NVIDIA's moat is CUDA's install base, and extreme co-design keeps scaling compute as AI becomes agentic.
Because the problem no longer fits inside one computer accelerated by one GPU. You add 10,000 computers but want a million times the speed, so you refactor the algorithm and shard the pipeline, the data and the model. Then Amdahl's Law bites: CPU, GPU, networking and switching all become the bottleneck.
It consumed all of the company's gross profit dollars. NVIDIA was worth roughly $8 billion, maybe six or seven, before CUDA launched; afterwards the market cap fell to about $1.5 billion and stayed down for a while. Costs rose about 50% at a company running 35% gross margins.
The install base of its computing platform, specifically CUDA. Someone arriving with a Gouda or a Tuda would make no difference, because it was never only the technology: 43,000 people and several million developers built on it. Target CUDA and you reach a few hundred million computers, every cloud, every industry.
Because inference is thinking, and thinking is far harder than reading. Pre-training is memorisation and generalisation, looking for patterns and relationships. Test-time scaling means decomposing a problem, reasoning from first principles or prior experience, exploring and searching and planning. That cannot possibly be compute-light; it is intensely compute intensive.
One agentic system does research, hits databases, uses tools, and spawns a whole bunch of sub-agents, so you are creating large teams. Jensen says it is easier to scale NVIDIA by hiring more employees than to scale himself. Agents multiply AI, and their experiences feed back into pre-training and post-training.
The grid is sized for worst-case conditions, a few days in winter, a few in summer, extreme weather, and otherwise runs around 60% of peak, so 99% of the time excess power sits idle. He wants contracts where data centers gracefully degrade, shift workloads, or simply run slower.
Started 30 years ago, speed of light is his shorthand for the limit of what physics can do. Everything is tested against it: memory speed, math speed, power, cost, time, effort, number of people, manufacturing cycle time. Rather than trimming 74 days to 72, ask what building from scratch takes.
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. One NVL72 rack alone is 1.3 million components, 1,300 chips, 4,000 pounds.
Agentic systems can access sensitive information, execute code, and communicate externally. NVIDIA gives you two of those three capabilities at any time but never all three, adds access control based on whatever rights the enterprise grants you, and connects it to the policy engines enterprises already run.
About 50% of the world's AI researchers are Chinese and most are still in China. The tech industry grew up in the software era, provinces and cities compete fiercely so there are many EV and AI companies, and family-friend-schoolmate culture shares knowledge instantly, which makes open source natural.
“Install base defines an architecture.”
“And that was always illogical to me because inference is thinking.”
“The speed of light is not just about the speed.”
“OpenClaw is the iPhone of tokens.”
“Jensen Huang: NVIDIA - The $4 Trillion Company & the AI Revolution | Lex Fridman Podcast #494.” https://www.youtube.com/watch?v=vif8NQcjVf0. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/jensen-huang-nvidia-the-4-trillion-company-the-ai-revolution.html.