Perplexity registers for Nvidia’s Vera CPU as the semiconductor manufacturer moves beyond AI accelerators.

      Perplexity has emerged as one of the first notable AI companies to announce its plans to utilize Nvidia's new Vera CPU, a general-purpose chip that Nvidia believes could help extend its reach beyond the accelerators that have established it as the most valuable company globally. The AI search company indicated its intention to execute its agent workloads on Vera, joining an early group of users that Nvidia is eager to highlight.

      Vera is the processing component of Nvidia’s Vera Rubin platform, designed on an Arm architecture featuring numerous custom cores aimed at supporting the next generation of Nvidia's accelerators. This move marks Nvidia’s most significant foray into the market for standard server processors, a domain traditionally ruled by Intel and AMD, and it comes as Nvidia has already identified Anthropic and OpenAI as its early users.

      Perplexity is positioning the chip primarily for its speed in performing specific tasks. According to Nate Kupp, the vice president of enterprise infrastructure for the company, Vera can execute agentic coding tasks approximately 1.5 times faster than the conventional CPUs previously used by Perplexity, praising its single-threaded performance as closely suited for the job.

      The scale of this work is substantial; Perplexity currently manages over 400 million search queries each month, each processed through an inference pipeline that has relied on Nvidia hardware, including H100 GPUs, as well as Nvidia’s Triton and TensorRT software, from the outset.

      The CPU’s agentic capabilities are particularly significant in this context. In AI systems that link multiple model calls, tool utilizations, and code executions to complete tasks, the general-purpose processor coordinating these processes can become a bottleneck, similar to the accelerator carrying out intensive computations.

      Vera is a custom-designed Arm processor, featuring numerous Nvidia-developed cores, fast low-power memory, and a high-bandwidth connection to the company's accelerators. The advantage is that a chip optimized by the same manufacturer of the GPUs can facilitate data transfer between the two with less resistance than a generic option.

      For Nvidia, the endorsement from Perplexity is as valuable as the sale itself. The company has informed investors that it anticipates generating about $20 billion in revenues from Vera CPU sales in this fiscal year, representing its initial foray into a general-purpose computing sector it estimates to be around $200 billion.

      This initiative is also somewhat defensive; several of Nvidia's primary clients, including OpenAI, are in the process of developing their own AI chips. Introducing a CPU line enables Nvidia to maximize sales across server racks, even in areas where its accelerators might face competition. Additionally, Nvidia has been providing startups with compute power upfront and allowing deferred payments to secure demand.

      The ecosystem surrounding the Vera Rubin platform is rapidly developing. On the memory front, Nvidia and SK Hynix have finalized a multi-year agreement for HBM4 to support the platform, among several supply contracts that bolster the development roadmap.

      The choice made by Perplexity also reflects the direction AI search is taking. As competitors integrate generative answers into their offerings, a transition that has already transformed search behaviors, the expense associated with scaling these answers has emerged as a competitive issue.

      Cost-efficient inference is key. Faster CPUs enable companies to handle more queries without increasing spend or to delve into agentic features that link multiple calls without escalating expenses.

      For Perplexity, which has experienced rapid growth and raised significant funds to sustain this expansion, minimizing the cost per query is nearly a matter of survival. Every efficiency gained from the hardware translates to margin that doesn't need to be sourced from investors or through higher charges to users.

      Neither Nvidia nor Perplexity has provided timelines for the mass production of Vera, nor has Perplexity disclosed the extent of its transition to the new chips. A greater challenge awaits later, as Vera must earn the trust of enterprises that already have established relationships with Intel and AMD, entities they are currently more familiar with than Nvidia’s CPUs.