NVIDIA brings RTX Spark Superchip with 128GB unified memory to PCs

NVIDIA's RTX Spark Superchip aims to reinvent the personal computer as a local AI agent machine. The chip combines Blackwell-generation GPUs with up to 6,144 CUDA cores and as much as 128GB of unified memory, enough to run agentic AI applications directly on Windows laptops and desktops without round-tripping to the cloud. Jensen Huang detailed the platform at his GTC Taipei 2026 keynote, timed to COMPUTEX.

The strategic bet is that as agents become always-on and token-hungry, running workloads locally cuts both latency and the runaway cloud token costs the industry is now grappling with. 128GB unified memory is the key enabler—it lets a single machine hold large models and long agent contexts resident in memory.

NVIDIA leaned into the gaming and developer community to seed adoption, launching RTX Spark in South Korea with game maker KRAFTON, NC and reigning League of Legends champions T1, celebrating across the country's PC Bangs. Competitively, RTX Spark goes head-to-head with Microsoft's new agent-tuned silicon and with Apple's unified-memory Macs, and dovetails with HP's GB300 workstation at the high end. The caveat is software maturity: local agentic AI is only as useful as the on-device model and tooling ecosystem, which is still nascent compared with cloud frontier models.