 16 or 32 gigabytes of memory and far less bandwidth. Even on a qualifying machine, the size of model it can run depends on the model's format and configuration.
Which computers qualify?
AMD's Ryzen AI Halo is the only hardware Microsoft named. AMD says the June model "can run local models up to 200 billion parameters, depending on model format and quantization," TweakTown reported when the machine went on sale. Quantization stores a model's parameters at lower numerical precision so the model takes up less memory, usually with some loss of accuracy. At IFA, AMD showed a next-generation Halo built on the Ryzen AI Max+ 495 with up to 192 gigabytes of memory, Wccftech reported, without a price or ship date.
Lenovo announced another high-memory mini PC at IFA, the ThinkCentre X Ultra, with AMD's newer Ryzen AI Max+ Pro 495 and up to 128 gigabytes of memory, shipping in November from $3,699, TechRadar reported. The 128 GB configuration meets Microsoft's memory requirement, but Microsoft's post names only the Ryzen AI Halo, and neither Microsoft nor Lenovo has said the ThinkCentre will ship with Zenith.
Nvidia sells a competing small box, the DGX Spark, and Apple's Mac Studio has offered 128 gigabytes or more of unified memory for years. Neither appears in Microsoft's post.
Advertisement
How fast do the models actually run?
How quickly a model runs depends partly on its architecture. Mixture-of-experts models are built as many smaller sub-networks, and only a few of them activate for each token the model generates. This reduces the computation required for each token, though speed still depends on the hardware and software. Dense models use all of their parameters for every token, so memory bandwidth can limit how quickly they generate text.
The clearest public numbers come from ServeTheHome's review of a Beelink mini PC with the same chip and memory as the Halo, published in October 2025. Tokens are chunks of text, sometimes shorter than a word. OpenAI's gpt-oss 120B, a mixture-of-experts model, ran at 31.41 tokens a second. Meta's Llama 3.3 70B, a dense model, ran at about 5 tokens a second in the same test. Community benchmarks compiled by DataHardware put a 30-billion-parameter mixture-of-experts model at 70 to 100 tokens a second on the same chip, and dense 7-to-13-billion-parameter models at 30 to 45.
Those measurements aren't Microsoft's, and none were run on a Project Zenith machine. Microsoft's post gives no performance figures at all. Its claim is that 30-billion-plus parameter models run locally, not that they run at any particular speed. The same chip can post different numbers with different software, power settings, and model formats. Large dense models may run too slowly for some workflows.
What does this mean for the cloud token business?
Microsoft chose the phrase "metered cloud tokens" itself, and Microsoft sells them. Many cloud AI services charge according to token usage, and a developer who runs a 30-billion-parameter model on a desk box avoids a cloud provider's per-token inference charge, though the electricity and the hardware still cost money. Whether this affects cloud revenue depends on how much developers can accomplish with models that run well locally, rather than the frontier models that top the benchmarks.
Last week, we looked at Nvidia's PAIR tool, which routes AI jobs across home computers, and last month at the shift in spending from training toward running models. Project Zenith adds a local-PC angle to that broader shift toward running AI models. AMD closed at $477.45 on Friday, up 4.7%, and Microsoft at $499.68, down 2.0%, according to Robinhood market data.
Advertisement
