Microsoft has reportedly launched a Surface Laptop Ultra with NVIDIA’s RTX Spark chip, a machine aimed at running demanding AI workloads locally rather than sending everything to the cloud. 36Kr says the system can be configured with up to 128GB of unified memory and is designed to run models with as many as 120 billion parameters.
Microsoft AI Watch analysis
What happened
36Kr reports that Microsoft worked with NVIDIA on the laptop’s platform, including Windows-level power, memory and hardware coordination. It says the RTX Spark combines a Grace CPU with a Blackwell RTX GPU, supports CUDA and offers up to one petaflop of FP4 compute.
The same report says pre-orders began at a company event in San Francisco, with shipping scheduled for 16 October. That date had already passed when the article was published on 26 October, so the launch timeline needs clarification.
Why it matters
If the reported specifications and model capacity hold up, this would put serious local-inference ambitions into a consumer laptop format. Developers could prototype and test AI workloads on-device, with less reliance on cloud instances for at least some tasks. The 128GB unified-memory ceiling is the practical headline: large models are often constrained as much by memory as by raw compute.
This is a reported product launch, not an independently established performance result. The supplied report gives headline specifications but no benchmark showing how quickly the laptop runs a 120-billion-parameter model, or under what settings.
Our read
The interesting shift is not another laptop wearing an AI badge. It is the prospect of Windows hardware designed around local model workloads from the outset, with CUDA and unified memory in the mix. But specifications are invitations, not test results; the real story will be what developers can run, at what speed and for how long before the fans begin auditioning for a speaking role.
What to watch
- Whether Microsoft or NVIDIA publishes full specifications, pricing and availability details.
- Independent benchmarks for local inference, including the model size and precision tested.
- Whether developers can use the CUDA stack without awkward compromises on a thin laptop.
Discussion spark: Would you pay a premium for a laptop that can run large models locally, or are cloud access and a lighter, cheaper machine still the better trade?
Sources and evidence
- Microsoft’s Azure Growth Looks Built For The AI Buildout – Finimize (26 October 2026, 12:02 UTC)
not affiliated with or endorsed by Microsoft