Beyond the Chatbot: What Patent Data Reveals About the Next Phase of GenAI

Public discussion often presents generative artificial intelligence (GenAI) as a race among a small group of highly visible models and chatbot providers. Yet the technologies likely to shape GenAI’s future extend well beyond the well-known chat interface. Computing infrastructure, electricity, industrial adoption, skills, international diffusion, trust, and intellectual property (IP) will all influence who can develop the technology, where it can be deployed, and how its economic value is distributed.

The latest WIPO Technology SPARK Report: Patent Trends Update in GenAI provides an early view of this broader transition. Published GenAI patent families rose from 18,862 in 2024 to 37,808 in 2025, with more than 56,000 published over those two years, exceeding the total of the preceding decade. GenAI’s share of all AI patenting also increased from 6.1% in 2023 to 8.7% in 2025 ^[1].

Patent publications do not offer a real-time ranking of technological leadership. Patent applications typically become public after a delay, and patent counts alone do not establish quality, commercial success, or market share. Their analytical value lies in recording organizational decisions to seek protection for technical solutions. Consequently, the latest data provides a delayed but revealing perspective on how the surge in research and investment following the emergence of widely used large language models (LLMs) is being translated into longer-term innovation strategies.

The GenAI race is becoming several interconnected races

The technological shift is evident in patent data. LLMs have overtaken generative adversarial networks (GANs) as the largest GenAI model category. LLM-related patent families increased from 881 in 2023 to more than 14,100 in 2025, while diffusion models and multimodal systems also expanded rapidly. Reasoning-oriented models and agentic systems are beginning to appear in filings as potential future directions. This rapid progress is also reflected in the Stanford AI Index 2026. According to the report, industry developed more than 90% of the notable frontier models released in 2025, while AI adoption among organizations rose to 88%. The report also discusses how closely matched the United States and China have become, with models from both countries repeatedly overtaking one another and only a small performance gap remaining by March 2026 ^[2]. In combination, the trends in these two reports suggest that AI models are continuing to improve quickly, but that it remains a field that is highly dynamic and somewhat unsettled.

As a result, durable competitive advantage may depend on more than producing a leading model at a given moment. It may also depend on access to advanced chips, cloud platforms, data centers, energy, distribution channels, specialist talent, proprietary data, and sector-specific IP. The OECD’s analysis similarly conceptualizes AI as a multilayered value chain in which control over infrastructure and complementary assets can shape competitive dynamics ^[3].

Infrastructure is moving to the center of the picture

The changing composition of leading patent applicants is particularly informative. Alongside leading internet and software companies, the ranking now includes SoftBank, Nvidia, State Grid Corporation of China, China Southern Power Grid, Inspur, and Bosch. These organizations operate across telecommunications, semiconductors, servers, electricity systems, and industrial engineering. This broader applicant base complements the OECD’s 2026 analysis of AI markets. The OECD describes an AI value chain extending from specialized chips and cloud infrastructure to models, applications, and distribution. It identifies high fixed costs, economies of scale, vertical integration, and control of complementary infrastructure, including data centers, electricity generation, and grid connections, as potential sources of market power. In 2023, the three largest cloud providers accounted for 74% of the global cloud market ^[3].

Energy is therefore not a peripheral issue. The International Energy Agency (IEA) estimates that electricity consumption by data centers could rise from 485 terawatt-hours (TWh) in 2025 to approximately 950 TWh in 2030, with consumption by AI-focused data centers tripling over that period. It also notes that reasoning, video-generation, and agentic tasks can require hundreds or thousands of times more energy per query than simple text generation, even as hardware and software efficiency improves ^[4].

WIPO’s patent data provides a complementary perspective on this relationship. State Grid Corporation of China and China Southern Power Grid are among the leading GenAI patent applicants, indicating that electricity providers are not only responding to the additional demand created by AI but are also developing GenAI applications for grid optimization, predictive maintenance, equipment monitoring, and infrastructure planning. AI may therefore place growing pressure on electricity systems while simultaneously offering tools to improve their efficiency and resilience. The next phase of the GenAI race may depend partly on how effectively these digital and energy systems develop together.

Deployment may matter as much as frontier development

Producing a frontier model and deriving economic value from GenAI are distinct challenges. The entry of utilities, industrial firms, and infrastructure providers into the patent landscape indicates that innovation is diffusing into established sectors. Growth in patenting related to software and code, 3D modelling, and life-science applications points in the same direction: toward systems designed for specific technical tasks, operational environments, and industry needs.

