At GTC in San Jose on 16 March 2026, Jensen Huang presented NVIDIA's future as a chain running from vast AI factories through agent software to robots and personal assistants. The keynote's real announcement was not one chip. It was the claim that NVIDIA should furnish nearly every floor of the building.
Jensen Huang Watch analysis
What happened
Huang told the audience he saw at least $1 trillion in NVIDIA revenue across 2025 to 2027, tying that forecast to soaring demand for accelerated computing. It was Huang's outlook, not an independently audited order book, and it supplied the commercial drumbeat for everything that followed.
The centrepiece was Vera Rubin, which NVIDIA described as seven chips and five rack-scale systems designed to operate as one agentic-AI supercomputer. The company said the chips were in full production and partner products would begin arriving in the second half of 2026. Its own release also warned that features and timings could change.
Why it matters
The hardware was only the ground floor. NVIDIA also introduced OpenShell, an open runtime intended to enforce policy, network and privacy controls around autonomous agents. Its wider Agent Toolkit packages models, search, evaluation and deployment tools for company use. The pitch moves NVIDIA closer to the rules governing what agents may actually do, not merely the processors doing the sums.
DSX pushes in the other direction, down into the physical facility. NVIDIA's reference design and simulation blueprint cover compute, networking, storage, power and cooling before an AI factory is built. Taken together, the roadmap asks customers to value one coordinated route from electricity to agent action. The obvious trade is simpler integration for a deeper dependency on one supplier.
Our read
There is real appeal in buying the engine, wiring diagram and safety clipboard from people who designed them together. There is also a point where the convenient toolbox starts invoicing for the shed. GTC 2026 made that boundary, rather than any single specification, the useful question.
What to watch
- Whether Vera Rubin partner systems arrive on the stated second-half timetable.
- Whether independent testing supports NVIDIA's efficiency and cost claims under ordinary workloads.
- Whether OpenShell's guardrails remain useful across competing models, clouds and hardware.
- Whether customers can swap individual layers without unpicking the whole coordinated stack.
Discussion spark: Does NVIDIA's coordinated route from data-centre power to agent policy solve a genuine integration problem, or make switching suppliers too costly once every layer interlocks?
Sources and evidence
- GTC 2026 Keynote (20 March 2026, 06:50 UTC)
- NVIDIA GTC 2026 Keynote Replay (17 March 2026, 10:00 UTC)
- NVIDIA GTC 2026: Live Updates on What's Next in AI (19 March 2026)
- NVIDIA Vera Rubin Opens Agentic AI Frontier (16 March 2026)
- NVIDIA Ignites the Next Industrial Revolution in Knowledge Work With Open Agent Development Platform (16 March 2026)
- NVIDIA Releases Vera Rubin DSX AI Factory Reference Design and Omniverse DSX Digital Twin Blueprint With Broad Industry Support (16 March 2026)
Independent WittyWires tracker for public updates about Jensen Huang. Not affiliated with or endorsed by Jensen Huang; this is not an official account.