Jensen Huang opened the Computex 2026 keynote in Taipei on June 1 with the RTX Spark, a superchip pairing a 20-core Arm Grace CPU with a Blackwell GPU carrying 6,144 CUDA cores, all sharing 128 GB of LPDDR5X unified memory on a single TSMC 3nm package. The Verge's report confirms a fall 2026 launch across more than 30 laptops and 10 desktops from Microsoft, Dell, HP, ASUS, Lenovo, and MSI, with Microsoft's Surface Laptop Ultra anchoring the line.
The CPU was co-designed with MediaTek. Memory bandwidth tops out around 300 GB/s, on par with workstation-class accelerators of two generations ago. Nvidia is no longer asking a workload to choose between data center and laptop. It is shipping a near-data-center silicon budget on a client board.
A 20-core Arm CPU paired with a Blackwell GPU running 6,144 CUDA cores does not fit a 35 W ultrabook envelope. The Surface Laptop Ultra and the ProArt and OmniBook variants will land in the 45-to-140 W class depending on form factor, with desktops running higher still. That is a different cooling problem from the rack-density story the industry has spent two years writing about. It is the same story moved to a chassis the user holds. Vapor chambers, hybrid heat pipes, and the high end of fan-based thermal design get a sudden volume bump from a single Nvidia platform decision.
The Spark is purpose-built for on-device agentic workloads. A 128 GB unified memory pool is enough to hold the kind of model footprints that have been forcing inference back into colocated edge sites. If a meaningful share of inference moves to the client, the edge thermal upgrade cycle does not slow down. It splits. Some workloads still need the edge data center. Some collapse into the laptop or workstation. The cooling implication for vendors who serve hyperscale is that volume forecasts have a new variable they do not control.
Nvidia is committing TSMC 3nm wafer capacity to a client product at the same moment Vera Rubin and Feynman are also pulling on TSMC's leading edge. The fight for advanced-node wafers and CoWoS-style packaging just gained another large claimant. For thermal vendors, the practical read is that any cooling product whose lead time assumes a stable Nvidia hyperscale wafer share should be stress-tested against a scenario where Spark consumes a few percentage points of leading-edge supply. The wattage roadmap now has a client branch, and that branch competes with the data center branch for the same fab capacity.