RTX Spark puts up to 128GB of unified memory into Windows laptops from six makers

Availability is the settled part. Six makers ship at launch ^[7], and NVIDIA says more systems will follow ^[2]. The specs are ceilings. The release says "up to" 128GB and "up to" 1 petaflop, and the petaflop is rated at FP4 ^[3]. NVIDIA says that is enough for developers to run frontier AI models locally ^[12]. In our view, a model served at higher precision will not see the FP4 rate.

The 128GB is one pool. The Blackwell RTX GPU and Grace CPU share it on a single superchip, with MediaTek collaborating on the custom CPU design ^[4]. NVIDIA says the shared memory lets demanding creative projects and local AI models run at the same time ^[11]. That cuts both ways for a dev machine. The model's budget is 128GB minus Windows, an editor, a browser and the context cache, so we'd expect the usable figure to be well under the headline.

Speed is the open question. Token generation for a single user is typically limited by memory bandwidth, not peak compute. The announcement does not state memory bandwidth, which configurations carry 128GB, or sustained throughput ^[9]. Those three numbers decide whether a 74B model is usable in an editor loop.

NVIDIA brings its own test case: Nemotron 3.5 Bolt 74B-A9B 24GB, a 74-billion-parameter model that NVIDIA says is tuned for calling tools and working through tasks in several steps ^[5]. It is offered for "supported RTX Spark configurations" ^[5]. That wording implies not every configuration qualifies. Guided setup covers agents including GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy, without a command line ^[6].

Local agents that write code need a permission boundary. NVIDIA points to its OpenShell secure runtime and Microsoft Execution Containers for that ^[8]. Yusuf Mehdi, an executive vice president at Microsoft, framed the pitch as hybrid: "We're building Windows to be the platform for hybrid intelligence, where AI can run locally when it makes sense and reach the cloud when needed," he said ^[10]. The condition in that sentence is "when it makes sense", and a team has to test it per workload.

We think the right move is a single-unit pilot with the team's own model, its own quantization and a realistic context length. Fleet purchases can wait for the bandwidth and sustained-throughput numbers.