RTX Spark, Vera, data centres and more

• Nvidia announces RTX Spark for Windows PCs at GTC Taipei.
• Nvidia also details Vera, Vera Rubin, DSX, and DOCA updates for AI data centres.

Nvidia used GTC Taipei at Computex to outline new hardware and infrastructure plans for Windows PCs and AI data centres.

The announcements included RTX Spark, a new processor for Windows laptops and small desktops. Nvidia also highlighted Vera, a data centre CPU designed for AI agents, reinforcement learning, and data processing.

RTX Spark is designed to run AI models directly on Windows PCs, with target use cases including creative applications and gaming, according to Nvidia. Nvidia announced the chip during CEO Jensen Huang’s keynote in Taiwan. The company has also referred to RTX Spark as N1X.

RTX Spark combines Nvidia’s Blackwell GPU technology with an Arm-based CPU design. It is being developed with MediaTek as a system-on-chip for Windows PCs. The chip includes a Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores. The GPU is connected to a 20-core Grace CPU through Nvidia’s NVLink-C2C chip-to-chip interconnect.

The first systems are scheduled to arrive in fall 2026. Microsoft, Dell, HP, ASUS, Lenovo, and MSI are among the hardware partners named by Nvidia.

Intel and AMD currently dominate the Windows laptop processor market. Qualcomm also sells Arm-based Snapdragon chips for Windows laptops, while Apple has used its own Arm-based processors in Macs since 2020. Shares of Intel, AMD, and Qualcomm fell after the announcement, while Nvidia shares rose.

RTX Spark will target laptops first. Early systems will support up to 128GB of memory, according to Nvidia.

The company has not announced pricing for RTX Spark laptops. It said lower-powered versions with less memory will be available for lower-priced notebooks.

RTX Spark notebooks are expected to be about 14 millimetres thick and include HD webcams. The systems are expected to come in 14- to 16-inch sizes and weigh around three pounds. The laptops will use aluminium chassis and include OLED displays with G-SYNC support, according to Nvidia.

RTX Spark focuses on local AI

Nvidia is positioning RTX Spark for local AI workloads. These include AI models and agentic applications running directly on the device instead of relying only on cloud infrastructure.

The chip supports up to 1 petaflop of AI compute, according to Nvidia. It also supports up to 128GB of unified memory.

Unified memory allows the CPU and GPU to access the same memory pool. This can reduce data movement between separate memory systems and help larger AI models run locally.

RTX Spark laptops will be suited for AI agents, content creation, and games, according to Nvidia.

Nvidia and Microsoft are also developing software controls for AI agents running on Windows PCs. The companies said the platform will use new Windows security primitives and Nvidia’s OpenShell runtime.

Those controls are intended to manage identity, containment and user permissions. OpenShell is also designed to route queries to local models based on privacy settings.

OpenShell can mask personal information in queries sent to cloud models, according to Nvidia. The work is part of Nvidia’s Microsoft collaboration on Windows agent controls.

Creative and gaming support

RTX Spark will support Nvidia graphics technologies like DLSS.

Adobe is working with Nvidia to optimise Premiere, Photoshop, and Substance 3D for RTX Spark. Updates to those applications are expected to begin rolling out with RTX Spark systems. More than 100 Windows software providers and game developers are supporting the platform, according to Nvidia. The list includes Adobe, Blender, CapCut, Riot Games, Xbox, and Remedy Entertainment.

RTX Spark can support 90GB 3D scenes and 12K 4:2:2 video editing, according to the company. It also said the chip can run 120-billion-parameter language models with a 1 million-token context. RTX Spark can also support AAA games at 1440p above 100 frames per second, according to Nvidia. The company did not provide independent benchmark results.

Nvidia is working with major game developers to support games and anti-cheat tools on RTX Spark systems.

Apple, Qualcomm, Amazon, Google, and Microsoft have adopted Arm-based chips in parts of their hardware or cloud businesses. Intel and AMD x86 processors still dominate Windows PCs.

Creative Strategies analyst Ben Bajarin estimated that Nvidia’s networking business alone would be at least 20 times larger than Nvidia’s PC chip business.

Nvidia’s data centre business remains much larger than the expected near-term PC chip opportunity. IDC estimated that 296 million PC chips shipped in 2025, up for the first time in three years. Shipments remained below the 2021 peak of 361 million.

