Nvidia Is Its Own First Customer. Nobody Has Figured Out Who Pays When the Agent Gets It Wrong.

Nvidia is using Cadence's AI agent to verify its own chips. Nobody has answered who is liable when it gets it wrong.

The Cadence ChipStack AI Super Agent is now operational inside Nvidia's own verification workflow — Nvidia is the first customer using ChipStack to autonomously verify its chip designs, reducing verification cycles by more than 40 times compared with manual processes, according to Nvidia's press release and CIO reporting. That 40-times figure comes from Cadence's own reporting to CIO.

Chip verification is the process engineers use to confirm a chip design will work before it gets manufactured — the step where a billion-dollar mistake becomes either a product or a disaster. Cadence built ChipStack to automate that work. Nvidia is now running it on real silicon.

What the press release does not foreground is the full stack Nvidia built around the deal. The Nemotron 3 Ultra model — a 550-billion-parameter mixture-of-experts system — is the reasoning engine powering the agent, according to the Nvidia press release. The NemoClaw toolkit manages how the agent is configured and monitored. The Vera Rubin GPU — delivering 3.3 times the FP8 throughput of the Blackwell B200 — runs the workload. Nvidia designed every layer of this system and is now deploying it on its own silicon.

Nvidia is the seller of the infrastructure, the architect of the agent's reasoning, and now the first customer deploying it at scale on real silicon. The Cadence ChipStack deal is the clearest example yet of the circularity that AI agent infrastructure is heading toward: the company building the agent OS is also the company most aggressively putting it to work.

The question nobody has answered is what happens when it fails.

Chip design errors that slip through verification are not abstract software bugs. A missed short circuit or an incorrect memory interface can cost a year of development time and hundreds of millions of dollars in respins. If an AI agent running on Nvidia's stack approves a flawed design, the liability chain is not clear. Cadence sold the agent. Nvidia built the infrastructure. The customer followed the recommended deployment configuration. Who is responsible?

The enterprise software vendors Nvidia is now asking to build on its agent infrastructure — ServiceNow, Salesforce, SAP, and others — are watching this question carefully. Each of those companies has spent years building agent governance layers on top of their own platforms, assuming they would control the interface between AI agents and enterprise workflows. Nvidia's move is to go below that interface: a secure runtime, called OpenShell, that enforces access policies at the operating system level, sandwiched between enterprise software and the agents running on top of it.

"Most of the runtime controls were at the agent process level," said Yugal Joshi, a partner at Everest Group who covers enterprise AI. "Nvidia is going a level below, making it more embedded and harder to escape." Joshi's framing, reported by CIO, describes a structural repositioning: whoever controls the secure runtime controls what the agent can see, do, and reach.

That is the same position Cadence's ChipStack agent now occupies inside Nvidia's own verification workflow — and it is the position Nvidia is asking every enterprise ISV to accept when they adopt NemoClaw and OpenShell as the infrastructure beneath their agent products.

OpenShell is open source, already integrated into Windows, Ubuntu, and Red Hat, and installs in a single command. It is not vaporware. The cross-platform adoption is real and in progress. But adoption of the standard and acceptance of the liability gap are different things, and enterprise procurement teams have not yet received a legal framework that answers the question Nvidia's own Cadence deployment raises by example.

The inference cost claims Nvidia makes for its agent stack — up to five times faster and 30 percent lower cost versus open frontier models for Nemotron 3 Ultra — are Nvidia's own benchmarks, according to the Nvidia press release. The 40-times verification cycle reduction for ChipStack is Cadence's figure, published in Nvidia's press release. Neither has been independently audited to a published methodology.

What to watch next is whether enterprise IT procurement teams start requiring liability clarity as a condition of agent infrastructure adoption, or whether the convenience of Nvidia's stack wins before the legal question gets answered. Nvidia's own use of ChipStack on real silicon is either the best proof of concept or the most motivated beta test in the industry, depending on how you read it.