Nvidia Part II: The Machine Learning Company (2006-2022)

The hosts frame NVIDIA's later Omniverse vision as the outcome of a long expansion beyond gaming. GPUs can simulate physical systems, from airflow around a wing to cell interactions and climate models, but those uses depend on hardware, developer tools, user-facing software, and services working together. The computational scale is difficult to picture: even a single speech-recognition model can require more arithmetic operations than there are grains of sand on Earth. The episode therefore treats NVIDIA's evolution as a shift from selling graphics cards to building a full stack for multiple industries.

By 2004-2006, NVIDIA had survived two major threats: intense competition in graphics cards and Intel's attempt to commoditize the GPU. Its response was to make GPUs programmable, build programmable shaders with the GeForce 3, work with Microsoft on CG, and supply graphics technology for Xbox. NVIDIA also shipped new chip generations roughly every six months, while competitors often took eighteen months or two years. That pace helped it establish a strong position in gaming, but the longer-term consequence was a technical platform on which new workloads could be built.

A less visible choice was to write and maintain NVIDIA's own graphics drivers. Other companies often left driver development to downstream partners; NVIDIA wanted to control the user's experience across operating systems and hardware configurations. This created a costly fixed investment, but it also built a base of low-level software expertise inside a chip company. At the same time, programmable shaders began a direct relationship with developers: programmers could target NVIDIA hardware and gain capabilities unavailable through generic operating-system software. Together, driver control and developer engagement widened NVIDIA's role beyond manufacturing chips.

By 2007, NVIDIA was a successful public company with a market capitalization near $20 billion, and gaming alone could have supported a strong business. Jensen Huang was not satisfied with that outcome. A Stanford researcher reportedly found that GeForce cards could run a quantum-chemistry model ten times faster than a lab supercomputer, an early signal that GPUs might serve scientific computing. The researcher had to express the problem through graphics-oriented CG shaders, however, exposing how awkward the tools were. Serving general-purpose computing properly would require much more than a graphics API: NVIDIA would need to build a broader programming stack for developers.