NVIDIA launches Vera CPU for agentic AI, claims 1.8x sustained per-core performance over x86

NVIDIA released Vera, a data-center CPU designed specifically for agentic AI workloads where tool-calling, code execution, and data-processing steps run sequentially between model calls. The chip features NVIDIA's custom Olympus cores, which deliver 50% higher instructions-per-cycle than NVIDIA's prior Grace CPU, paired with 1.2 TB/s LPDDR5X memory bandwidth and 3.4 TB/s core-to-core bandwidth via a monolithic compute die. NVIDIA claims Vera delivers 1.8x sustained per-core performance over x86 CPUs in loaded agent workloads and supports up to 88 cores.

Early customer testing reinforces the pitch: Perplexity, running real coding workflows (repo cloning and test suite execution), saw ~1.5x faster completion time and 1.9x faster concurrent sandbox startup versus x86. Vera is positioned to reduce idle GPU time in AI factories—every nanosecond a CPU spends on tool execution is compute capacity the operator could monetize. The chip also enables 3x faster SQL analytics and up to 6x lower latency on retrieval workloads when paired with systems like Starburst.

For AI infrastructure builders, Vera is the clearest signal yet that CPU design is fragmenting by workload. While GPU throughput dominates headlines, agents live in the latency of the agent loop: single-threaded CPU performance per step now directly maps to revenue per GPU. Perplexity is targeting Vera in production; watch for adoption among other agentic inference platforms.