At Computex 2026, NVIDIA CEO Jensen Huang unveiled RTX Spark, an Arm-based Windows PC superchip built with Microsoft and MediaTek, and announced Vera CPU for data centers is now in full production. RTX Spark features a Blackwell GPU with 6,144 CUDA cores connected via NVLink-C2C to a 20-core Grace CPU, designed for on-device AI agents; it debuts this fall on laptops from Microsoft Surface, Dell, HP, ASUS, Lenovo, and MSI. Concurrently, Vera CPU is entering mass production with first units delivered to Anthropic, OpenAI, SpaceXAI, and Oracle Cloud Infrastructure, with customers including NYSE, ByteDance, and CoreWeave.
Vera marks a structural shift: the CPU becomes the control plane for agentic workloads. NVIDIA claims Vera delivers 1.8x faster task completion than x86 processors and 1.8x faster token generation than prior-generation configurations, with early production reaching 22,500 concurrent environments per Vera CPU rack. The Vera Rubin platform—five purpose-built racks running as one AI supercomputer—is also in full production with supply chains now 2x larger than Grace Blackwell and rack assembly time slashed from 2 hours to 5 minutes. OCI plans to deploy hundreds of thousands of Vera units beginning in 2026.
Together, RTX Spark and Vera represent NVIDIA's full-stack pivot from GPUs alone to complete edge (RTX Spark on PCs) and hyperscale (Vera Rubin) systems. RTX Spark's PC entry threatens AMD, Intel, and Qualcomm market share; Vera's CPU leadership aims to lock in agentic inference workloads before rivals can scale alternatives. The moves signal NVIDIA's bet that owning both the edge and the orchestration layer—not just the accelerator—is the next frontier.