In 2024, on-device AI shifted from a product differentiator to a baseline expectation.
As major consumer electronics brands began embedding real-time, privacy-first AI into their flagship products, every player in the market received the same signal: close the gap or fall behind.
Edge AI patent filings surged significantly in 2024. Companies were claiming territory in model efficiency, continuous learning, and agent capabilities faster than most R&D teams could track.
A consumer electronics company needed clear answers to three questions:
- Where was the real competition happening across key on-device AI domains?
- Which primary competitors were building something serious?
- Where did genuine opportunity still exist?
GreyB helped the client explore the on-device AI patent landscape across multiple competitors, several domains, and a multi-year window of worldwide filings. The goal was not to count patents. It was to identify where competitive pressure was real, where it was fading, and where the client still had room to act.
Patent Counts Could Not Show Which Competitors Were Still Building
A patent count shows who has filed. The questions that matter for investment decisions, such as the depth of a competitor’s commitment, where pressure is still active, and which domains have room, do not appear in a count.
The study read each portfolio as a story.
Not all filing activity signals the same intent. A competitor can file heavily in a domain for two years and stop, consistent with a shift from active building to consolidation. Another can be quiet for years and then resume, signalling renewed interest in a domain it had previously set aside.
A third can hold dozens of patents in a domain, all of them originating from a single acquired company. The IP came through acquisition.
None of these distinctions appears in a headcount.
Five Decision Layers Converted Filing Activity Into Roadmap Intelligence
The analysis operated in five sequential layers. Each layer answered a practical R&D or strategy question and reduced the risk of drawing the wrong conclusion from raw patent volume.
| Layer | Question Asked | Decision It Addressed |
| L1 | What falls within scope, and what does not? | GreyB locked the boundary before collecting data. Only patents with explicit edge-device or on-device application qualified. Cloud-based architectures, general quantization methods without an edge use case, and agent systems built exclusively for automotive were excluded. A loose scope would have flooded the dataset and buried the signal. |
| L2 | How should the search framework be constructed to capture relevant filings without over-inclusion? | GreyB combined keywords, classification codes, and domain-specific concepts, then tested and refined them across multiple rounds. Sample patent sets went to the client team at several points to pressure-test scope alignment. The team treated this as an ongoing checkpoint throughout the analysis. |
| L3 | What do filing trends, grant status, and geographic coverage reveal about each competitor’s position within each domain? | This layer established the competitive baseline. Year-by-year filing trends showed momentum and retreat. Grant status by domain marked where the field had settled and where room remained. Geographic coverage and lapsed protections pointed to where competitors had reassessed commercial relevance. |
| L4 | Who is behind the filings, and what do inventor profiles and collaboration patterns reveal about strategic intent? | This layer distinguished genuine research investment from filing activity maintained to preserve competitive presence. Inventor background, prior work, and external university partnerships gave a read on commitment depth that filing volume alone could not. |
| L5 | How do all four companies compare at the technique level within each domain, and what does this mean for the client’s own roadmap? | This layer produced technique-level scoring across all three domains with year-by-year evolution analysis and direct strategic recommendations for the client’s R&D priorities. |
Filing Trends Showed Where Pressure Was Real and Where It Was Fading
Year-by-year filing analysis showed that some competitors had filed heavily in specific domains and then stopped, consistent with a shift to portfolio consolidation.
Others had resumed filing in a domain after a multi-year gap, signalling renewed strategic interest.
Neither dynamic appears in a total patent count.
For the client, this helped separate active competitive pressure from historical filing activity. It also showed that the on-device AI race was not equally intense across every domain. Some areas were still active. Others had already moved into consolidation.
Grant Status Separated Established Territory From Open Windows
Across the domains studied, grant rates varied significantly.
Some domains had a high proportion of already-granted patents, indicating established and crowded territory. Others still had most filings in pending status, meaning the competitive shape was not yet fixed.
This distinction matters directly to investment timing.
In a domain where most relevant patents are already granted, a new entrant is building on established ground. In a domain where most filings are still pending, the competitive shape is not yet fixed.
The analysis showed that AI Agents were largely granted, while Quantization and Incremental Learning were mostly pending. This helped the client understand where differentiation would be harder and where the window for technical positioning was still open.
Inventor Profiles Revealed the Difference Between Research Depth and Visibility Filing
Across several competitors, a small group of researchers accounted for the majority of filings.
The analysis identified three recurring patterns:
- Dedicated internal research: In some cases, lead inventors were core members of long-standing internal AI labs with a focused body of work over several years. The IP was built organically.
- Deployment-experienced inventors: In other cases, lead inventors had production deployment experience, bringing practical system-level knowledge into their patent work.
- Acquisition-sourced IP: In at least one instance, a competitor’s entire patent position in a domain originated from researchers at an acquired company. None of the IP was built in-house.
The origin of IP changed the competitive read.
A portfolio built by a dedicated lab over several years carries a different risk profile than one assembled through acquisition. Filing volume does not indicate depth of commitment. Inventor origin does.
Collaboration Patterns Signaled What Competitors Could Build Next
Some competitors had active university partnerships on problems their internal teams were not handling alone, including optimizing neural network efficiency for resource-constrained devices and advancing multi-agent system architectures.
Others worked entirely in-house.
University collaborations signal capability development that will not appear in commercial filings for two to three years. For the client, this provided an early view of where competitors may be building the next wave of on-device AI capabilities before those capabilities appear in products or major commercial filings.
Five Intelligence Dimensions Turned the Landscape Into Strategic Direction
| Dimension | Signal | Strategic Implication |
| Filing Trends | Two competitors consolidating; one re-entering after a gap | Active competitive pressure is concentrated in fewer areas than total counts suggest |
| Grant Status | AI Agents largely granted; Quantization and Incremental Learning mostly pending | Timing of entry matters: the window for differentiation is domain-specific |
| Inventor Profiles | Varies from dedicated lab research to acquired IP to deployment-experienced inventors | Filing volume does not indicate depth of commitment; inventor origin does |
| Geographic Coverage | US dominant; some lapsed protections in secondary markets | Lapsed protection signals reduced commercial confidence in specific geographies |
| Collaborations | Some competitors have active university partnerships; others fully self-contained | University activity is a two-to-three-year leading indicator of emerging capability |
Patent Landscapes Need to Be Read as Competitive Stories, Not Filing Counts
The competitive landscape in on-device AI cannot be read from patent counts.
A headcount answers one question: who has filed the most.
It leaves unanswered the questions that matter for investment decisions: who has built a genuine research capability, which domains are still open, and where a new filing would land in already-claimed territory.
Reading a portfolio as a story is what produces actionable intelligence.
Filing trends show where momentum is building and where it has peaked. Inventor profiles measure commitment depth. Grant status marks which races are already over. Geographic coverage shows where a competitor will enforce. Collaboration patterns signal what is two to three years away.
What AI whitespaces could become your next advantage?
Our AI patent landscape shows that the strongest activity is concentrated in computational intelligence and computer architecture. AI infrastructure, chips, model efficiency, and scalable systems are now key areas of competition. At the same time, lower-density fields such as healthcare, automation, and speech processing may offer greater white-space opportunities.
A focused patent landscape analysis can help R&D, IP, and strategy teams move beyond patent noise and understand where their next filing, partnership, or product roadmap decision will land. GreyB can help you move from broad research and early-stage signals to validated, commercially useful direction for roadmap planning, investment decisions, and competitive positioning.
Schedule a Consultation With Our Experts Today
Get In Touch​
