Computex 2026: Vera Rubin ships, and Nvidia moves into Windows PCs

Nvidia used its Computex keynote in Taipei on Monday, June 1, to make three claims: its next-generation data center platform is in full production, a new class of Windows PC chip (RTX Spark) is coming, and, in Jensen Huang's words, "our customers don't want to buy computers, they want to build AI factories" ^[1]. This post separates the announcements that are specifications from the ones that are vendor estimates, and notes what each means for a business that buys AI capacity rather than racks.

What was announced

Vera Rubin NVL72 and the five-rack platform

The Vera Rubin NVL72 connects 36 Vera CPUs and 72 Rubin GPUs in one rack. Nvidia says it delivers up to 10x higher inference performance per watt and up to 10x lower cost per token. Paired with a Groq 3 LPX inference rack, Nvidia claims up to 35x higher throughput per watt for trillion-parameter models ^[1].

Vera Rubin is positioned as a five-rack platform: the NVL72 compute rack, a Vera CPU rack, Groq 3 LPX, Spectrum-6 SPX Ethernet, and Vera BlueField-4 STX storage. Nvidia said hundreds of supply chain ecosystem partners are ramping it across more than 350 factories in 30 countries, and quoted CEO Jensen Huang: "The supply chain we created for Vera Rubin is twice as large as Grace Blackwell" ^[1].

Two engineering details are worth noting because they affect data center planning rather than software. The NVL72 is described as fully liquid-cooled, designed for 45 degrees C warm-water inlet temperatures, and compute tray assembly time drops from two hours to five minutes ^[1]. Nvidia also describes 800 VDC power architecture and MGX-compatible 800 VDC racks, a third-generation MGX rack design for Vera Rubin, and says it has contributed the MGX rack-scale design to the Open Compute Project ^[1].

The CPU, networking and storage pieces

Early adopters of the photonics Ethernet switches include CoreWeave, Lambda and Oracle Cloud Infrastructure ^[1]. The networking line matters beyond the headline: in its latest quarter Nvidia's data center networking revenue was $14.8 billion, up 199% year over year ^[2].

RTX Spark: the PC angle

RTX Spark pairs a Blackwell RTX GPU with a custom 20-core Grace CPU built with MediaTek and targets slim Windows laptops and compact desktops running Microsoft Windows. A separate DGX Station for Windows, a deskside AI system, was also announced ^[1]. Adobe said it is re-engineering Photoshop and Premiere for the chip, with 2x faster AI and graphics performance ^[1]. Nvidia's page gives no availability date or pricing for RTX Spark.

Models and agent software

Nvidia also released Nemotron 3 Ultra, a 550-billion-parameter mixture-of-experts model it says offers up to 5x faster inference and up to 30% lower cost for agentic tasks. It is available on Hugging Face, OpenRouter and build.nvidia.com. Separately, an Agent Toolkit and an OpenShell runtime provide sandboxing and centralized policy enforcement for autonomous agents ^[1].

Reading the claims

Every performance multiple above is "up to," measured by Nvidia against its own prior platform, on workloads Nvidia chose. They are useful for direction, not for budgeting. A cost-per-token improvement only reaches a customer if their cloud provider passes it on, and Nvidia's own numbers show how much of the industry is still being built on the current generation.

The demand side is consistent. In late April, Meta raised its 2026 capital expenditure range to $125-145 billion, Alphabet raised its to $180-190 billion, and Microsoft guided to about $190 billion for calendar 2026, including $25 billion tied to higher component pricing; Amazon said its plan, which in January it expected to approach $200 billion, was largely unchanged ^[3].

What to do with this

Treat these announcements as a roadmap, not a purchase signal. For on-premises or edge deployments, the DGX Station and RTX Spark lines show Nvidia expecting trillion-parameter models to run on a desk, but with pricing and availability dates not given on Nvidia's page, plans should not depend on them yet. For everyone else, ask your cloud or model provider which hardware generation your workloads run on, and revisit unit pricing when Vera Rubin capacity reaches general availability.