The World Bank’s concept of “Small AI” offers a useful perspective through which to understand this transition. Small AI does not refer only to models with fewer parameters. It describes an approach to deployment that is affordable, accessible, resource-efficient, and adapted to a clearly defined context. Such systems may use smaller datasets, operate on smartphones or laptops, function with limited or intermittent connectivity, and be tailored to specific languages, institutions, or operational needs. Rather than attempting to perform a broad range of general-purpose tasks, they are designed to solve a bounded problem in the environment where the solution will actually be used ^[5].

This distinction is particularly relevant to this blog’s “beyond the chatbot” title. A general-purpose chatbot is designed to serve many users and tasks through a largely standardized interface. Small AI applications are more likely to disappear into existing products and processes. They may help a farmer identify a crop disease from a photograph without a continuous internet connection, provide localized advice through basic mobile services, or deliver tutoring through platforms already used by students. Their significance lies less in presenting a new public-facing model than in embedding GenAI capability into a specific workflow ^[5].

Small AI should not, however, be understood as entirely separate from frontier AI. Some applications may rely on compact, task-specific models developed locally; others may adapt, fine-tune, or distil larger open or proprietary models. Frontier advances can therefore lower the cost and improve the capabilities of downstream systems, while the application layer determines whether those advances become useful in particular sectors and communities. The two forms of innovation may be complementary: large models provide increasingly capable foundations, while smaller and more specialized systems translate those capabilities into locally relevant services.

The World Bank’s four foundations of AI readiness, connectivity, compute, context, and competency, help explain why the ability to make this translation remains uneven. Small AI can reduce dependence on hyperscale computing and constant broadband access, but it does not eliminate the need for reliable electricity, suitable devices, relevant data, technical skills, institutional capacity, and mechanisms for maintaining and governing deployed systems. Its advantage is therefore not that it removes all infrastructure constraints, but that it may lower the threshold for meaningful participation ^[5].

WIPO’s patent data does not directly identify which inventions qualify as Small AI, and application-specific patents should not automatically be treated as evidence of lightweight or locally deployed systems. Nevertheless, the growing activity of utilities, industrial companies, and infrastructure providers is consistent with a broader movement from general-purpose model development toward technical adaptation and deployment. Patent information can help reveal where organizations are seeking protection for the systems, interfaces, workflows, and sector-specific applications through which GenAI may ultimately create value.

This may also broaden the range of countries and firms able to participate in the GenAI economy. Not every economy needs to train a frontier model to develop useful GenAI applications. Some may specialize in adapting existing models to local languages, public services, industrial processes, or resource-constrained environments. Only around 9% of GenAI patent families published in 2025 had appeared in at least two jurisdictions, compared with approximately 15% of patent families worldwide across all technologies. The difference should however be interpreted with some caution, since many GenAI patent portfolios are recent and may not yet have completed the process of seeking protection in additional jurisdictions ^[6].

The longer-term influence of GenAI may therefore be determined not only by a small number of highly visible models, but by a much larger and less visible layer of specialized systems embedded in agriculture, health, education, energy, manufacturing, and public services. The chatbot may remain the most recognizable expression of GenAI, while Small AI illustrates how its capabilities could become more distributed, contextual, and operational.

From patent diffusion to wider economic adoption

WIPO’s patent data indicates that GenAI innovation is expanding beyond frontier models and consumer-facing tools into software development, infrastructure, industrial systems, and other specialized applications. However, this expansion remains geographically concentrated. This distinction is relevant to the World Trade Organizaiton’s (WTO) analysis of AI and trade. AI could lower trade costs through translation, logistics, regulatory compliance, and access to information, while trade can provide access to semiconductors, computing infrastructure, digital services, and other technologies needed to develop and deploy AI. The WTO estimates that global trade could be 34–37% higher by 2040 under different scenarios but emphasizes that the distribution of these gains will depend on infrastructure, skills, technological access, and policy conditions ^[7]. Read alongside the patent findings, this suggests that the international diffusion of GenAI will depend not only on where inventions originate, but also on whether firms and economies can access, adapt, commercialize, and protect them across markets.

The growing application of GenAI across software, utilities, industrial firms, and other sectors also provides a link to the changing organization of work. The patent data does not measure employment effects, but it helps identify the technical fields and industries in which new systems are being developed. The International Labor Organization (ILO) estimates that one in four workers is employed in an occupation with some exposure to GenAI and concludes that task and job transformation is generally more likely than complete replacement ^[8]. As GenAI inventions move into operational settings, their effects will therefore depend on how organizations redesign workflows, allocate responsibilities between humans and automated systems, and develop the skills required to use them effectively.

