AI search startup Perplexity announced plans to adopt NVIDIA's newly released CPU chip, Vera. This chip is not an extension of NVIDIA's traditional GPU product line, but rather a new architecture designed from scratch for the era of AI agents, targeting the deep pain points of traditional CPUs in AI agent scenarios.

The CPUs from Intel and AMD are "legacy of the laptop era."

As autonomous programs, AI agents can continuously write code, perform tasks, and coordinate complex workflows without pausing or resting like human users. However, Intel and AMD CPUs were developed during the laptop era, with their architecture initially designed to serve intermittent human-computer interactions. When it comes to the continuous loop computations required by AI agents, they often fail to achieve ideal performance.

In its official blog, NVIDIA further explained the bottlenecks of traditional CPUs: slower execution speeds reduce the number of effective evaluations per reinforcement learning cycle, thereby extending training time and increasing service latency per user. Additionally, insufficient speed may lead to KV cache loss, increasing compute costs and creating a vicious cycle. Nate Kupp, Vice President of Infrastructure at Perplexity, stated that in actual testing, the NVIDIA Vera CPU executed AI agent coding tasks approximately 1.5 times faster than traditional CPUs, showing high compatibility with its core architecture.

Core speed increased by 1.8 times, significantly reducing training cycles

Vera is designed to directly address these pain points. Its core speed has been increased by 1.8 times, enabling up to 85% of evaluation tasks to be completed within the same time window. At the same time, faster execution speed provides richer reinforcement learning feedback, generating higher-quality training signals, significantly shortening the overall training cycle and reducing negative impacts related to compute power.

Perplexity did not disclose the exact purchase quantity, but NVIDIA previously confirmed that companies such as OpenAI, Anthropic, and Oracle also plan to adopt this CPU. As leading AI laboratories launch their own chip development plans, NVIDIA is actively expanding its hardware layout, and Vera is a key step in its transition from "GPU dominance" to a "full-stack computing provider." When AI agents become the core form of next-generation software, custom-built underlying hardware for them may redefine the competitive landscape of the entire computing market.