NVIDIA Debuts RTX Spark and Confirms Vera CPU at Computex

NVIDIA used its Computex keynote on June 1, 2026, to unveil RTX Spark — the company’s Windows‑on‑Arm PC superchip — and to confirm that its Vera data‑center CPU is now in production.

The chip shown as RTX Spark is the long‑rumored N1X family: a Windows‑on‑Arm platform developed with partners including Microsoft and MediaTek and positioned as a new class of PC silicon for agentic AI workloads.

NVIDIA and Microsoft framed RTX Spark as a hybrid CPU‑GPU “superchip” built to run local AI agents. Public specs highlighted up to 1 petaflop of AI compute, roughly 6,144 Blackwell RTX GPU cores, up to 20 Arm CPU cores, and as much as 128 GB of unified memory for model context and caching.

The company showed how those on‑device resources aim to let agents run in the background, switching among apps and handling long token contexts without constant cloud round trips. NVIDIA also announced software optimizations — including an on‑device agent framework and faster inference on popular local runtimes — to speed multi‑token prediction.

Hardware partners have already lined up first‑wave designs. Microsoft, Dell, HP, Asus, Lenovo, and several others are listed among OEMs planning RTX Spark laptops, mini‑PCs and desktops for a fall 2026 rollout; Microsoft’s Surface team showed a prototype device during the event.

NVIDIA and Microsoft say software compatibility is central to the launch. Microsoft’s Windows Experience Blog describes runtime and emulator work to run legacy x86 apps on Arm, plus built‑in hooks so agent features can surface from the desktop taskbar and system UI. Developers will be able to target existing frameworks with updated toolchains.

Alongside the PC push, Jensen Huang confirmed Vera — NVIDIA’s in‑house data‑center CPU designed for agentic workloads — is now in production and shipping to customers. The company markets Vera as a CPU tuned for high single‑thread performance and data‑heavy AI tasks.

Vera is part of NVIDIA’s Rubin platform, which pairs Vera CPUs with Rubin GPUs in rack‑scale systems connected by high‑bandwidth NVLink interconnects. Public materials cite an 88‑core Vera SKU in Rubin racks and coherent CPU‑GPU bandwidth figures intended to accelerate agent orchestration at scale.

Taken together, the announcements underline a two‑pronged strategy: push substantial agent compute onto consumer and enterprise endpoints, while boosting data‑center CPU supply and integration for large‑scale agent orchestration. That mix lets agents run locally for privacy or latency‑sensitive tasks, and fall back to cloud resources for heavyweight model training or multi‑agent coordination.

Industry analysts at the show flagged practical constraints. Memory modules, flash storage, and power budgets remain bottlenecks for very large local models, and pricing will influence how broadly RTX Spark devices penetrate mainstream PC segments. NVIDIA’s ties to Taiwanese fabs and MediaTek are meant to ease supply and design hurdles, but competition from x86 incumbents and other Arm vendors will be fierce.

NVIDIA also used the stage to pitch its software ecosystem: model support for Nemotron‑class networks, optimized local runtimes such as llama.cpp and vLLM, and partnerships with ISVs to rebuild apps for local AI acceleration. Adobe and Blender were cited as early third‑party optimizations that demonstrate practical gains for creators and developers.

What to watch next: first RTX Spark notebooks and mini‑PCs arriving in fall 2026, early reviews of real‑world agent workflows, and Vera‑based Rubin systems ramping in data centers later in 2026. The announcements mark a clear bet by NVIDIA that agentic AI will reshape both endpoint experiences and data‑center economics.