Vera extends Nvidia’s data centre CPU push

Nvidia also used the event to highlight Vera, its data centre CPU. The chip is now in full production, and Vera-only rack servers are expected to be available this fall.

Vera is designed for agentic AI, reinforcement learning, and data processing workloads. The CPU delivers 1.8 times faster task completion than x86 CPUs, based on benchmark data cited by Nvidia. The CPU is powered by Nvidia’s custom Olympus core. It includes 88 Olympus cores, Spatial Multithreading, and an LPDDR5X memory subsystem with up to 1.2TB/s of bandwidth.

The design is intended to reduce CPU-bound delays in agentic workloads, according to Nvidia. The CPU also serves as the host processor for the Vera Rubin platform.

Vera connects to GPUs through second-generation NVLink-C2C. The interconnect provides up to 1.8TB/s of coherent bandwidth between CPU and GPU, according to Nvidia.

Vera CPUs will be available in dense, liquid-cooled racks for large-scale AI and reinforcement learning workloads. Nvidia also plans two-socket air-cooled systems for enterprise and data processing deployments.

Customers exploring or planning to deploy Vera include NYSE, Anthropic, OpenAI, ByteDance, CoreWeave, Oracle Cloud Infrastructure, Cloudflare, and Vultr, according to Nvidia. Dell Technologies, HPE, Lenovo, and Supermicro will offer Vera in standalone CPU server configurations. Other system builders in Taiwan are also working on Vera-based systems, according to Nvidia.

Nvidia adds Vera Rubin platform details

The broader Vera Rubin platform is also ramping into full production, according to Nvidia. It is designed for data centre workloads involving AI agents.

Vera Rubin combines Vera CPUs, Rubin GPUs, NVLink networking, BlueField-4 DPUs, Spectrum-6 Ethernet, and storage infrastructure. These components are packaged into rack-scale systems, according to Nvidia.

The company described Vera Rubin as a five-rack POD-scale platform. It said Vera Rubin NVL72 systems operate as part of a single AI infrastructure system. The platform delivers 10 times the agent throughput of Nvidia’s previous-generation Grace Blackwell platform, according to the company. The figure is based on Nvidia’s own comparison.

The platform also includes Spectrum-X Ethernet Photonics, Nvidia’s co-packaged optics switching technology. The switches use 200Gb/s SerDes links for larger AI clusters, according to Nvidia. The switches are intended to improve power efficiency and deployment times compared with networks using traditional transceivers, according to Nvidia.

Nvidia also introduced DSX, a platform for designing and operating AI data centre infrastructure. DSX includes reference designs, software libraries, APIs, simulation tools, and accelerated computing systems.

The platform includes DSX OS for lifecycle management and infrastructure operations. DSX Sim is used to model and validate infrastructure decisions in planning and operations.

DSX MaxLPS can allow operators to run up to 40% more GPUs at their most energy-efficient operating point, according to Nvidia. CoreWeave, Lambda, Nebius, and Nscale are among the companies deploying parts of the DSX platform stack, according to Nvidia.

Nvidia also announced new DOCA security abilities for Vera BlueField-4 STX, its storage architecture for agentic AI infrastructure. The platform uses Nvidia’s DOCA software stack and BlueField-4 silicon to enforce network and file access policies.

Vera BlueField-4 STX can support policy enforcement at speeds of up to 800Gb/s, according to Nvidia. The company also introduced DOCA components for file access permissions, workload visibility, and network isolation.

Broadcom pursues custom AI chips

Broadcom is also working with large customers on custom AI chips designed for specific workloads. Broadcom’s work includes its collaboration with Alphabet on Tensor Processing Units.

The chips are designed for AI workloads in Google’s infrastructure. Broadcom has said its custom AI chip business could generate $100 billion in revenue in 2027. The figure reflects company guidance and remains separate from Nvidia’s PC chip push. Bajarin said Nvidia’s PC chip business is expected to be much smaller than its data centre business in the near term.

Intel and AMD remain the main x86 CPU suppliers for PCs and data centres.

Want to experience the full spectrum of enterprise technology innovation? Join TechEx in Amsterdam, California, and London. Covering AI, Big Data, Cyber Security, IoT, Digital Transformation, Intelligent Automation, Edge Computing, and Data Centres, TechEx brings together global leaders to share real-world use cases and in-depth insights. Click here for more information.