Perplexity backs Nvidia’s Vera CPUs for AI agent workloads

Perplexity has said it will use Nvidia’s new Vera CPUs for AI agent coding workloads, giving Nvidia an early customer endorsement as it pushes beyond its dominant AI accelerator business and into general-purpose processors. Perplexity’s infrastructure team found Vera ran those coding tasks about 1.5 times faster than traditional CPUs, and Perplexity vice president Nate Kupp called it a “dead-on fit” for the company’s core workloads.

The decision matters because Nvidia is trying to win ground in a CPU market long controlled by Intel and Advanced Micro Devices, where servers, laptops and standard enterprise software have historically run on general-purpose chips. Vera is Nvidia’s bid to make that market relevant to AI agents, which can operate continuously and demand fast, low-latency computing that is efficient on cost per task as well as raw speed.

Nvidia says Vera is built for the CPU work behind agentic AI and reinforcement learning, including code execution, tool use, sandboxing, analytics, data pipelines and orchestration beyond the model itself. The company says the chip has 88 custom Olympus cores, 1.2TB/s of memory bandwidth and 50% faster per-core performance. Jensen Huang introduced Vera at GTC San Jose in March, and Nvidia said its first Vera systems were delivered on May 18 to Anthropic, OpenAI, Oracle Cloud Infrastructure and SpaceXAI.

The Perplexity win gives Nvidia a visible test case as it tries to widen its revenue base while AI companies increasingly build more of their own infrastructure and shift spending from pure model training toward more complex agentic tasks. Nvidia has said it expects Vera CPU sales to reach $20 billion by the end of its fiscal year, a target that depends on convincing major AI buyers that the company can supply not just GPUs for model building, but CPUs that sit inside the production systems running those models every day.

Perplexity has not disclosed how many Nvidia CPUs it plans to buy, leaving the scale of the deployment unclear. But the choice itself signals that AI companies are weighing fewer suppliers, tighter integration and more leverage over the server stack as agent workloads become a bigger part of real-world computing.