NVIDIA-Verified Agent Skills bring capability governance to AI agents

Verified Agent Skills is NVIDIA's first serious move into the agent governance layer, sitting alongside MCP as a portable capability spec. The idea: an agent skill (e.g. "query Postgres," "call internal CRM," "summarize PDF") gets a verified manifest covering inputs, outputs, side effects, and safety attestations, so platforms can grant or revoke skills uniformly across open-model agents.

Mechanically, the program is positioned as complementary to Anthropic's MCP rather than competitive — MCP defines how tools connect, Verified Skills define what they're allowed to do. NVIDIA's incentive is straightforward: it wants enterprise agent deployments standardized on its accelerator stack, and governance is the missing piece beyond raw inference performance.

The broader Nvidia week tested its inference-era story. The April-quarter earnings are expected to show ~79% revenue growth and $42.97B adjusted profit, but analysts question whether the ecosystem lead holds as workloads shift from training to inference where AMD, custom silicon, and TPUs all compete more credibly. Dell unveiled its AI Factory with NVIDIA plus Deskside Agentic AI workstations, citing a 320x rise in token consumption. Challenger Zyphra is raising $500M to train models entirely on AMD hardware — an explicit ecosystem bet against Nvidia.

Also this week: NVIDIA released Kimi-K2.6-NVFP4 on Hugging Face, a pre-quantized version of Moonshot AI's Kimi-K2.6 (1T total / 32B activated parameters) optimized for NVIDIA GPU inference. The pattern — Nvidia shipping ready-to-deploy quantized versions of frontier open-weight models — is a quiet but important way it stays sticky as the inference market opens up. Skeptical takes: agent governance frameworks proliferate (OpenAI, Google's Antigravity, Anthropic's Claude Code) and Verified Skills' adoption depends on whether the major model labs back it.