In 2012, three University of Toronto researchers trained an image-recognition system on two NVIDIA graphics cards. The network needed five to six days to learn from roughly 1.2 million images. Their work, now known as AlexNet, helped show what deep learning could do when researchers had enough computing power and a practical way to use it.

The cards were important. So was the software that let the researchers make them do something other than draw graphics.

That second part is central to Jensen Huang's contribution. Under his leadership, NVIDIA invested in making its processors useful to people solving problems far beyond video games. CUDA, the programming platform announced in 2006, helped turn specialized hardware into something other researchers and companies could build on.

It is worth understanding that achievement now. NVIDIA announced Huang's departure as president, CEO and a board member on November 12; its regulatory filing dates the resignation to November 11. Colette Kress became interim CEO while retaining her CFO role. The Associated Press has reported calls for Huang's return, including support from Elon Musk and Tim Sweeney.

The board's independent members have announced an investigation. Those announcements do not establish its findings, and admiration for Huang's technical leadership cannot settle it. His contribution can still be assessed on its own evidence. There is a substantial positive case to make.

Making Graphics Hardware Useful to Scientists

A graphics processing unit, or GPU, is good at performing many similar calculations at once. That is useful when drawing an image, but also when working through the repeated arithmetic in a neural network. A conventional processor still handles work that needs a different pattern of execution. The opportunity is to give each kind of processor suitable work.

Before CUDA, researchers could already use GPUs for non-graphics tasks. The awkward part was often expressing their problem through interfaces designed for graphics. They had to make a scientific calculation look enough like graphics work for the hardware to accept it.

NVIDIA's CUDA 1.0 programming guide, dated June 2007, documents the change. Programmers could send computations to the GPU more directly, using extensions to the C programming language. The platform included the driver, runtime and mathematical libraries needed to do useful work, not just a fast processor on a specification sheet.

This did not make every program faster. Developers still had to divide suitable work into pieces the GPU could execute efficiently. But it gave them a clearer route from a problem to a working implementation.

For a scientist, that distinction matters. Hardware capability is only useful if the time and expertise needed to harness it are within reach. The software is part of the scientific instrument.

A Bet Made Before the Demand Was Obvious

The patient part of the story comes from Huang's own account. In his 2024 conversation with Patrick Collison at Stripe Sessions, he recalled adding costs to NVIDIA's chips before a large market for CUDA applications existed. He described years of investment before the strategy paid off.

“I believed in CUDA because you reasoned about it,” he said.

That is a founder's retrospective, not an independent audit of every early business decision. What makes it persuasive is the product NVIDIA actually built: a programming system that allowed people outside the company to put its hardware to new uses.

Huang did not need to predict every successful application. The strategic judgment was broader: many important problems involved large amounts of computation, and graphics hardware could help with some of them if developers had usable tools. Supporting those tools gave researchers a reason to experiment before the commercial outcome was obvious.

NVIDIA is marking CUDA's twentieth anniversary this month. The interesting anniversary is not simply that a product survived. It is that a bet on making hardware programmable left room for other people to discover what it was good for.

Two Graphics Cards Made the Opportunity Concrete

Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton supplied one of the clearest demonstrations. Their original 2012 paper describes training a large neural network using two NVIDIA GTX 580 GPUs, each with three gigabytes of memory.

The paper's five-to-six-day training run is a tangible measure of what was possible. A research team could use graphics cards to train a system on a large collection of labeled images. Separately, the team's competition entry, which combined several networks, won that year's ImageNet classification challenge decisively.

The achievement belonged to the researchers. It depended on their network design, training methods, data and engineering as well as the processors. Huang's role was not to invent their result. It was to lead a company that built part of the foundation on which they could produce it.

That is a stronger account of his contribution than crediting one person with inventing modern AI. It explains how useful leadership works at a distance: someone else gets to make the discovery.

The growth in AI training compute since AlexNet also makes the scale of that early experiment striking. The systems and tasks have changed enormously.

The Achievement Other People Can Build On

An AI system ultimately depends on much more than its most visible model. Compute infrastructure includes the processors, memory, networks, power and cooling that make the work possible. Software connects that machinery to the people trying to solve a problem.

Huang's strongest legacy sits at that connection. NVIDIA's engineers built tools around the hardware, and developers used those tools to do work the company did not have to invent itself. Sustaining that effort required a longer view than shipping the next faster graphics card.

The most compelling reason to appreciate Jensen Huang is therefore concrete. He helped make a powerful kind of machine usable by people with ideas of their own. The discoveries were theirs. Helping give them the means to make those discoveries is an achievement worth celebrating.