NVIDIA’s Long Road from Graphics to AI Computing

Sequoia’s Crucible Moments recounts NVIDIA’s early graphics business and later move into accelerated computing. Jensen Huang, Chris Malachowsky, Mark Stevens, Alfred Lin and Andrew Ng describe the decisions from their own perspectives.

Huang, Malachowsky and Curtis Priem had worked on workstation graphics and semiconductors. Their professional collaboration became a friendship and eventually a conversation about founding a company.

Huang recalls meeting the other founders at a Denny’s restaurant to explore ideas. They saw personal computers moving toward the home and believed games could create demand for much better 3D graphics.

The founders had workstation experience but needed to learn the PC environment. Huang says they bought a PC and studied its hardware and software as they developed their idea.

Investor Mark Stevens recalls that established companies already competed in graphics chips, while PC gaming was not yet viewed as a large serious market. The founders were betting on demand that had not yet developed.

Huang describes deciding from first principles: if the team believed 3D graphics could change computing, they had to act despite an uncertain market. The company was founded in 1993.

NVIDIA’s first chip, NV1, combined 3D graphics with audio and other functions. The founders say the technical achievement did not match what customers wanted: a simpler, cheaper graphics product.

Huang recalls poor retail sales and returned inventory. He frames the episode as an early lesson in product-market fit, positioning and pricing rather than a failure of engineering effort alone.

The company also worked with Sega on a next-generation console. Its graphics architecture used a different approach from the triangle-based direction that the PC industry was adopting.

As Microsoft’s DirectX and other vendors followed the emerging approach, NVIDIA faced a difficult choice: finish a project built around its existing architecture or abandon that direction while running short of cash.

Huang says he explained the problem to Sega and asked to be released from the contract while still receiving payment. He credits Sega’s leadership with giving NVIDIA enough time to pursue a new path; this is his recollection of the negotiations.

Stevens recalls the board’s concern that the company might run out of money. Building chips required expensive engineering talent, and the team had already endured a failed product and an abandoned direction.

For the next chip, RIVA 128, the team believed it could afford only one manufacturing attempt. Huang says they worked backward from that constraint and completed software and validation before fabrication.

The team acquired an emulator from another company and used it to test the design before the physical chip was ready. Huang recalls the first chip working; the episode presents the story as an example of planning under severe constraints.

The episode reports strong early RIVA 128 sales and says the development process helped NVIDIA increase its release cadence. The numbers are participants’ historical recollections, not audited sales figures in the episode.

The program calls GeForce 256 the first GPU. NVIDIA’s [1999 timeline](https://www.nvidia.com/content/timeline/time_99.html) used that label; its [2001 timeline](https://www.nvidia.com/content/timeline/time_01.html) identifies GeForce3 as its first programmable GPU, a distinction from the episode’s broader wording.

As the PC graphics market matured, NVIDIA looked for uses beyond games. The idea was to use a graphics processor’s parallel computation for other workloads, while continuing to compete in its original market.

The company introduced CUDA in 2006 as a platform for general-purpose GPU computing. NVIDIA’s [programming guide](https://docs.nvidia.com/cuda/archive/13.2.0/cuda-programming-guide/01-introduction/introduction.html) places that development in the broader history of accelerated workloads.

Huang recalls researchers assembling inexpensive GPU systems for scientific work. Such examples helped the company see demand for tools beyond graphics, although each field had different computational needs.

Andrew Ng describes early experiments using GPUs for neural-network training. He reports large speedups in the workloads his team studied, while the actual gain depends on the model, hardware and implementation.

The guests describe AI computing as an uncertain market when NVIDIA committed resources to it. The company had to fund new skills and products without knowing how quickly demand would develop.

Huang says the shift required attention across the company, not merely one product team. Ng recalls that the leadership committed early; both accounts are retrospective interpretations of the decision.

Huang emphasizes celebrating intermediate signs of progress, including researchers finding the platform useful. Near-term feedback helped sustain a long project whose full commercial payoff was uncertain.

He argues that progress across chips, systems and algorithms can compound. The episode cites a large historical speedup and extrapolates it; such comparisons depend on the workload and baseline and do not guarantee a fixed rate of future progress.

The speakers close with a lesson about long time horizons and repeated adaptation. NVIDIA’s path included setbacks, a change in graphics architecture and a later expansion into new workloads; their account emphasizes choices as well as fortunate circumstances.