The limits of AI scaling laws - NVIDIA CEO explains | Jensen Huang and Lex Fridman

Pre-training continues via synthetic data

Fears that high-quality human data would run out were unfounded, as AI can now generate synthetic training data from ground truth, shifting the constraint from data availability to compute capacity.

Test-time scaling is compute-intensive thinking

Unlike pre-training's pattern matching, inference now involves complex reasoning, planning, and search, making test-time scaling far more computationally demanding than previously assumed.

Agentic scaling multiplies AI workers

Individual AI agents spawn sub-agents to form teams, creating a multiplicative effect on computational requirements that far exceeds scaling a single model.

The closed-loop data cycle

Agentic systems generate experiences that feed back into pre-training as memorized knowledge, creating a continuous cycle where compute drives intelligence across all four stages.

Anticipating AI architectures 2-3 years ahead

With AI models evolving every 6 months but hardware requiring 3-year cycles, NVIDIA must predict future needs, as demonstrated by designing Vera Rubin racks for agentic computing before OpenClaw existed.

Flexibility versus specialization balance

CUDA provides the necessary adaptability to support rapidly changing algorithms like mixture of experts while maintaining the specialization required for acceleration beyond general-purpose CPUs.

Infrastructure shifts from LLM to agent processing

Grace Blackwell systems optimized for large language model inference are giving way to Vera Rubin architectures that include storage accelerators and IO subsystems necessary for tool-using agents.

Digital workers use existing tools

Rather than replacing software, future AI agents will access files, conduct research, and operate tools like OpenClaw through APIs, functioning as humanoid robots that read manuals rather than reinventing functionality.

The two-out-of-three security principle

NVIDIA's security framework restricts agents to only two capabilities at any time—accessing sensitive data, executing code, or external communication—to prevent dangerous combinations while maintaining utility.

OpenClaw as a ChatGPT moment

Open-source agentic systems represent a pivotal inflection point for AI adoption, similar to ChatGPT's impact on generative AI, requiring robust security integration like NVIDIA's Nemo Claw and Open Shell.

Power efficiency as the primary constraint

While power availability remains challenging, the immediate focus is extreme code design to improve tokens-per-watt efficiency, driving token costs down an order of magnitude annually despite rising hardware prices.

Supply chain coordination at unprecedented scale

NVIDIA works closely with ASML, TSMC, and DRAM manufacturers to address lithography, packaging, and HBM memory bottlenecks, having successfully convinced the industry to shift from DDR to HBM for data centers.