News

CUDA turns 20, with 11 million developers

Nvidia marks CUDA’s science and AI legacy.

Jon Peddie

Happy anniversary, CUDA. Nvidia marked 20 years of its GPU programming platform at SC26 this week. Nvidia announced CUDA in November 2006 and brought it to scientists at SC06. Today, more than 11 million developers build on it, and over 1,000 CUDA-X libraries serve fields from drug discovery to finance. Nvidia credits CUDA for work from AlexNet to the HIV capsid structure and AlphaFold. We review the history and explain what the milestone means for ISVs, silicon teams, and CIOs.

Nvidia marked the 20th anniversary of CUDA at SC26 this week. Nvidia announced the platform in November 2006 and brought it to the scientific community at SC06. The anniversary gives ISVs, silicon teams, and CIOs a reason to examine what a two-decade software platform contributes to a hardware business. Nvidia invites SC26 attendees to visit a 20-Year CUDA Celebration Event Wall at its booth and to start building with the CUDA Toolkit.

A programming model for scientists

CUDA gave scientists and developers a familiar, high-level programming model for GPUs. Nvidia says the model extended accelerated computing beyond graphics into demanding applications for science and simulation. A familiar model lowers the training burden for domain scientists, who write physics and chemistry codes and have little time to learn graphics programming interfaces. Over two decades, developers used CUDA to advance Nobel Prize-winning AI research, to accelerate scientific applications on supercomputers, and to support visual effects in Oscar-winning films. Nvidia counts more than 11 million developers who build on CUDA across the computing stack. Table 1 lists the milestones in Nvidia’s account.

Figure 1. CUDA connects Nvidia GPUs, development tools, libraries, and scientific and AI applications.

CUDA’s development over the past two decades has created a software ecosystem that connects Nvidia GPUs with programming tools, specialized libraries, and applications spanning scientific computing, simulation, and AI.

Table 1. CUDA milestones in Nvidia’s 20-year account.

Nvidia frames the list as a chain in which one generation’s work creates opportunities for the next.

These milestones trace CUDA’s evolution from a GPU programming model into a software platform supporting scientific research and AI development. Each generation expanded the applications developers could accelerate, while Nvidia’s growing library collection reduced the effort required to bring new workloads onto GPUs.

Science and AI case studies

Klaus Schulten’s team at the University of Illinois Urbana-Champaign helped connect GPU computing with molecular biology through Nanoscale Molecular Dynamics and Visual Molecular Dynamics, tools for molecular simulation and visualization. In 2013, researchers at Illinois, the University of Pittsburgh, and Vanderbilt University used CUDA-accelerated tools to help resolve the atomic structure of the HIV capsid. They combined experiments with simulations on Blue Waters, a large supercomputer of its day, in Nvidia’s description.

University of Toronto researchers Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton used CUDA to train AlexNet on two Nvidia GPUs. Their image-recognition system won the 2012 ImageNet competition and helped demonstrate the potential of GPU computing for deep learning. AlphaFold links computing to scientific discovery in the same way. Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for protein structure prediction, and the team’s publicly released AlphaFold 2 software uses CUDA for GPU execution. Table 2 summarizes the three cases.

Table 2. Science and AI cases Nvidia cites.

Nvidia selected these examples to show reach, and none of them measures CUDA’s share of any field.

What the platform means for buyers

CUDA-X now includes more than 1,000 libraries. Nvidia says developers use them to discover new medicines, design factories, advance autonomous vehicles, model financial markets, and work on quantum computing. At SC26, Nvidia added cuPotential for molecular simulations and cuSpa for spatial biology calculations, and it ties its new Rubin HPC GPU, expected in 2028, to the same library family. The pairing shows how Nvidia sells hardware and software roadmaps together.

Libraries carry the porting cost for ISVs. A vendor that calls a CUDA-X routine inherits tuning that Nvidia maintains, and each new library extends that benefit to a new field. For silicon teams, the installed base sets the bar any competing accelerator must clear. AMD’s pitch for its Instinct MI400 family, for example, leans on ROCm progress and open software for that reason. Independent testing supports the point. In MLPerf Training 6.0 in June 2026, AMD’s MI355X came within about 5% to 6% of Nvidia’s B200 on two LLM tasks, and commentary on those results, says Nvidia, leads in software ecosystem and very large cluster deployments.

CIOs should read the milestone as both a benefit and a risk. A large developer base shortens hiring and porting timelines, and lowers integration risk for Nvidia hardware. It also raises switching costs for any team that writes directly to CUDA. We recommend that CIOs map which code paths call CUDA-X libraries, which use portable frameworks, and which use custom kernels, because that map determines what a second GPU source costs to add. A small pilot of one workload on a second software stack would price that effort with real data.

What Nvidia leaves out

Nvidia’s account omits several measures that buyers need. Nvidia does not disclose how it counts a developer, how many applications run on CUDA, or how much revenue the platform supports. The 11 million developer count and the 1,000-plus library count are Nvidia’s own figures, and Nvidia gives no methodology for either. The case studies come from researchers at universities and laboratories, and Nvidia gives no data on commercial deployments. The announcement also gives no download counts for the CUDA Toolkit and no count of developers that use competing stacks, so the figures say nothing about share. Anyone who cites them should attribute them to Nvidia.

The anniversary celebrates the developer community, and the developer community is the asset that matters to a technical buyer. ISVs should inventory their CUDA dependence now and identify which libraries bind them to one supplier. Silicon teams should study which CUDA-X routines their customers call frequently, because those routines set the order in which a software stack must reach parity. CIOs should add CUDA exposure to vendor-risk reviews and ask suppliers for portability data on the workloads they run.

What do we think?

CUDA’s 20 years show how software shapes a hardware market. The 11 million developers and over 1,000 libraries give Nvidia a durable position, and rivals must answer with open software and measured results. Nvidia’s figures are self-reported and carry no definitions today. ISVs should map their CUDA dependence, and CIOs should price the cost of adding a second GPU source.

Inflection point. The anniversary does not create an inflection point in AI because it records past results and announces no new technology. It confirms that an earlier inflection point arrived when researchers trained AlexNet on two GPUs in 2012 and when AlphaFold earned a Nobel Prize in 2024. The next inflection point depends on whether open software stacks match CUDA’s libraries, and ISVs and CIOs alike will see that in early measured portability data through 2027. Until then, the milestone records history.

LIKE IT? THINK YOUR FRIENDS AND ASSOCIATES MIGHT? PLEASE SEND IT TO THEM WITH OUR BEST WISHES.