A leading global technology company needed an analysis of NVIDIA’s patent portfolio to determine where there was still space for independent innovation. The client wanted to know where NVIDIA’s most consequential researchers were taking the platform next.
NVIDIA controls approximately 80% of the AI training chip market. Hardware explains part of that position. The rest sits in CUDA, NVIDIA’s software platform. CUDA has become the standard programming environment for AI development, with more than four million developers building on it.
NVIDIA also holds over 17,000 global patents, with AI and machine learning as the dominant focus. For companies operating in AI computing, tracking where NVIDIA is building and where it is heading is a competitive requirement.
Patent counts could not show where CUDA was defended, sparse, or still contested
KEY REFRAME
A portfolio count tells you how large a competitor’s position is. It does not tell you where the position is densest, where it is sparse, which sub-domains are still contested, or what the company’s most significant inventors are working on next. The study had to read NVIDIA’s CUDA portfolio across four distinct lenses simultaneously.
Counting NVIDIA’s patents shows how large its position is. It cannot show where that position is densest, where gaps remain, which sub-domains are still contested, or where the company’s most significant inventors are heading.
GreyB structured the study around four questions: how NVIDIA builds and defends its CUDA position in AI computing; which sub-domains are already crowded; where whitespace remains; and what the key inventors driving CUDA’s evolution are working on next. .
5 analysis layers separated CUDA AI patents from unrelated chip filings
The study covered NVIDIA’s CUDA patent portfolio from 2022 onwards, global filings both granted and pending, scoped to software technologies with AI-related applications.
| Layer | Question Asked | Decision It Addressed |
| L1 | How should the CUDA domain be structured and bounded for this analysis? | GreyB built the taxonomy before collecting data. Sub-domains (Programming Model, Hardware Abstraction, Compilation, and others) were defined with explicit inclusion and exclusion criteria drawn from NVIDIA’s own whitepapers and technical documents. Without those boundaries, CUDA AI-specific patents mix with unrelated chip engineering filings. |
| L2 | How should search logic be constructed to capture relevant filings without over-inclusion? | GreyB combined domain-specific keywords, IPC and CPC codes, and technical terminology from NVIDIA’s literature, then tested and refined them against sample patent sets. Terms such as “Occupancy” required careful scoping; without tighter criteria, searches return unrelated hardware patents. The team treated scope alignment with the client as an ongoing checkpoint throughout the analysis, not a one-time sign-off. |
| L3 | Which patents in the initial set are genuinely in scope after human review? | A manual screening pass removed false positives. Reviewers examined each patent’s title, abstract, and claims to catch filings that appeared relevant but had no AI-specific application within the CUDA stack. Running AI-assisted categorisation without this step produces unreliable results. |
| L4 | How should patents be assigned to sub-nodes and relevance-scored at scale? | GreyB’s AI auto-categorisation tool assigned patents to sub-nodes and scored relevance. A second agentic AI layer evaluated specific innovation focus areas including Quantized Fusion and Speculative Decoding. The tool flagged borderline patents for human review. By this point, relevance accuracy exceeded 80% before detailed analysis began. |
| L5 | What do the portfolio, technology, and bibliographic lenses reveal about NVIDIA’s strategy and trajectory? | Three parallel reads of the same dataset: an IP portfolio view covering filing trends, geographic strategy, citation quality, and foundational patent indicators; a technology view tracing sub-domain evolution, emerging areas, and innovation rate changes; and a bibliographic view identifying key inventors, tracking how their focus had shifted, and surfacing collaboration and acquisition signals. |
Sub-domain mapping showed where R&D should build, design around, or wait
The portfolio view produced a sub-domain map distinguishing crowded areas, where filing density was high and the competitive space was already defended, from sparse areas where meaningful room remained.
This distinction is not visible from a total patent count. It is the foundation for any investment decision about where to direct original R&D versus where to design around existing IP.
The technology view traced how specific CUDA sub-domains had evolved year by year, identified where innovation rates were accelerating or slowing, and surfaced technological gaps not yet addressed by any player’s filings.
Filing spikes revealed product-linked, long-horizon, and external capability signals
The analysis plotted NVIDIA’s filings within each sub-domain against the company’s product and market activity over the same period.
That mapping identified which patent clusters tied to commercial launches and which represented longer-horizon R&D not yet reflected in product announcements.
Inventor-driven filing spikes and sudden volume surges each carried a distinct signal, pointing to different kinds of priority shifts at NVIDIA. Collaboration-linked patents pointed to a third: capabilities being built outside that would not appear in internal filings for years.
 Inventor Analysis: Where the Consequential Work Is Concentrated
KEY FINDING
Within NVIDIA’s CUDA portfolio, the innovations with the highest strategic significance concentrate around a small group of senior researchers. Their filing rates are unremarkable, however, the coherence and direction of what they file is what matters.Â
| Inventor | Research Focus | Strategic Signal |
| Jan KautzVP, Learning and Perception Research | Physics-guided diffusion models and multimodal large language model agents | His research pushes CUDA to evolve at the compilation and inference layer, requiring new optimised processing pipelines to support real-time AI performance. NVIDIA is building the software stack for the next generation of AI workloads. |
| Ronny KrashinskyDistinguished Engineer | New thread hierarchy models and distributed memory architectures | His work expands the CUDA programming model structurally, enabling larger-scale parallelism and more efficient memory coordination across AI workloads. It signals foundational changes in how the platform handles scale. |
The bibliographic view also tracked collaboration patterns. Where NVIDIA had active external partnerships, with universities or acquired companies, those signals pointed to capabilities that would not appear in internal filings for several years.
Three portfolio reads turned patent data into competitive directionÂ
| Lens | What It Examined | What It Produced |
| IP Portfolio View | Filing trends, geographic protection strategy, citation quality, foundational patent indicators | Sub-domain density map distinguishing hotspots from whitespace; geographic enforcement priorities |
| Technology View | Sub-domain evolution year by year, emerging areas, innovation rate changes, technological gaps | Technology roadmap with trigger analysis connecting filing spikes to product and market activity |
| Bibliographic View | Key inventors, focus shifts over time, collaboration and acquisition signals | Forward-looking intelligence on where NVIDIA’s most consequential researchers are working now and next |
Want to decode a competitor’s patent strategy before investing in R&D?
Understanding a dominant player’s IP strategy requires more than counting its patents. The number tells you the scale of the position. It does not tell you where that position is densest, where it remains open, or where the company’s most significant researchers are taking it next.
GreyB ran the same portfolio through multiple lenses. Filing trends showed where NVIDIA was building momentum and where it had pulled back. Sub-domain density mapped open versus defended territory. Inventor trajectories separated genuine research investment from portfolio maintenance. Technology roadmap analysis connected past filing patterns to future product direction.
If your R&D decisions sit in a space where one player controls both the hardware architecture and the software programming model, skipping that level of analysis means making investment decisions with half the picture.
We help technology strategy and R&D teams decode competitor IP portfolios to support investment decisions, white space analysis, and competitive positioning. Contact our specialists to discuss how this framework applies to your specific competitive landscape.
Schedule a Consultation With Our Experts Today
Get In Touch​
