40,000 Employees Today and 4 Million Agents Tomorrow: Nvidia CEO Jensen Huang Believes Agentic AI Will Soon Become the Norm in the Workplace

Nvidia projects up to $20 billion of Vera CPU revenue in its current fiscal year, with potential for that figure to more than double by fiscal 2028, positioning the company as a major supplier of chips for agentic AI workloads. CEO Jensen Huang said AI usage has crossed into a predominantly agentic model and expects companies to deploy vastly larger fleets of autonomous agents. Nvidia claims Vera delivers 1.8x industry-standard performance and 5x data-center CPU bandwidth per watt, while Bank of America forecasts the CPU market could expand fivefold from $35 billion this year to $170 billion by 2030. The opportunity is tempered by cybersecurity risks as businesses grant AI agents access to files, browsing, and workflow systems.

Analysis

The investable implication is not simply incremental NVDA demand: agent deployment shifts the bottleneck from training accelerators toward inference orchestration, networking, memory and trusted execution. NVDA's ability to bundle CPU, GPU, networking and software could raise system-level content per AI cluster, but its CPU revenue ambition remains a management assertion rather than externally validated demand; the near-term risk is CPU cannibalization of incumbent server spend rather than wholly incremental data-center capex. AMD and INTC face the clearest multiple risk if hyperscalers validate a non-x86 control-plane stack, while AVGO and ANET benefit if agent workloads create materially higher east-west traffic.

The immediate stock reaction should be limited because the agentic-AI narrative is already embedded in NVDA's premium valuation. Over the next 1-3 months, catalysts are hyperscaler capex commentary, OEM adoption disclosures, benchmark evidence against AMD EPYC/INTC Xeon, and any disclosed design wins. Over 6-18 months, recurring agent workloads could materially increase inference utilization, but autonomous access also creates an underappreciated security-control cost: spending may migrate toward PANW, CRWD and ZS before broad enterprise agent deployment converts into compute orders.

Consensus likely overstates the direct substitution of CPUs for GPUs. Agents still require expensive model inference and reasoning; CPU attach is valuable only if NVDA can win the orchestration layer without customers disaggregating components. A failed or high-profile agent-security incident could delay production permissions, reducing utilization and elongating the revenue ramp despite continued model experimentation.

Key Decisions for Investors

• Maintain NVDA as a core long, but add only on a 10-15% pullback or after independently disclosed Vera design wins; target a 6-12 month catalyst path through platform attach-rate expansion. Falsifier: hyperscaler capex guidance weakens or NVDA data-center gross margin falls below guidance on CPU/networking mix.
• Initiate a 3-6 month pair: long AVGO or ANET / short INTC, sized market-neutral. Agentic production deployments should favor networking/custom silicon exposure while INTC bears architecture-share and margin risk; cover if enterprise server demand materially reaccelerates or Intel reports major hyperscaler CPU wins.
• Use PANW or CRWD as a second-order agentic-AI hedge rather than a direct AI-beta substitute. Add after evidence of enterprise agent rollout; privileged-access governance and monitoring are likely gating purchases before broad autonomous workflow adoption, with a 6-18 month horizon.
• Avoid treating BAC's cited research as a standalone catalyst; its exposure is indirect and the relevant signal is whether its AI-capex/semiconductor estimates are adopted by hyperscaler guidance. No trade absent revised estimates or sector positioning data.