Nvidia Part III: The Dawn of the AI Era (2022–2023)

In 2021, Nvidia described its market opportunity from the top down: industries such as cars and factories represented roughly $100 trillion, and capturing just 1% would amount to a $1 trillion total addressable market. The estimate depended on applications that had not yet arrived, including autonomous vehicles, Omniverse, and robotics. In the hosts' earlier discussion of Nvidia's valuation, they argued that a trillion-dollar case required those physical-world applications to happen soon.

Another possibility was that Nvidia could benefit from the continued growth of the internet and software without waiting for machines to become ubiquitous. Social recommendations had already shown that a new computing layer in the digital world could create enormous value. Looking back, the hosts suggest that this once-unanswered question may now have an answer: generative AI has created a new source of compute demand, visible in Nvidia's revenue and earnings.

The opportunity arrived at an unusual moment. In 2022, rising interest rates pressured stocks, startups, and real estate; the crypto and Web3 bubbles burst, and Nvidia recorded an inventory write-down. At the same time, large language models, after years of laboratory work, reached the public. ChatGPT's rapid adoption turned generative AI into a product opportunity and raised the episode's central question: how did this shift happen, and why is it built on Nvidia's hardware and software?

The story returns to AlexNet in 2012. The ImageNet competition asked systems to classify 14 million manually labeled images; at the time, machine learning could not reliably perform this task, and large numbers of low-paid Mechanical Turk workers supplied the labels. AlexNet reduced the error rate from roughly 25% to 15%, a clear leap rather than a small incremental gain.

AlexNet did not invent neural networks: neural networks, especially convolutional neural networks, had existed for decades. The obstacle was the cost of training them. The Toronto team used two consumer GeForce GTX 580 cards and CUDA to train a convolutional network for about $1,000 in hardware. GPUs' parallel execution made a computation that had seemed to require a supercomputer laboratory practical.

Graphics are well suited to parallel processing because each screen pixel can be computed independently. Technology Nvidia first developed for graphics later applied to machine learning, cryptocurrency, and other matrix operations. AlexNet was therefore more than an image-recognition result: it marked the arrival of accelerated computing in AI research.

The technology quickly entered commercial systems. AI recommendations could determine what appeared in social feeds and unlock billions of dollars in value. Google and Facebook began buying large numbers of Nvidia GPUs, showing how parallel computing had moved from a research tool to production infrastructure for major internet companies.

The people behind AlexNet became part of the next chapter. The team included Alex Krizhevsky, his advisor Jeff Hinton, and fellow doctoral student Ilya Sutskever. Hinton's family history reaches back to George and Mary Boole, whose work underlies Boolean algebra; Sutskever later became an OpenAI co-founder and chief scientist. Technical breakthroughs, talent movement, and new companies began to intersect.

AlexNet's researchers and others in the field soon became targets for major technology companies. Google and Facebook recruited them and began buying more Nvidia GPUs. Ilya Sutskever, one of the AlexNet researchers, later co-founded OpenAI and became its chief scientist. That movement of talent connects early machine-learning breakthroughs to the later development of generative AI.

Google absorbed the AlexNet breakthrough into its research organization: team members joined the new Google Brain group and worked alongside other leading researchers. Google acquired DeepMind, while Facebook recruited Yann LeCun, concentrating top AI talent in a small number of large technology companies. These systems were not general intelligence; they were good at bounded tasks such as feed recommendations.

Bounded tasks could still produce enormous business value. Google Brain researchers helped reshape YouTube's algorithm, contributing to its transformation from a money-losing acquisition into a giant platform. AI-powered feed recommendations also helped make Instagram an exceptionally valuable Facebook asset. Targeted ads and recommendations generated profits while consuming Nvidia GPUs, tying AI research, cloud computing, and internet business models together.