Nvidia marked the 20th anniversary of CUDA at the SC26 supercomputing conference this week, announcing that more than 11 million developers now build on the platform.
The milestone arrives at a crossroads for the company. CUDA arrived in November 2006 at SC06. It gave scientists a familiar, high-level programming model for GPUs and extended accelerated computing beyond graphics into simulation and molecular biology.
Eventually it reached deep learning. Today it connects Nvidia GPUs with more than 1,000 CUDA-X libraries spanning drug discovery, finance, autonomous vehicles and quantum computing.
Nvidia’s anniversary narrative traces a chain of breakthroughs, each built on the previous generation’s CUDA work. In 2012, University of Toronto researchers Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton trained AlexNet on two Nvidia GPUs. They won the ImageNet competition and demonstrated GPU computing for deep learning.
In 2013, teams at the University of Illinois, the University of Pittsburgh and Vanderbilt University combined experiments with CUDA-accelerated simulations on the Blue Waters supercomputer. They resolved the atomic structure of the HIV capsid.
In 2024, Demis Hassabis and John Jumper shared the Nobel Prize in Chemistry for AlphaFold 2. Its publicly released software uses CUDA for GPU execution. Nvidia also notes CUDA’s role in accelerating scientific applications on supercomputers and in visual effects for Oscar-winning films.

Huang co-founded Nvidia in 1993 and steered its shift from gaming graphics to AI infrastructure. His departure, announced November 12, left Colette Kress, the chief financial officer, as interim CEO while the board searches for a permanent successor.
At SC26 in Chicago, Nvidia launched a desk-size cluster of four DGX Spark computers linked directly without a network switch. The system carries 512GB of memory and runs open AI models as large as DeepSeek V4.1 with up to 700 billion parameters.
🚨 NVIDIA CUDA 13.1 brings the largest update to the #CUDA platform in 20 years. Discover the new features and updates for improving performance and driving #acceleratedcomputing, including: ✅ Announcing CUDA Tile: Future-proof kernels by abstracting Tensor Cores. ✅… pic.twitter.com/srswJEVqrt
Forbes reported that a University of California, Berkeley lab used one DGX Spark to predict bacterial traits from genome sequences. The run cut an 11-hour cluster queue wait and delivered results 15 times sooner.
The kit is sold through Dell, HP, Lenovo and Nvidia’s online store. No price was disclosed. Jon Peddie Research said Nvidia introduced cuPotential for molecular simulations and cuSpa for spatial biology calculations at the conference. The firm tied them to the Rubin HPC GPU expected in 2028.
Jon Peddie Research emphasized that the 11 million developer count and 1,000-plus library count are Nvidia’s own figures with no disclosed methodology. The firm said the case studies come from university and laboratory researchers without commercial deployment data.
.jpg)
The firm recommended that CIOs map which code paths call CUDA-X libraries, which use portable frameworks, and which use custom kernels. That map determines the cost of adding a second GPU source. A small pilot on a second software stack would price that effort with real data. ISVs, it said, should inventory their CUDA dependence now and identify which libraries bind them to one supplier.
“Today, more than 11 million developers build on CUDA across the computing stack,” Nvidia said in its anniversary blog post.