NVIDIA's edge is extreme co-design plus CUDA's install base, letting compute scale as intelligence keeps scaling.
Because the problem no longer fits inside one computer accelerated by one GPU. You add 10,000 computers but want a million times more speed, so you shard the pipeline, the data and the model. Then Amdahl's law bites: CPU, GPU, networking and switching all become problems at once. Otherwise you scale only linearly, with Moore's law, which has slowed.
Answered around 7:14The install base of its computing platform, meaning CUDA. A rival called a Gouda or a Tuda would make no difference, because it was not three people but 43,000 people and several million developers who made CUDA successful. Second is the ecosystem: one architecture in Google Cloud, Amazon, Azure, CoreWeave, cars, robots and satellites.
Answered around 1:21:28Because install base defines an architecture, and developers go where the reach is. GeForce was already selling millions of GPUs a year, so CUDA went into every single PC whether customers used it or not, as a starting point for cultivating install base, while NVIDIA went to universities, wrote books and taught classes.
Answered around 19:14It consumed the company's gross profit dollars. Cost went up about 50% at a company running 35% gross margins, and gamers would not pay for it. NVIDIA was worth roughly six to eight billion dollars; after CUDA launched the market cap fell to about one and a half billion, and the payoff took a decade.
Answered around 20:42Four: pre-training, post-training, test time and agentic. Pre-training was thought limited by high quality data, and when Ilya said we are out of data the industry panicked that AI was over. That was wrong: most training data will be synthetic, so training is now limited by compute rather than by data.
Answered around 29:18People assumed pre-training was hard and inference easy, so inference chips would be little tiny commodity chips. Jensen found that illogical: inference is thinking, and thinking is harder than reading. Pre-training is memorization, generalization and pattern finding, while test time is reasoning, planning, search and decomposition, so it is intensely compute intensive.
Answered around 31:50During test time the agentic system does research, bangs on databases, uses tools, and spawns subagents, so you are creating large teams. It is multiplying AI: agents spin off as fast as you want. The experiences they create feed back into pre-training, post-training and test time, and the loop goes on and on.
Answered around 33:38By pushing extreme co-design to improve tokens per second per watt by orders of magnitude every single year. In the last ten years Moore's law would have progressed computing about 100 times; NVIDIA scaled it by a million times. Computer prices go up, but token cost comes down an order of magnitude every year.
Answered around 44:06The grid is designed for worst case conditions but runs around 60% of peak, so 99% of the time excess power sits idle. Three fixes: customers stop demanding six nines of uptime, data centers gracefully degrade to about 80% and shift workloads, and utilities sell power at different guarantee levels and prices.
Answered around 53:36Seven 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. One NVL72 rack alone is 1.3 million components, 1,300 chips and 4,000 pounds, and about 200 pods ship weekly.
Answered around 1:05:46“Install base is everything.”
“The speed of light is my shorthand for what's the limit of what physics can do.”
“It is the most complex computer the world has ever made.”
“And so this is the fastest innovating country in the world today.”
“Jensen Huang on NVIDIA's AI Revolution.” https://media.blubrry.com/takeituneasy/ins.blubrry.com/takeituneasy/lex_ai_jensen_huang.mp3. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/jensen-huang-on-nvidia-s-ai-revolution.html.