Perplexity Moves Agent Sandboxes to NVIDIA Vera

The search company is shifting Perplexity Computer code execution to custom Olympus-core processors, pointing to single-threaded CPU delays as the main drag on agent loops.

Perplexity has started moving the code execution environments behind Perplexity Computer onto NVIDIA Vera processors. CEO Aravind Srinivas announced the migration on X, arguing that controlling sandboxes directly and pairing them with newer CPU silicon allows the company to vertically integrate its agent systems. Srinivas called the architecture “far better than x86.” The change targets a practical bottleneck in autonomous agents: the heavy reliance on single-threaded CPU performance while expensive GPU clusters wait for code outputs.

The Sequential Bottleneck in Agent Loops

Standard search interfaces spend most of their server time running neural network inference on GPUs. Autonomous systems change that dynamic. When Perplexity Computer tackles complex research or coding requests, it does not just produce tokens. It spins up isolated operating system environments, runs Python scripts, manipulates files, tests terminal commands, and evaluates the returned data before calling the language model again.

Those steps run sequentially on host CPUs. When the central processor takes too long to spin up a sandbox or parse output, the entire reasoning loop stalls. Expensive GPU clusters sit idle waiting for the next prompt. Srinivas had already pointed out this friction in earlier technical discussions, arguing that sequential execution latency has become the primary constraint on agent speed.

MicroVMs on NVIDIA Vera Silicon

The infrastructure in question supports SPACE, Perplexity’s proprietary isolation layer built to run untrusted code inside Firecracker microVMs. SPACE separates permissions across three tiers, holds sensitive credentials entirely outside the execution runtime, and captures real-time state snapshots to support session forking.

To speed up these microVMs, Perplexity partnered with NVIDIA to adopt the Vera CPU. Built for agent loops and reinforcement learning workloads, the chip carries 88 Olympus cores based on the Arm v9.2-A instruction set, 176 hardware threads, 1.2 TB/s of aggregate LPDDR5X memory bandwidth, and a 3.4 TB/s coherent interconnect mesh.

According to internal benchmarks released by both companies, real Perplexity Computer workloads ran roughly 1.5 times faster on Vera than on their production x86 baseline. Concurrent sandbox boot times improved by up to 1.9 times. Because these numbers come straight from the two commercial partners and lack independent third-party verification, real-world gains across diverse production workloads still need independent confirmation.

Vertical Integration Limits and Open Questions

While Srinivas described the transition as vertical integration, Perplexity is not building proprietary chips. The startup took direct ownership of its sandbox software stack, virtualization, and execution orchestration through SPACE, but it continues buying commercial processors from NVIDIA.

Neither company has shared a rollout timeline or clarified what share of Perplexity Computer traffic already runs on Vera hardware versus legacy x86 machines. Contract terms, infrastructure expenses, and the cloud hosting providers housing the hardware also remain undisclosed.