The same movement from experimentation to sector-specific deployment makes trust increasingly important. The International AI Safety Report 2026 finds that general-purpose AI systems can perform strongly on complex tasks while continuing to exhibit factual, reasoning, cybersecurity, and reliability weaknesses ^[9]. These limitations may become more consequential as GenAI applications move into infrastructure, healthcare, industrial processes, and other settings where failures can have significant effects. Although WIPO’s patent report does not evaluate the safety or reliability of individual inventions, the widening range of applicants and applications suggests that risk management, oversight, and dependable performance will increasingly influence which patented technologies achieve adoption.

Together, these perspectives extend the patent story from invention to diffusion. Patent data provides early evidence of where organizations are developing and seeking to protect GenAI technologies; trade conditions, workforce adaptation, and trust will help determine how widely those technologies are subsequently deployed and who benefits from them.

GenAI patent portfolios can serve different strategies

The comparison between SoftBank and OpenAI illustrates why patent volume should be interpreted as one indicator among several, rather than as a direct measure of GenAI leadership. SoftBank has recently become the largest GenAI patent holder in the WIPO report, with 2,985 patent families, virtually all arising from applications filed in 2023 and published in 2025 ^[1]. This concentrated wave of patent filings coincides with SoftBank’s wider ambition to become an AI platform provider spanning models, semiconductors, data centers, power infrastructure, and robotics ^[10]. Although patent data cannot establish the purpose or value of individual filings, the breadth of the portfolio appears consistent with a strategy of building positions across several layers of the AI value chain.

OpenAI presents a different pattern. The WIPO report identified only 35 patent filings globally as of late 2025, focused mainly on product-level developments such as multimodal interfaces, code generation, image generation, and text editing rather than foundational model architectures. OpenAI has also publicly pledged to use its patents defensively ^[11]. Its smaller disclosed portfolio may therefore indicate a more selective role for patents, complementing other sources of competitive advantage such as trade secrets, rapid product development, proprietary systems, and access to computing infrastructure. However, the relative importance of these mechanisms cannot be determined from public patent information alone.

The comparison is not simply between a company that patents extensively and one that does not. It shows that firms operating in the same field may use patents for different strategic purposes. SoftBank’s portfolio appears aligned with a broad platform and infrastructure strategy, while OpenAI’s portfolio and defensive patent pledge point toward more selective protection of particular products and implementations. Neither approach, by itself, establishes which company holds the stronger technological or commercial position.

Patent leadership is one indicator, not the final score

The WIPO report does not predict a single winner, and patent data alone cannot determine future outcomes. However, it illuminates the structural dynamics emerging beneath the visible model competition. GenAI patenting is expanding rapidly across both model types and application modes, with strong growth in LLMs and diffusion models, as well as in text, image, video, and software/code applications. At the same time, utilities, infrastructure providers, and industrial firms are becoming more active in the field. Complementary evidence on markets, energy, development, trade, labor, and safety reinforces this broader perspective.

All of this suggests that the next phase of GenAI will be shaped not only by who develops the most capable models, but also by who can provide the necessary infrastructure, deploy the technology across sectors, manage associated risks, and protect or disseminate the resulting innovations. For policymakers, this means that GenAI readiness extends well beyond frontier model capability; for firms, patent filings can reveal strategic intent in ways that product announcements and benchmark results do not. More broadly, the growing presence of utilities, infrastructure providers, and industrial firms in the GenAI patent landscape may be one of the clearest signs that the foundational competition has moved beyond the model itself.

The chatbot remains the most visible part of the story, but patent data suggests that it may not be the most consequential.

---

References

[1] World Intellectual Property Organization (WIPO), WIPO Technology SPARK Report: Patent Trends Update in GenAI, 2026.

[2] Stanford Institute for Human-Centered Artificial Intelligence (HAI), AI Index Report 2026, 2026.

[3] Organisation for Economic Co-operation and Development (OECD), Artificial Intelligence Markets, 2026.

[4] International Energy Agency (IEA), Key Questions on Energy and AI, 2026.

[5] World Bank, Digital Progress and Trends Report 2025, 2025.

[6] WIPO, World Intellectual Property Indicators 2025: Patents Highlights, 2025

[7] World Trade Organization (WTO), World Trade Report 2025, 2025.

[8] International Labour Organization (ILO), Generative AI and Jobs: A Refined Global Index of Occupational Exposure, 2025.

[9] International AI Safety Report, International AI Safety Report 2026, 2026.

[10] SoftBank Group Corp. SoftBank Group Report 2026, 2026

[11] OpenAI, Our Approach to Patents, 2025