Nvidia Vera CPU Targets AI Inference Workloads

Nvidia said Vera delivers 1.6 times more agents per watt than the fastest x86 CPU in a 100-megawatt data center comparison, while also doubling agent speed versus the most energy-efficient x86 part. The CPU’s 88 Olympus cores, 176 hardware threads and 164 megabytes of unified L3 cache are designed to cut contention and keep AI agents moving between CPU and GPU tasks with less delay. For operators building “AI factories,” the appeal is straightforward: more completed jobs without a proportional increase in power draw or rack space.

The company is also backing the launch with early customer tests that suggest the chip is not just a lab demonstration. Perplexity said it saw 1.5 times faster agentic coding tasks, while DeepInfra reported up to 2.2 times faster agent orchestration than the best x86 processor. ClickHouse said it scaled nearly linearly across 176 cores without Vera-specific tuning, and Daytona said sandbox initialization time fell from 88.4 seconds on x86 to 10.8 seconds with Vera. Those kinds of gains matter because small delays compound quickly when thousands of agents are running simultaneously.

Investors will read the launch as part of Nvidia’s broader effort to extend its dominance beyond GPUs and deeper into the data-center stack. The more tightly Nvidia can couple CPUs, GPUs and interconnects, the harder it becomes for cloud operators and AI developers to mix and match components from competing vendors. Vera’s integration with Nvidia’s NVLink-C2C fabric, which the company says provides 1.8 terabytes per second of bandwidth, reinforces that strategy by making the CPU a structural part of Nvidia’s platform rather than a commodity add-on.

The timing also comes as the AI hardware trade remains volatile. Nvidia shares have been under heavy pressure in recent sessions, with the stock falling to 169.64 on Nov. 10 from 206.69 on Nov. 3, while its 14-day RSI dropped to 19.4, a level that typically indicates deeply oversold conditions. Adalytica’s proprietary gauge showed NVIDIA Earnings Sentiment at 4.0, or “Extreme Fear,” even as awareness remained at 100.0. That combination suggests the market is still grappling with headline risk and valuation compression even as the company keeps pushing product boundaries.

For rivals, the message is less flattering. Intel and AMD remain tied to the x86 base that Vera is targeting, and any widening gap in agent throughput per watt could force customers to rethink procurement around total cost of ownership rather than CPU price alone. For Nvidia’s customers, the benefit is more immediate: lower operating costs and better utilization in a sector where power has become as important as silicon.

The bigger question is whether Vera can translate benchmark and pilot wins into meaningful deployment at scale. If it does, Nvidia could strengthen its grip on the AI data-center market just as the industry starts optimizing for efficiency, not just speed. If it does not, the launch still raises the bar for competitors and underscores how central energy economics have become to the AI boom.