Half a trillion wasn't enough. Try one.
Speaking at Nvidia's GTC developer conference in San Jose on Monday, Huang told a packed arena that the company now sees at least $1 trillion in orders for its Blackwell and next-generation Vera Rubin chips through 2027. In October, he had cited $500 billion in visibility. In January, CFO Colette Kress said that figure was "too conservative."
"I am certain computing demand will be much higher than that," Huang added on stage.
For context: no company in history has generated $1 trillion in annual revenue. Walmart, the world's largest company by sales, brought in $681 billion last year. Nvidia is projecting cumulative chip revenue that would surpass that — in about two years.
Huang's thesis for why the build-out keeps accelerating centers on a shift from training AI models to running them. Companies spent years buying chips to teach AI how to think. Now they're deploying AI to actually do work — and that requires far more compute, continuously.
"An AI that could generate became an AI that could reason, an AI that could reason became an AI that could do work," Huang said. "Inference inflection has arrived."
He also predicted software companies will stop selling subscriptions and start selling AI agents that do the work instead — what he called "AaaS," or agentic AI as a service.
Huang unveiled the first chip from Nvidia's $20 billion Groq acquisition: the Groq 3 Language Processing Unit, designed to work alongside Vera Rubin GPUs and shipping in Q3. He also previewed Kyber, the next rack architecture after Vera Rubin, expected in 2027.
Nvidia stock briefly spiked after the $1 trillion comment before pulling back. It's still down about 1.8% year-to-date in 2026 — but has beaten analyst revenue estimates for 14 straight quarters.