Nvidia Stopped Selling Chips — It’s Selling the Agent Now

Jensen Huang’s Computex 2026 keynote announced six new chips, a PC superchip, a robotics platform, and a 500-billion-parameter open model. But the real announcement was structural, not silicon: Nvidia stopped selling the fastest chip. It started selling the unit of computation itself — the agent.

This is the most significant strategic pivot in Nvidia’s history. And most of the coverage missed it entirely.

The Agentic Expansion Cascade

AI computing is moving through a four-phase progression that changes what a “computer” even means:

Phase 1: Sandbox. AI runs inside a text box. You type a prompt, it returns text. ChatGPT — as explored in the intelligence factory race between AI labs — in 2023. No tool access, no persistence, no agency. The AI is a calculator.

Phase 2: Tool use. AI gains access to APIs, databases, code interpreters. Claude with computer use. GPT with function calling. The AI becomes a research assistant — it can look things up and take limited actions, but it still operates within predefined boundaries.

Phase 3: Computer operation. AI takes over the interface. It clicks buttons, navigates applications, fills forms, manages workflows. This is where we are today — Anthropic’s computer use, OpenAI’s Codex operating autonomously, Microsoft’s Copilot managing entire workflows across Office. The AI operates the human’s tools.

Phase 4: Convergence. The intelligence layer becomes the interface. There is no desktop, no app, no menu. You state a goal. The agent executes it. The computer disappears behind the intelligence. This is where Nvidia is building toward.

What Jensen Actually Announced

Read Nvidia’s Computex lineup through this lens and the strategy becomes clear:

Vera Rubin (data center) doesn’t just deliver 5x more inference performance. It delivers 10x lower cost per token. That’s not a chip upgrade — it’s making agents economically viable to run 24/7. When inference costs drop 10x, you don’t just run the same workloads cheaper. You run agents that never stop.

Rubin CPX (inference-specific) is purpose-built for million-token context windows. That’s not about answering longer questions. It’s about agents that maintain persistent memory across entire workflows — agents that remember everything they’ve done and everything they need to do.

RTX Spark (client) puts 1 petaFLOP of AI compute in a laptop. Jensen said it would “turn Windows into an agentic AI OS.” That’s Phase 4 language — the computer becomes the agent.

Nemotron 3 Ultra (500B open model) is optimized for “multi-step task execution with minimal human oversight.” Not question-answering. Not content generation. Autonomous task execution. Agent behavior.

GR00T N2 (robotics) extends the agent from digital to physical. The same progression — sandbox → tools → computer operation → convergence — applied to the real world. The robot is the physical agent.

The Business Model Shift

Nvidia’s revenue has historically been measured in chips sold, GPUs shipped, data center racks deployed. The Computex announcements reveal a different endgame: Nvidia wants to be measured in agents running.

Consider the economics. A chip is a one-time sale (with replacement cycles). An agent that runs 24/7 on Nvidia infrastructure — as explored in the economics of AI compute infrastructure — consumes compute continuously. Every token processed, every inference run, every agent action generates consumption revenue — either directly (via Nvidia’s cloud partnerships) or indirectly (via hardware demand from the hyperscalers running these agents).

This is why Nvidia built the full stack — not just the GPU, but the CPU (Vera), the networking (NVLink 6, ConnectX-9), the security (BlueField-4), the switching (Spectrum-6), the software (CUDA, NeMo, DSX), and now the client device (RTX Spark). Every layer is a toll on every watt the world converts into tokens.

The Second Revolution in Computing

The first revolution in computing was the graphical interface. Before the GUI, computers required specialized knowledge — command lines, programming languages, punch cards. The GUI made computers accessible to humans by creating a visual metaphor (desktops, folders, windows) that translated human intent into machine operations.

The second revolution eliminates the metaphor entirely. Instead of a human operating a visual interface that controls a computer, an AI agent receives a goal and operates the computer directly. The interface doesn’t get better. It disappears.

Jensen’s statement — “40 years of traditional PCs is now at an end” — is not about hardware specs. It’s about this structural shift. The PC was built for human-computer interaction via interfaces. The next device is built for goal-agent interaction via intelligence. RTX Spark is the first device designed for this paradigm.

Who Wins in the Agent Economy

If the intelligence layer becomes the interface, three types of companies capture value:

Infrastructure providers (Nvidia, TSMC, Broadcom) — they sell the compute that agents consume. Every agent running is revenue. This is why Nvidia is worth $5.23 trillion.

Model providers (Anthropic, OpenAI, Google) — they provide the intelligence that agents use. But intelligence is commoditizing faster than infrastructure. The model wars are approaching parity.

Orchestration platforms (Microsoft, Salesforce, Palantir) — they provide the workflows that agents execute inside. The company that owns the enterprise workflow owns where the agent lives.

The losers are companies that sell interfaces — traditional software vendors whose value proposition is “a better UI.” When the agent operates the computer, the UI is irrelevant. The best-designed dashboard in the world doesn’t matter if no human looks at it.

Nvidia’s Computex 2026 wasn’t a product launch. It was a declaration that the company understands this shift better than anyone — and has built every layer needed to profit from it.

For the deep analysis of agent-computer convergence and the five-layer AI economy, read The AI Agent Is Taking Over The Computer on Business Engineer.