NVIDIA Vera CPU: faster agents, more agents per watt

What happened?

On November 5 NVIDIA published a Vera CPU blog post on agent workloads. The claim is that AI agents keep switching between GPU reasoning and CPU execution (search, database queries, writing and testing code). AI factories are also power-constrained, so work per watt matters.

What NVIDIA claims

In a 100-megawatt data center comparison, NVIDIA says Vera beats both the fastest and the most energy-efficient x86 CPU:

• versus the fastest x86: industry-leading agent speed plus 1.6× more agents per watt,
• versus the most energy-efficient x86: 2× faster agent speed with leading energy efficiency.

The pitch is one CPU for both fast and dense agent work, instead of mixing fast and dense x86 parts.

Partner results

• Perplexity: agentic coding tasks 1.5× faster than on traditional CPUs; plans to deploy Vera in a production system.
• DeepInfra: up to 2.2× the best x86 on agentic orchestration among four architectures tested.
• ClickHouse: the standard binary, with no Vera-specific tuning, scaled near-linearly across 176 cores and topped the ClickBench hardware leaderboard.
• DuckLabs: about 1.5× the median TPC-H SF1000 of an Intel Xeon 6 baseline.
• Daytona: 1,000 agent sandboxes online in 10.8 seconds (88.4 s on x86), and 12.7 completed jobs/s at 2,000 concurrent sandboxes.
• Modal: up to 2× sandbox throughput per core versus comparable x86, with gains that grow under load. We covered Modal's own write-up separately.

Why it matters

If the CPU side of an agent run is slow, expensive GPUs sit idle. Vera's pitch is fast response and high density per watt on the same chip for AI factories. The figures are NVIDIA's and its partners' own measurements: validate on your workload before planning capacity around them.