Jensen Huang: NVIDIA - The $4 Trillion Company & the AI Revolution | Lex Fridman Podcast #494

Lex Fridman states that NVIDIA is one of the most important and influential companies in the history of human civilization.

Lex Fridman states that NVIDIA has propelled itself into a new era in AI, moving beyond chip scale design to rack scale design.

Lex Fridman states that winning for NVIDIA used to be about building the best GPU possible, and they still do.

Lex Fridman states that NVIDIA has expanded its focus to extreme co-design of GPU, CPU, memory, networking, storage, power, cooling, software, the rack itself, the pod, and even the data center.

Jensen Huang states that if computation represents 50% of a problem and is sped up infinitely, the total workload is only sped up by a factor of two due to Amdahl's Law.

Jensen Huang states that in distributed computing at NVIDIA's scale, the CPU, GPU, networking, switching, and distributing the workload across computers are all problems.

Jensen Huang states that Moore's Law has largely slowed because Dennard scaling has slowed.

Jensen Huang states that the reason for extreme co-design is to distribute the workload to exceed the benefit of just increasing the number of computers.

Jensen Huang states that his direct staff at NVIDIA is 60 people, and he does not have one-on-ones with them.

Jensen Huang states that almost all of his 60+ direct reports have a foot in engineering, being experts in memory, CPUs, optical, GPUs, architecture, algorithms, and design.

Jensen Huang states that NVIDIA started as an accelerator company, but the problem with accelerators is that the application domain is too narrow.

Jensen Huang states that NVIDIA decided to put CUDA on GeForce, which was selling millions of GPUs a year, to cultivate an install base by putting a supercomputer in the hands of every researcher and student.

Jensen Huang states that people initially thought the amount of high-quality data would limit AI intelligence, but this pre-training scaling law showed larger models with more data result in smarter AI.

Jensen Huang states that NVIDIA had NVLink 72 instead of NVLink 8 when mixtures of experts came out, allowing a 4-10 trillion parameter model to run in one computing domain.

Jensen Huang states that OpenClaw did for agentic systems what ChatGPT did for generative systems.

Jensen Huang states that agentic systems can be kept safe by allowing only two out of three capabilities (access sensitive info, execute code, communicate externally) at any time, not all three.

Lex Fridman states that the Vera Rubin pod includes seven chip types, five purpose-built rack types, 40 racks, 1.2 quadrillion transistors, nearly 20,000 NVIDIA dies, over 1,100 Rubin GPUs, 60 exaflops, and 10 petabytes per second of scale bandwidth.

Jensen Huang states that NVIDIA GPUs are the first GPUs in space.

Jensen Huang states that intelligence is a commodity.

Was it like $8 billion or something?