Three AI Inflection Points: From ChatGPT to Agents and Physical Intelligence

At a March conference in San Francisco, NVIDIA CEO and founder Jensen Huang described three critical inflection points in AI development over the past two years. His main takeaway: compute has stopped being a supporting resource and has become the foundation of a new economy, where AI directly affects corporate revenue and national GDP. Each inflection point radically changed the technology's usefulness, the volume of tokens consumed, and electricity consumption. Huang is convinced that the software development industry will soon become fully token-driven.

Why NVIDIA Grew at Hyperspeed

From the very beginning, NVIDIA focused not on individual chips but on building a complete computing stack — from algorithms to architecture to ecosystem. Back in the video game era, the company established an approach of full control: low-level algorithms, deep integration into software engines, and systems architecture.

The company's culture was originally built around annually updating the entire stack — from processors and interconnects to architecture. Only full control over every layer allows innovation to be deployed at this speed. When a company owns the entire infrastructure, it can change it every year instead of coordinating components from different suppliers.

Three Inflection Points in AI

First: generative AI. The ability to translate information from one form to another and auto-generate tokens. GPT-3 existed for a long time as a model before it was wrapped in the user-friendly ChatGPT interface. However, generative systems suffered from hallucinations — they didn't rely on up-to-date context or verified data.

Second: reasoning AI (the O1 model). This brought reliance on research, facts, and RAG (retrieval-augmented generation). AI learned conditional generation: answers formed based on context and ground truth. Self-analysis and real-time self-correction emerged. Usefulness surged, the volume of generated tokens grew by hundreds of times, and compute consumption grew by a thousand times. ChatGPT went from a novelty to a reliable tool.

Third, current: agentic AI. AI can now work with files and tools, perform searches, and plan and execute complex tasks. Prompts changed: instead of questions, people now give commands — "create," "build." Agents independently read documentation, browse the web, and apply tools. A single agent consumes a million times more tokens than a previous chatbot and runs continuously in the background. OpenClaw became the fastest-growing open-source project in history, overtaking Linux in download numbers in just three weeks.

The Token Economy and the Shift to Token-Driven

Huang suggested thinking of data centers not as data storage facilities but as AI factories whose primary product is tokens. The more compute power, the more tokens, the higher the revenue. The volume of computing power directly correlates with companies' revenues and countries' GDP.

The key metric is tokens per watt. Given limited available power (1-2 gigawatts per factory), architectural efficiency determines the size of a company's revenue in the following year. The decision on which computer architecture to choose is now made at the CEO level.

According to Huang's forecast, this will transform the entire IT industry. Today, it mostly rents tools. Tomorrow, every software company will build agents — combining open and closed models, hiring "digital workers," and renting out specialized tokens. A trillion-dollar industry that barely consumed tokens will turn into one of their largest consumers, creating enormous new demand for compute. The internet industry has already shifted all of its capex toward AI systems.

The Next Frontier: Physical AI

In the coming years, agentic systems will take a back seat. Physical AI will take over, requiring an understanding of the laws of physics, cause-and-effect relationships, gravity, inertia, and object permanence — a fundamental grasp that objects continue to exist even when they disappear from view.

NVIDIA is actively developing models in this area, from physics simulation to robotics and digital biology. In two to three years, physical AI will become the defining theme of the next decade. It will unlock AI applications beyond data centers — in robots, manufacturing, autonomous vehicles, and the real world. Huang is convinced that compute growth cannot be contained. In the future, every company will need computing power to generate revenue, and every country will need it to maintain the competitiveness of its national economy.