NVIDIA Vera CPU Delivers the Best of Both Worlds: Highest Throughput and Fastest Agents

Summary

NVIDIA's Vera CPU is focused on AI agent workloads. The company claims it can run 1.6x more agents per watt compared to the fastest x86 chip and delivers 2x faster agent performance than the most efficient x86 alternative. These claims are supported by benchmarks from partners such as Perplexity, DeepInfra, and ClickHouse.

Why it matters

Agentic AI is shifting computational demand from pure GPU inference to CPU-heavy tasks (code, search, orchestration). This makes the CPU a first-class bottleneck in AI data centers. A single chip that simultaneously maximizes both speed per agent and agents per watt could simplify data center design and reduce the need to mix fast and dense x86 SKUs. This directly impacts scaled capex and power budgets.

Highlights

• 1.6x more agents per watt compared to the fastest x86; 2x agent speed compared to the most energy-efficient x86 (100 MW data center scenario)
• Daytona launched 1,000 agent sandboxes in 10.8 seconds versus 88.4 seconds on x86; Modal reported up to 2x faster sandbox transition speed per core under heavy load
• Perplexity ran agentic coding 1.5x faster and is planning production deployment; DeepInfra achieved 2.2x the performance of the best x86 in agentic orchestration

Critical take

All figures here were published by NVIDIA or its partners. There is no independent benchmarking, cost-per-core data, or information on software ecosystem maturity. Therefore, the 'best of both worlds' claim relies entirely on vendor-selected baselines and selective workloads that highlight the chip's architecture. Perplexity only having a 'deployment plan' and some results being 'production-like tests' rather than live production data further weakens the otherwise confident tone.