Most of what an AI agent does between thoughts runs on CPUs. NVIDIA's Vera pitch is a capacity warning

Watch a coding agent work for a minute. It thinks, then it searches a repository, runs a test suite, starts a container, queries a database, reads the output and thinks again. Only the thinking happens on a GPU. Everything between those moments runs on ordinary processors.

That is the argument NVIDIA, of all companies, made this month for its Vera CPU: as agents multiply, the CPU work around the model becomes a bottleneck of its own.

The claim and the evidence

NVIDIA says data centers built for AI are limited by power, so the processor that coordinates agents should maximize work per watt while serving many agents at once. For Vera, NVIDIA claims 1.6 times more agents per watt than the fastest x86 CPU, and twice the agent speed of the most energy-efficient x86 part.

The more interesting numbers come from partners, each describing its own workload:

• Daytona, which runs sandboxes for agents, brought 1,000 sandboxes online in 10.8 seconds on Vera, against 88.4 seconds on x86.
• Modal reports up to twice the sandbox throughput per core, with the gap growing under heavy load.
• DeepInfra measured up to 2.2 times the speed of the best x86 processor on agent orchestration for a production workload.
• Perplexity saw agentic coding tasks run 1.5 times faster and plans to deploy Vera in an upcoming production system.
• ClickHouse says its standard binary scaled almost linearly to 176 cores without Vera-specific tuning, and DuckLabs reports about 1.5 times the performance of an Intel Xeon 6 baseline on a large analytical benchmark.

These are vendor and partner figures, chosen to show the product well. They are still useful, because they show which workloads the people running agents at scale consider the bottleneck: sandbox startup, orchestration and the databases agents query.

Where an agent's work actually runs

Accelerator

• Reasoning and generation
• Embeddings

What most budgets count

CPU

• Sandbox start and teardown
• Code execution and tests
• Search and retrieval
• Database queries
• Orchestration and tool calls

What agents multiply

Why this matters even if you never buy Vera

Enterprise capacity plans for AI usually start from tokens: how many requests, how many tokens each, which model. Agents break that habit. A single user request to an agent can start a sandbox, run a build, call five tools and query three systems before the model writes a word of the answer.

So the systems that fail first under agent load are often not the inference servers. They are the CI runners, the sandbox pool, the read replicas and the search cluster, all sized for human tempo. We have argued that the first systems to fail under agent adoption are the ones sized for people, and NVIDIA’s partner list reads like a list of those systems.

It also changes the cost picture. NVIDIA’s own Minerva measures AI factory yield on the accelerator side. The CPU side needs its own yield metric: completed agent tasks per server, per watt or per dollar, not CPU utilization.

What to measure before the next expansion

For teams running agents on private infrastructure or reserved cloud capacity:

1. Profile one agent task end to end. Record how long it spends waiting on the model, on tools, on sandboxes and on databases. The split is rarely what the budget assumes.
2. Load-test the sandbox pool. Startup time under concurrency, not single-sandbox speed, decides how many agents can run at once.
3. Separate agent traffic on shared databases. Agents read more and faster than people. A replica or a quota for agent queries protects the systems people still use.
4. Price per completed task. GPUs, CPUs, storage and network together, divided by tasks that finished correctly.

Whether Vera wins those comparisons in your environment is a procurement question. That the CPU belongs in the agent cost model is no longer one.