AI · Oct 28, 2026
RTX Spark laptops get open speech models and a dial for background agentsTwenty years of CUDA: the software bet that made Nvidia hard to replace
· Updated · Signal Desk newsroom
Analysis of reporting by NVIDIA Blog. How we work

Nvidia marked CUDA’s 20th anniversary at SC26 by honoring the researchers behind AlexNet, AlphaFold and HIV capsid simulations. The tribute is earned. It also shows why Nvidia’s software moat outlasts any single executive.
Key takeaways
- Nvidia announced CUDA in November 2006 and brought it to scientists at SC06; it marked the 20th anniversary at SC26.
- Nvidia says more than 11 million developers now build on CUDA, with more than 1,000 CUDA-X libraries.
- The tribute highlights Klaus Schulten’s molecular dynamics tools, the 2013 HIV capsid structure, AlexNet’s 2012 ImageNet win on two GPUs, and AlphaFold 2.
- Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry; AlphaFold 2 uses CUDA for GPU execution.
- The anniversary lands during Nvidia’s leadership crisis, a reminder that its platform advantage is institutional, not personal.
A programming model, not a chip
Nvidia’s CUDA anniversary post is a celebration, and most of it is deserved. In November 2006 Nvidia announced CUDA and took it to the SC06 supercomputing conference, giving scientists a familiar, high-level way to program graphics chips for general computation. That simple idea, letting researchers write normal-looking code that ran on GPUs, extended accelerated computing from graphics into simulation and science. Twenty years later, Nvidia says more than 11 million developers build on CUDA, with more than 1,000 CUDA-X libraries on top.
The examples Nvidia picks are well chosen. At the University of Illinois, Klaus Schulten’s team connected GPU computing to molecular biology with NAMD and VMD, simulation and visualization tools. In 2013, researchers at Illinois, Pittsburgh and Vanderbilt used CUDA-accelerated tools on the Blue Waters supercomputer to help resolve the atomic structure of the HIV capsid. In 2012, Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton trained AlexNet on two Nvidia GPUs and won ImageNet, which helped kick off the deep learning era. And AlphaFold 2, whose creators Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry, uses CUDA for GPU execution.
The moat, stated plainly
What the post doesn’t say outright is the business lesson. CUDA is why Nvidia is hard to replace. Rivals can build fast chips, but two decades of libraries, tools, papers, courses and habits sit on top of CUDA. Every scientist who learned GPU computing on it, and every lab whose code depends on it, raises the cost of switching. Our coverage of SC26 shows Nvidia still extending that stack, from CUDA-X libraries packaged for AI agents to the Rubin HPC GPU for science and new FP64 emulation in cuBLAS.
That moat has costs. Science built on a proprietary platform depends on one company’s pricing, licensing and roadmap. Open alternatives and compiler layers that target several vendors exist precisely because researchers worry about that lock-in. A fair anniversary assessment credits CUDA for enabling Nobel-level work and also admits that the field is more dependent on a single vendor than is healthy.
An anniversary in an awkward month
The timing is notable. CUDA turns 20 just days after Jensen Huang, who championed the platform for two decades, resigned as Nvidia’s CEO. Nvidia’s post focuses on researchers and developers, not executives, and that focus is the right one. CUDA’s strength was never one person’s charisma. It’s millions of developers and decades of accumulated code. That’s also the best argument that Nvidia’s position will survive its leadership turmoil, at least in the near term.
Who it matters to
- Scientists
- CUDA remains the default path to GPU acceleration, with deep libraries and a huge community.
- Rival chipmakers
- Matching CUDA’s software ecosystem, not raw hardware speed, remains the main barrier.
- Students and new developers
- CUDA skills stay highly marketable, though portable alternatives are worth learning too.
- Nvidia investors
- The platform’s institutional strength offsets some leadership risk.
What to watch
- Growth in Nvidia’s reported CUDA developer count beyond 11 million.
- Adoption of vendor-neutral GPU programming layers in scientific computing.
- New CUDA-X libraries aimed at agent-driven science.
- Whether new Nvidia leadership changes CUDA licensing or openness.
FAQ
When was CUDA introduced?
Nvidia announced CUDA in November 2006 and brought it to the scientific community at SC06.
How many developers use CUDA?
Nvidia says more than 11 million developers build on CUDA.
What breakthroughs relied on CUDA?
Nvidia cites AlexNet’s 2012 ImageNet win, the 2013 HIV capsid structure work and AlphaFold 2, among others.
Companies: NVIDIA
More on NVIDIA
AI · Nov 5, 2026
NVIDIA’s four open Cosmos 3 Nano models let robot and driving teams simulate worlds that reactAI · Nov 16, 2026
Rubin HPC is Nvidia’s bet that science agents will be bottlenecked by simulation, not reasoningAI · Nov 9, 2026
Nvidia’s Korean startup bets span interconnect, open models, robot hands and document AIAI · Nov 8, 2026
Nvidia’s Dynamo recipes speed four Korean open models to production, tuned for B200s