NVIDIA GTC Keynote 2026

Jensen opened by declaring 2025 «NVIDIA's year of inference» and explained why computing demand has surged beyond all forecasts. He traced three inflection points: ChatGPT (generative AI era), OpenAI's O1 (reasoning models that reflect, plan, and decompose problems), and Anthropic's Claude Code (the first agentic model that reads, compiles, tests, and iterates code). Each wave multiplied both token input (context) and output (thinking) by orders of magnitude.

He quantified the shock: «The computing demand of the work has gone up by 10,000 times, and usage has gone up by 100 times. That's why computing demand has increased by one million times in the last two years.» This is no longer training-dominated—AI must now reason, generate, and act in production, which means continuous inference at massive scale. Every startup, every hyperscaler, every enterprise is constrained not by training capacity but by inference throughput.

The consequence: 100% of NVIDIA engineers now use a combination of ChatGPT, Claude Code, Codex, and Cursor. «There's not one software engineer today who is not assisted by one or many AI agents,» Jensen said. The shift from perception to generation to reasoning to action has made AI «able to do productive work»—and that productive work consumes tokens at rates no one anticipated even twelve months ago.

Jensen introduced the mental model every AI CEO will adopt: «Your data center is no longer a file repository—it's a factory that generates tokens.» He drew a two-axis chart: vertical = throughput (tokens per watt), horizontal = token speed (latency/interactivity). The smarter the AI (longer context, more reasoning), the lower the throughput; the faster the inference, the higher the price tier.

NVIDIA segments the market into four tiers: free (high throughput, low speed), $3/million tokens (medium), $6/million, and $45–$150/million for premium, ultra-low-latency research workloads. «As a research team, $150 per million tokens using 50 million tokens per day is not even a thing,» Jensen said. The key insight: Grace Blackwell lifted every tier, but Vera Rubin + Grok created an entirely new premium segment—35× faster than Blackwell at the highest-value tier. This architectural leap translates directly into revenue expansion: a simplified model shows Blackwell generating 5× more revenue than Hopper in the same gigawatt, and Vera Rubin another 5× over Blackwell.

Jensen's punchline: «What you do this year will show up precisely next year as your revenues. This chart is what it's all about.» Every hyperscaler, cloud provider, and AI startup will now manage their infrastructure as a manufacturing operation, balancing power allocation across tiers to maximize token margin.

Jensen called OpenClaw «the most popular open source project in the history of humanity… it exceeded what Linux did in 30 years, in just a few weeks.» Peter Steinberger's framework provides scheduling, resource management, multi-modal I/O, tool access, and agent orchestration—essentially an OS for agentic computers. Within days of launch, «Hundreds of people are queuing up,» and a ClawCon conference materialized.

The implication is massive: every SaaS company must now have an «OpenClaw strategy,» just as they once needed Linux, Kubernetes, or HTTP strategies. NVIDIA responded by releasing NemoClaw, a secure, enterprise-ready reference design built on OpenClaw. It adds OpenShell (network guardrails, privacy router, policy engine integration) to prevent agents from accessing sensitive data or executing unsafe code. «Access sensitive information, execute code, communicate externally—obviously this can't possibly be allowed,» Jensen warned.

NVIDIA also bundled its open frontier models (NemoTron, Cosmos, Groot, BioNemo, AlpaMayo, Earth-2) and launched the NemoTron Coalition—partners include Black Forest Labs, Cursor, Mistral, Perplexity, and others committing to co-develop NemoTron 4. The goal: enable every enterprise to deploy domain-specific agents safely, using world-class base models that NVIDIA will «keep working on every single day… NemoTron 3 followed by NemoTron 4, Cosmos 1 by Cosmos 2.»

NVIDIA introduced DSX, a digital twin platform where all vendors—compute, cooling, electrical, networking—meet virtually before hardware ships. Partners include Siemens (Star CCM Plus thermals), Cadence (Reality internal simulation), ETAP (electrical), PTC Windchill (PLM), Dassault (3D Experience systems engineering), Jacobs (custom Omniverse apps), and Procore (virtual commissioning).

Once live, the twin becomes an AI agent-driven operator. DSX MaxQ dynamically orchestrates infrastructure: Phaedra agents manage cooling and electrical, Emerald agents interpret grid demand and adjust power. Jensen: «There's no question in my mind there's a factor of two in here. And the factor of two at the scale we're talking about is gigantic.» At trillion-dollar capex scale, a 2× efficiency gain is worth hundreds of billions in recovered capacity and avoided buildout. DSX aims to eliminate squandered power, maximize token throughput, and compress construction timelines across every AI factory deployment globally.