Local AI compute just got a name. NVIDIA and Microsoft took the stage together at GTC Taipei / Computex 2026 on June 1, with Microsoft EVP Pavan Davuluri and NVIDIA CEO Jensen Huang announcing the RTX Spark platform, silicon designed from the ground up to run agentic AI workloads on Windows 11 PCs without a cloud connection.
The specs are vendor-claimed and haven’t been independently verified. According to NVIDIA and Microsoft’s announcement, RTX Spark is claimed to deliver up to 1 petaflop of local AI compute, with up to 6,144 Blackwell RTX cores and up to 128GB of unified memory. That last qualifier matters: “up to” means the ceiling, not the floor. No independent benchmark, no Epoch AI evaluation, no third-party testing, exists yet. These are the numbers NVIDIA and Microsoft say. Treat them accordingly.
Davuluri stated that Windows 11’s workload profile scheduler has been optimized for the RTX Spark architecture to manage local background agent execution. That’s the integration point that matters for enterprise teams: this isn’t a GPU dropped into a PC chassis. It’s co-engineered silicon with OS-level scheduler hooks, which means agents running on RTX Spark get Windows-native resource management rather than fighting the general scheduler. Whether that translates to reliable production behavior at the workloads enterprise teams actually run is the question independent testing will eventually answer.
As reported by Windows Central, the Microsoft Surface Laptop Ultra is described as the first flagship device built on the RTX Spark platform. Dual-fan cooling is reported alongside that claim. Both details come from a single T3 trade press source, no primary product page was available for verification at time of publication. Flag them as reported, not confirmed.
The catch is that pricing isn’t available yet. NVIDIA and Microsoft haven’t disclosed what RTX Spark hardware will cost at retail, which means enterprise procurement teams can’t model deployment economics. That gap matters because the business case for on-device agent execution, versus continuing to route workloads through cloud inference APIs, depends entirely on whether the hardware cost amortizes against the inference spend you’re replacing. No pricing, no math.
The architecture shift here is real even if the specs need independent confirmation. Moving agent execution onto client silicon changes three things simultaneously: the security perimeter (agents run without cloud visibility), the latency profile (local inference eliminates round-trip latency for agent tool calls), and the compliance boundary (data processed locally doesn’t transit external networks). For the software layer that runs on top, the story from last week is the complement: agent software executing locally on agent hardware is a complete stack, not two separate product announcements.
What to Watch
Don’t expect independent benchmark validation this week. Computex announcements run ahead of reviewers. The RTX Spark platform’s real performance numbers, latency under agent workloads, memory bandwidth at concurrent tool-call volumes, scheduler overhead at background execution, will surface when hardware reaches reviewers and when OEM availability is confirmed beyond the Surface Laptop Ultra. Until then, the 1 petaflop figure is NVIDIA and Microsoft’s claim, and that’s how it should be reported.
Watch for: pricing announcement, OEM expansion beyond Surface, and independent benchmark publication, particularly from Epoch AI or comparable evaluation bodies. Those three data points are what converts this from a Computex announcement to an enterprise procurement decision.