GTC 2026 DLSS 5 and 20 years of CUDA: the install-base flywheel

Updated 2026-10-01 · Video: NVIDIA, published 2026-03-16

This segment runs from 6:02 to 16:15 of the GTC 2026 keynote. Jensen Huang marks CUDA’s 20th anniversary, explains the install-base flywheel he says sums up NVIDIA’s whole strategy, traces the path from GeForce to RTX, and unveils DLSS 5, a neural renderer that blends traditional 3D graphics with generative AI. Watch it if you want the strategic foundation the rest of the keynote builds on, or if you build or play games on RTX hardware.

Key takeaways

1. CUDA turns 20. Jensen frames the anniversary around SIMT programming and the new Tiles feature for programming Tensor Cores (6:02).
2. The install base is the whole strategy. One slide, used since the beginning, shows install base attracting developers, developers creating breakthroughs, and breakthroughs creating markets that grow the base (7:04).
3. Old GPUs stay valuable. Jensen claims cloud rental prices for six-year-old Ampere GPUs are still rising, which he attributes to continuous software optimization (9:02).
4. GeForce was the on-ramp. Programmable shaders came 25 years ago and CUDA 20 years ago, distributed to millions of PCs through GeForce (11:21).
5. DLSS 5 fuses structured 3D data with generative AI. Jensen presents it as a template other industries will follow (13:47).

Chapter notes

6:02 – 7:04 Twenty years of CUDA

Jensen opens with CUDA’s 20th anniversary and credits the SIMT model, single instruction, multiple threads, for letting developers write code that looks single-threaded and scale it across thousands of threads. He contrasts this with SIMD, which is harder to program. He then mentions Tiles, a recent addition that helps developers program Tensor Cores directly.

Tiles matter more than the brief mention suggests. Most AI math runs on Tensor Cores, and until recently developers reached them mainly through libraries such as cuBLAS or cuDNN. A tile-level programming model lets more developers write custom kernels for new model architectures without dropping to low-level code, which keeps CUDA relevant as models change faster than libraries.

7:04 – 10:13 The install-base flywheel

This is the most important strategic passage of the keynote, and it lasts three minutes. Jensen shows a slide he says he has used from the start and calls it 100% of NVIDIA’s strategy. Install base attracts developers. Developers create breakthroughs such as deep learning. Breakthroughs create new markets. New markets bring more companies and grow the install base.

Jensen says the hardest part was the install base itself, which took 20 years and now numbers hundreds of millions of CUDA GPUs in every cloud and computer maker. That figure is his stage estimate; the official GTC 2026 releases do not give a current count.

He draws two consequences. First, because NVIDIA supports every architecture-compatible GPU, each software optimization benefits millions of users, so the cost of computing keeps falling after purchase. Second, accelerated infrastructure has a long, valuable life. As evidence he says Ampere GPUs from about six years ago still see cloud rental prices going up. We could not verify that pricing claim in NVIDIA’s releases, but it is the argument investors should note: NVIDIA is countering the worry that GPUs depreciate quickly by pointing to software-driven longevity.

10:13 – 13:00 From GeForce to RTX

Jensen calls GeForce NVIDIA’s most effective marketing campaign, joking that parents paid for a generation of future computer scientists to become NVIDIA users. The timeline he gives matches the DLSS 5 release: programmable shaders 25 years ago, CUDA five years later, and RTX with hardware ray tracing in 2018.

He says CUDA consumed most of NVIDIA’s profits when it launched and that NVIDIA kept investing through 13 generations over 20 years; the 13-generation count is from the stage and not the releases. GeForce then put CUDA in front of the researchers who used GPUs for deep learning, which he credits with starting the AI boom. The loop closes here: graphics brought AI into the world, and AI now comes back to change graphics.

13:00 – 15:08 RTX and the DLSS 5 reveal

Jensen recalls that RTX added two ideas: hardware ray tracing and the bet that AI would reshape computer graphics. He then introduces neural rendering, which he describes as fusing 3D graphics with AI, and plays the DLSS 5 demo reel.

The official release supplies the details the stage omitted. DLSS 5 takes each frame’s color and motion vectors as input and uses an AI model to add photoreal lighting and materials anchored to the source 3D scene and consistent across frames. It runs in real time at up to 4K. The model infers scene content such as skin, hair and fabric, plus lighting conditions, from a single frame, then renders effects like subsurface scattering on skin. Developers get controls for intensity, color grading and masking, and it integrates through the same Streamline framework as existing DLSS. NVIDIA says it arrives in fall 2026, with support from Bethesda, CAPCOM, Ubisoft, Tencent, NetEase, Warner Bros. Games and others.

15:08 – 16:15 Structured data plus generative AI

After the reel, Jensen explains the idea in general terms. Traditional 3D graphics is fully predictable structured data: the ground truth of a virtual world. Generative AI is probabilistic but very realistic. DLSS 5 uses the structured data to control the generative model, so the output is both beautiful and precise.

He then says this pattern, structured data as the trustworthy anchor for generative AI, will repeat across industries. That line is the bridge to the next segment of the keynote on accelerating structured and unstructured data with cuDF and cuVS. Read DLSS 5 here as more than a gaming feature. It is NVIDIA’s showcase example of grounding generative output in hard constraints, the same problem enterprises face when they want AI that respects their databases.

What changed since GTC 2025

GTC 2025 spent little keynote time on consumer graphics. NVIDIA’s big graphics launch that year came earlier, at CES in January 2025: the GeForce RTX 50 Series, DLSS 4 with Multi Frame Generation that produces up to three AI frames per rendered frame, and RTX Neural Shaders. At GTC 2025 itself, NVIDIA’s live blog quotes Jensen saying the CUDA install base was “now everywhere”, and the CUDA content centered on CUDA-X libraries and open-sourcing cuOpt.

The shift in 2026 is from performance to appearance:

• DLSS 4 and 4.5 made frames faster; DLSS 4.5, launched at CES 2026, uses AI to draw 23 of every 24 on-screen pixels, according to the DLSS 5 release.
• DLSS 5 makes frames look different. It generates lighting and material detail rather than only extra pixels or frames. NVIDIA calls it its biggest graphics advance since real-time ray tracing in 2018.
• Neural shaders became neural rendering. The 2025 neural shaders put small networks inside the shading pipeline; DLSS 5 applies one model to the whole finished frame.
• The CUDA story moved from “everywhere” to “economics”. GTC 2026 adds the argument that the install base makes old GPUs keep their value, which supports NVIDIA’s token economics pitch.

GTC 2025 had no DLSS 5 promise to keep, so this is a new commitment with a fall 2026 date.

Skip list

• 13:47 – 14:59: the DLSS 5 demo reel. Watch it once for the visuals; it has no narration.
• 10:40 – 11:10: the GeForce joke about parents paying for their kids’ graphics cards.

Glossary

• CUDA — NVIDIA’s parallel computing platform and programming model for GPUs, launched in 2006.
• SIMT — Single instruction, multiple threads: CUDA’s model, where code written per thread runs across many threads in lockstep.
• Tensor Core — Specialized GPU units for the matrix math that dominates AI workloads.
• Install base — The number of deployed GPUs that can run CUDA software, which Jensen treats as the engine of NVIDIA’s flywheel.
• DLSS — Deep Learning Super Sampling, NVIDIA’s AI rendering technology, first released in 2018 for upscaling and frame generation.
• Neural rendering — Using a neural network to produce or transform rendered images, guided by the game’s 3D data.

The keynote’s next topic, data processing with cuDF and cuVS, is summarized in the full keynote notes, and the event overview is on the GTC 2026 index.