NVIDIA Vera CPU ships: agent work is CPU work

NVIDIA's Vera CPU — the chip it built specifically for agent workloads — is now shipping at scale, with the first systems hand-delivered to AWS this month. NVIDIA updated its Vera delivery post on August 27 to mark it. The interesting part is not the silicon. It is the admission underneath it: when your AI stops answering and starts acting, the expensive work stops being matrix multiplication and starts being ordinary computing.

What actually happened

Per NVIDIA's launch release, Vera carries 88 custom Olympus cores, an LPDDR5X memory subsystem at up to 1.2 TB/s, and up to 1.8 TB/s of coherent bandwidth to the GPU over second-generation NVLink-C2C. NVIDIA's headline performance claim is 1.8x faster task completion than x86 CPUs on agentic workloads — and it names those workloads specifically: code compilation, Python, Java, and database processing. Systems ship from manufacturers and cloud partners starting fall 2026.

Read that workload list again. Compilation. Python. Databases. That is not model inference. That is the stuff wrapped around model inference.

NVIDIA's Ian Buck frames it as "a new CPU moment in the AI factory." The company's own argument is that every agent sandbox, every tool call, every orchestration hop, and every long-context retrieval is CPU work — and that a coding agent running a test suite spends most of its wall clock somewhere other than the GPU.

Named adopters include Anthropic, OpenAI, SpaceX AI, ByteDance, CoreWeave, Oracle Cloud Infrastructure, and NYSE, alongside providers such as Akamai, Cloudflare, Lambda, and Nebius.

Why the agent CPU shift matters for your business

You are not buying a Vera. You will rent one, indirectly, and the reason to care is that it reprices the thing you have been mismeasuring.

Most teams budget agents in tokens because tokens are what the invoice shows. But if you have ever watched a coding agent work, you know the shape: a burst of generation, then ninety seconds of npm install, test runs, file reads, and API calls. The token meter is idle for most of that. The wall clock is not, and neither is the compute you are paying for underneath the abstraction.

Three concrete moves.

Measure cost per finished task, not cost per token. An agent that costs twice as much per million tokens but finishes in half the turns is cheaper. If your only dashboard is a token counter, you cannot see that trade.

Profile where the time actually goes. Log the wall clock split between model calls and tool execution in your own workflows. We routinely find the slow part is a sandbox cold start or an unindexed query — a boring fix, invisible to anyone staring at model benchmarks.

Treat the orchestration layer as real infrastructure. Sandboxes, retrieval, queues, and tool execution are where agent workloads now live. NVIDIA just spent a chip generation making that point. If your agent stack runs tool calls through something you set up once and never instrumented, that is your bottleneck, not the model tier.

The vendor pitch is that faster silicon fixes this. Sometimes. But nobody sells you an upgrade for the tool call that takes four seconds because it hits a third-party API with no cache.

Sources: NVIDIA Blog — Delivering Vera, NVIDIA Newsroom — NVIDIA Unveils Vera, the CPU for Agents.