Nvidia Debuts Its First In-House CPU at GTC Taipei 2026
Nvidia formally introduced the Vera CPU at GTC Taipei 2026, presenting it as the company's first custom-designed processor engineered specifically for the agentic AI workloads that are placing growing pressure on conventional server architectures. The chip arrives as one of several major announcements from the event, signaling Nvidia's intent to move beyond GPUs and establish itself across the full computing stack. OpenAI, Anthropic, and SpaceXAI rank among the earliest organizations to take delivery of the processor, which Nvidia claims offers a 1.5x performance edge over the newest 128-core x86 chips when handling agentic tasks.
Architecture Designed Specifically for AI Agents
At the heart of the Vera CPU sit 88 custom Olympus cores built on the Armv9.2 instruction set. Rather than relying on licensed Arm microarchitecture, Nvidia designed these cores entirely in-house, a notable departure that reflects the company's ambition to control performance characteristics end to end. Each core incorporates a 10-wide instruction front-end alongside a neural branch predictor, producing what Nvidia describes as 50 percent faster per-core performance than comparable x86 designs across compilation, scripting, and compression tasks running inside agentic sandbox containers.
Memory Bandwidth and Power Efficiency
Memory bandwidth stands out as one of the chip's defining specifications. The Vera CPU delivers 1.2 TB/s through LPDDR5X across a 1,024-bit interface, allocating each core roughly 14 GB/s—approximately three times the per-core figure found in traditional data center CPUs. The processor also links to Nvidia's Rubin GPUs through NVLink-C2C, supplying 1.8 TB/s of coherent bandwidth between the two. On the efficiency front, Nvidia reports that Vera draws less than 30 watts for memory power, a sharp contrast to the more than 100 watts consumed by conventional DDR5 systems.
Early Deliveries and Adoption Already in Motion
Initial shipments of the Vera CPU are already in customers' hands. Nvidia vice president Ian Buck personally delivered the first Vera CPU systems to Anthropic, OpenAI, SpaceXAI, and Oracle Cloud Infrastructure in mid-May, according to multiple reports. CoreWeave has been confirmed as the first cloud provider to offer standalone Vera CPU access, while ByteDance is counted among additional hyperscalers preparing to adopt the chip.
During the GTC Taipei keynote, Jensen Huang noted that SpaceX is among the processor's early adopters. He also indicated that systems built around the Vera CPU from Dell, HPE, Lenovo, and Supermicro are expected to ship in the second half of 2026.
A Wider Platform Strategy Behind the Vera CPU
The Vera CPU forms one component of the broader Vera Rubin platform, which Nvidia says is now ramping into mass production and promising up to 10 times lower cost per token compared with the earlier Blackwell architecture. The same keynote introduced the RTX Spark PC chip, developed in partnership with MediaTek and manufactured by TSMC. That product marks Nvidia's entry into the Windows PC market, positioning the company in direct competition with Intel and Apple.