Jensen Huang walked into a packed San Jose hockey arena, leather jacket and all, and casually doubled NVIDIA’s AI revenue forecast to $1 trillion through 2027. Then he announced chips purpose-built for AI agents, a partnership with a former competitor, Disney robots, self-driving car deals, and — because apparently Earth isn’t big enough — data centers in space.

GTC 2026 wasn’t a product launch. It was a declaration of what comes next.

The Number That Broke Brains: $1 Trillion

Just last month, NVIDIA cited a $500 billion AI chip opportunity on its earnings call. Yesterday, Huang doubled it. The logic: training massive models was Phase 1. Inference — running those models for hundreds of millions of users in real time — is Phase 2. And inference scales with usage, not just investment.

More users. More queries. More autonomous agents doing things without human supervision. All of it means more chips.

“The inference inflection has arrived,” Huang told 18,000+ attendees. “And demand just keeps on going up.”

For the company that became the first to hit a $5 trillion valuation last October, this isn’t hubris. It’s a thesis — that AI compute is becoming a utility as fundamental as electricity. Whether you believe the number or not, the direction is undeniable.

Vera Rubin: Not a GPU. A Full-Stack Agent Machine.

The headline hardware is Vera Rubin, and calling it a chip undersells it. It’s an integrated platform — seven chips, five rack-scale configurations, and a complete supercomputer architecture — purpose-built for the age of agentic AI.

The most telling piece: a new Vera CPU designed specifically for agent workloads where traditional GPUs are overkill. This is NVIDIA admitting that not every AI task needs a sledgehammer. Agents routing emails, managing workflows, or making API calls need efficient, flexible compute — not brute-force training rigs.

Huang also teased what comes after Vera Rubin: the Feynman architecture, featuring a CPU called Rosa (after Rosalind Franklin), next-gen networking with BlueField-5, and co-packaged optics. NVIDIA isn’t iterating. It’s planning several generations out.

NVIDIA + Groq: Splitting Inference in Two

The most technically fascinating announcement was the Groq integration. NVIDIA is carving inference into two distinct stages:

  • Prefill (converting your request into AI tokens) runs on Vera Rubin GPUs
  • Decode (generating the AI’s response) runs on Groq’s specialized Language Processing Units

Think of it as a translator and a writer working as a team, each optimized for their job. It’s a tacit admission that GPUs aren’t the best tool for every part of the inference pipeline — and NVIDIA would rather integrate a competitor’s tech than lose the market.

This also positions NVIDIA against the custom silicon tide from Google, Amazon, and Meta. The message: we’ll build the best system, even if not every chip inside says NVIDIA.

OpenClaw: “The Operating System of Agentic Computers”

If one theme united every announcement, it was AI agents. And the platform Huang couldn’t stop talking about was OpenClaw, the open-source agent framework that’s exploded since launch.

“OpenClaw is the most popular open-source project in the history of humanity,” Huang declared. Bold? Absolutely. But the tools NVIDIA announced for the ecosystem are serious:

  • NemoClaw: Policy enforcement, network guardrails, and privacy routing for enterprise OpenClaw deployments
  • OpenShell: A secure runtime that lets agents access systems and files without running loose on corporate networks

Huang compared OpenClaw to HTML and Linux. “Every single company in the world today needs to have an OpenClaw strategy.” Whether you buy the comparison, the signal is clear: NVIDIA is betting agents — not chatbots, not image generators — are the next dominant computing paradigm.

Disney Robots, Self-Driving Cars, and Space

Because GTC wouldn’t be GTC without a “did that just happen” moment:

Disney’s Olaf joined Huang on stage. The Frozen robot showcased NVIDIA’s Newton physics engine for real-world movement, built with Disney and Google DeepMind. Last year it was Star Wars droids. The progression toward consumer-facing AI robotics is accelerating.

Four new autonomous driving partners — Hyundai, BYD, Nissan, and Geely — signed onto NVIDIA’s Drive Hyperion robotaxi platform. “The ChatGPT moment of self-driving cars has arrived,” Huang said. Level 4 autonomy, from data center training to in-vehicle compute, all on one stack.

And then: space. NVIDIA revealed Space-1 Vera Rubin, a module designed to bring AI data centers into orbit. As tech giants run out of real estate and power for terrestrial data centers, orbital computing with solar power is apparently no longer science fiction. It’s on the roadmap.

What This Actually Means

Strip away the showmanship and three trends emerge:

AI is shifting from creation to action. The training era produced impressive models. The inference era puts them to work. Agents that book travel, manage operations, and run business processes aren’t theoretical — the infrastructure is being purpose-built for them right now.

Platforms beat models. NVIDIA isn’t selling chips anymore; it’s selling a complete stack from silicon to software. The winners in this next phase won’t have the smartest model. They’ll have the most complete ecosystem.

Compute is becoming a utility. When NVIDIA talks $1 trillion, they’re really saying AI compute will be as ubiquitous as electricity. Every company, device, and service will need it. The question isn’t if but how much.

The Bottom Line

GTC 2026 was NVIDIA repositioning itself as the infrastructure layer for a world where AI agents are as common as apps, inference demand dwarfs training, and the ceiling for compute demand doesn’t exist.

Whether Huang’s $1 trillion forecast proves conservative or wildly optimistic, the direction is locked in. The AI industry just left the lab and entered the factory.

NVIDIA intends to supply every tool on the assembly line.