Perplexity has become one of the first big AI names to say it will build on Nvidia’s new Vera CPU, the general-purpose chip Nvidia is betting can carry it beyond the accelerators that made it the world’s most valuable company. The AI search firm said it plans to run its agent workloads on Vera, joining an early group of adopters Nvidia is keen to show off.

Vera is the processor half of Nvidia’s Vera Rubin platform, an Arm-based design with dozens of custom cores built to feed the company’s next generation of accelerators.

It is also Nvidia’s clearest move yet into the market for ordinary server processors, a space long dominated by Intel and AMD, and it arrives with Nvidia having already named Anthropic and OpenAI among its first users.

Perplexity’s pitch for the chip is speed on a specific job. Nate Kupp, the company’s vice-president of enterprise infrastructure, said Vera ran agentic coding tasks about 1.5 times faster than the traditional CPUs Perplexity had been using, calling its single-threaded performance a close fit for the work.

That work is not small. Perplexity now handles more than 400 million search queries a month, each one running through an inference pipeline that has leaned on Nvidia hardware, including H100 GPUs and the company’s Triton and TensorRT software, from early on.

Agentic features are what make the CPU matter here. When an AI system chains together many model calls, tool uses and code runs to complete a task, the general-purpose processor coordinating all of it becomes a bottleneck as real as the accelerator doing the heavy maths.

Vera itself is a custom Arm design, pairing dozens of Nvidia-built cores with fast, low-power memory and a high-bandwidth link to the company’s accelerators. The pitch is that a chip tuned by the same firm making the GPUs can move data between the two with less friction than an off-the-shelf part.

For Nvidia, the endorsement counts for as much as the order. The company has told investors it expects to generate around $20bn from Vera CPU sales this fiscal year, its opening bid for a general-purpose computing market it sizes at roughly $200bn.

The push is partly defensive. Several of Nvidia’s largest customers, OpenAI among them, are designing their own AI chips, and a CPU line lets Nvidia sell more of every rack even where its accelerators face competition. It has also been offering startups compute now and payment later to lock in demand.

The wider ecosystem around Vera Rubin is filling in quickly. On the memory side, Nvidia and SK Hynix have sealed a multi-year HBM4 deal to supply the platform, one of several supply agreements underpinning the roadmap.

Perplexity’s choice also says something about where AI search is heading. As rivals fold generative answers into everything, a shift that has already reshaped how people search, the cost of running those answers at scale has become a competitive problem in itself.

Cheaper inference is the lever. Faster CPUs let a company serve more queries for the same spend, or push into agentic features that string together many calls without the bill spiralling.

For Perplexity, which has grown fast and raised heavily to fund that growth, holding down the cost per query is close to an existential question. Every efficiency it can wring out of the hardware is margin it does not have to find from investors or from charging users more.

Neither side has said when Vera ships in volume, and Perplexity has not detailed how much of its fleet will move across. The harder test comes later, when Vera has to win over enterprises that already know and trust Intel and AMD in a way they do not yet know Nvidia’s CPUs.