Perplexity has registered for Nvidia's Vera CPU as the chip manufacturer moves beyond AI accelerators.

      Perplexity has emerged as one of the initial major AI companies to commit to using Nvidia’s new Vera CPU, a general-purpose chip that Nvidia hopes will propel it beyond the accelerators that have established it as the world's most valuable company. The AI search company announced plans to run its agent workloads on Vera, becoming part of an early group of users that Nvidia is eager to showcase.

      Vera serves as the processing component of Nvidia’s Vera Rubin platform, an Arm-based architecture boasting numerous custom cores designed to support the company’s upcoming generation of accelerators. This marks Nvidia's most significant venture into the realm of standard server processors, a market traditionally led by Intel and AMD, and it follows the company naming Anthropic and OpenAI as its initial clients.

      Perplexity emphasizes speed for specific tasks when promoting the chip. Nate Kupp, the firm’s vice-president of enterprise infrastructure, stated that Vera executes agentic coding tasks approximately 1.5 times faster than the conventional CPUs Perplexity had been utilizing, noting that its single-threaded performance aligns well with the requirements of their work.

      This work is substantial; Perplexity currently processes over 400 million search queries monthly, with each query traversing an inference pipeline that has relied on Nvidia hardware, including H100 GPUs, as well as the company’s Triton and TensorRT software from the outset.

      The agentic capabilities are what make the CPU significant in this context. When an AI system integrates multiple model calls, tool usages, and code executions to complete a task, the general-purpose processor coordinating these actions can become a bottleneck as substantial as the accelerator handling the complex calculations.

      Vera is a custom Arm architecture that combines numerous Nvidia-engineered cores with fast, low-power memory and a high-bandwidth connection to Nvidia’s accelerators. The idea is that a chip optimized by the same company manufacturing the GPUs can transfer data between the two with less friction than a standard component.

      For Nvidia, the support carries as much weight as the order itself. The company has indicated to investors that it anticipates generating approximately $20bn from Vera CPU sales during this fiscal year, marking its initial attempt at a general-purpose computing market which it estimates at around $200bn.

      This initiative is partly defensive, as several of Nvidia's largest clients, including OpenAI, are developing their own AI chips. Having a CPU line enables Nvidia to maximize sales per rack, even in instances where its accelerators face competition. Additionally, it has been extending offers to startups for computing resources now with payment deferred to secure demand.

      The broader ecosystem surrounding Vera Rubin is rapidly taking shape. On the memory front, Nvidia and SK Hynix have secured a multi-year HBM4 agreement to supply the platform, among several supply deals supporting its development.

      Perplexity’s selection also reflects the direction of AI search. As competitors integrate generative responses into their offerings — a transformation that has already altered search behaviors — the challenge of running these responses at scale has become a competitive issue in itself.

      Reducing inference costs is essential. More efficient CPUs enable a company to handle a greater number of queries without increasing expenses or delve into agentic functionalities that link numerous calls without escalating costs significantly.

      For Perplexity, which has rapidly expanded and secured substantial funding to support its growth, keeping the cost per query down is nearly a matter of survival. Every efficiency derived from the hardware translates into margin that they do not need to source from investors or by increasing charges to users.

      Neither party has disclosed a timeline for when Vera will be available in large quantities, and Perplexity has yet to specify how much of its infrastructure will transition. The more challenging test will come later when Vera seeks to win over enterprises that already have established trust in Intel and AMD, a confidence that Nvidia’s CPUs have yet to achieve.