Leaked Windows AI PC Roadmap: 8GB to 128GB Tiers and Copilot+ Requirements

Microsoft has confirmed the broader strategy in public, and that part is solid. That strategy is what it calls hybrid intelligence. The official numbers show where the leak lines up and where it doesn't.

What the leaked roadmap reportedly says​

Wccftech describes a diagram titled "Windows is the home for hybrid intelligence". It sorts devices by memory and task: a Windows PC (8GB) for local actions, a Copilot+ PC (16GB) for quick questions, a Builder PC (32GB/64GB) for delegated and long-running tasks, and a Supercomputer (128GB+) for building and serving models locally.

The article's own prose groups the tiers a little differently from that diagram, so treat the exact boundaries as provisional. Here is the prose version:

Wccftech says many Tier 0 devices already sell for about $500. It also argues that 8 GB is a bottleneck for anything beyond light multitasking. That is the outlet's opinion, not a Microsoft statement.

What Microsoft has confirmed​

Microsoft's October 7 Windows Experience Blog post backs the overall shape of the roadmap without using the tier names.

• Strategy: Microsoft says it is building Windows as the home for hybrid intelligence. Agents run locally when that makes sense, reach the cloud when needed, and operate under security and manageability controls.
• Copilot+ PCs: Copilot will gain local context, local actions and local models on these machines, with permission. Microsoft expects rollout "in the coming months", and says timing may vary by device, market and silicon.
• Adoption figures: Microsoft says Copilot+ PCs run over 2 trillion inferences locally each month. It also says more than 40% of laptops being built for business are Copilot+ PCs.
• Builder PCs: Microsoft describes these as needing powerful silicon, unified memory and optimized AI experiences. It names RTX Spark systems from ASUS, Dell, HP, Lenovo, MSI and Surface.
• Surface Laptop Ultra: It offers up to 128 GB of unified memory and can run models above 120 billion parameters locally. Microsoft said it would be available from October 16.
• DGX Station for Windows: It uses the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip. Microsoft's footnote cites up to 748 GB of coherent memory and 20 petaflops of FP4 compute, for models up to 1 trillion parameters. Dell Pro Precision with GB300 and the HP ZGX Fury AI Station were named for later in 2026.

Microsoft's blog post does not mention Tier 0 to Tier 3, the 8/16/64/128 GB thresholds, "Wildcat Lake Refresh", "Medusa Point" or "Threadripper Halo Station". I also found nothing in my search that confirms the Crescent Island memory range as part of a Microsoft plan.

Does 16 GB mean Copilot+?​

No, and this is the most useful point for buyers. Microsoft's current requirements page says a Copilot+ PC needs:

• a processor with an NPU capable of 40+ TOPS;
• 16 GB of DDR5 or LPDDR5 RAM;
• 256 GB of SSD or UFS storage.

The page also says a PC with a 40+ TOPS NPU but less than 16 GB of RAM, or less storage, is not a Copilot+ PC. Such machines can still run compatible AI apps locally, but they miss Copilot+ exclusives like improved Windows search and Click to Do. So 16 GB is one condition among several.

Nor is 16 GB a universal minimum for AI on Windows. Even in the leak, the 8 GB tier is still called an AI PC, though a basic one. Windows 11's own baseline RAM requirement is still 4 GB.

Why the tiers make sense (my analysis)​

This part is my reading, not Microsoft's.

Memory is the practical limit for local models. Weights have to fit in RAM or unified memory. Context windows and the rest of the OS also need room. That is why the ladder is built on memory rather than NPU speed alone.

Microsoft's own examples support this. It says a 2-bit Nemotron model over 70 billion parameters needs just over 20 GB. It says MAI Code 1.1 Flash, at 137 billion total parameters, was shrunk by nearly 80% using 3-bit precision. The dev-box and supercomputer scenarios in the post lean on 128 GB-class and larger pools.

Microsoft's performance claims need care. It reports up to 2.1x faster time to first token against a MacBook Pro 16-inch with M5 Pro and 64 GB. Its footnotes say this came from Microsoft-commissioned September 2026 tests on preproduction hardware, using llama.cpp with Qwen3.5 27B at Q4K-Medium and an 8,192-token prompt. It adds that performance varies significantly by configuration. Those are company benchmarks on one workload, not general results.

Caveats​

• Source quality: The tier chart is a leak relayed by a hardware outlet. Wccftech also calls it Microsoft's "internal roadmap" without showing the document. Chip names like "Intel Core Ultra Next" and "Snapdragon Next" are placeholders, not products.
• Copilot+ branding: Wccftech notes Microsoft has been rolling back parts of Copilot+. It also reports that RTX Spark laptops won't ship under the Copilot+ name. Microsoft's May post, though, says RTX Spark PCs "will join the Copilot+ PC category". Naming looks unsettled.
• Local versus cloud: Microsoft frames hybrid intelligence as a mix, with cloud budgets one of the reasons. Nothing in the evidence says local AI replaces the cloud.
• Price: One third-party report puts the Surface Laptop Ultra at $2,599 and the Surface RTX Spark Dev Box at $5,999. I haven't verified these against a Microsoft source, so check current pricing before you buy.

What to do with this​

• Buying for general office work: Don't pay for a higher tier because of the roadmap. If you want Copilot+ features, check for a 40+ TOPS NPU, 16 GB of RAM and 256 GB of storage.
• Buying 8 GB machines: Expect cloud-reliant AI and no Copilot+ exclusives. Treat Wccftech's bottleneck warning as plausible, not proven.
• Developers and builders: Memory capacity should drive your choice. Compare configurations against the model sizes you actually plan to run.
• IT administrators: Microsoft ties its agent controls to Microsoft Execution Containers (generally available on Windows 11), Microsoft Agent 365 and Intune. Plan around those. Microsoft says governance at scale may need additional services, and it points to Ignite in November for more detail.

The tiered ladder is a believable way to describe where Windows hardware is heading. Until Microsoft publishes it, it remains a leaked chart.