The AI race is moving beyond language models. Patent data shows the emerging battleground in world models, multimodal systems, and embodied intelligence.
The AI landscape everyone’s watching isn’t the one that matters. While Silicon Valley obsesses over who has the best chatbot, patent filings tell a radically different story about where the actual innovation is happening—and who’s positioning to win.
442 patents. That’s how many multimodal AI filings hit the USPTO and EPO in just three years (2023-2025). For context, that’s a 404% growth rate in an AI category most people can’t even define.
Multimodal AI—systems that process vision, language, audio, and video simultaneously—isn’t a research curiosity anymore. It’s becoming the foundation layer for the next generation of AI products. And the patent data shows who’s taking it seriously.
Google leads with 40 patents, but here’s where it gets interesting: DeepMind filed 9 patents, all in 2025. That acceleration pattern? That’s what serious R&D looks like when a company stops theorizing and starts productizing.
But multimodal is just the appetizer. The real meal is what’s happening in three other categories that don’t get enough attention: visual generation, world models, and application-layer intelligence.
Everyone knows ByteDance owns TikTok. Almost nobody knows they’re quietly dominating visual generation AI with patent filings.
327 patents filed by the ByteDance ecosystem (ByteDance + Lemon Inc. + Zitiao Network Technology) in diffusion models, image synthesis, and video generation since 2023. That’s 127% more than Google (144 patents) and 130% more than Adobe (142 patents).
Adobe—the company whose entire business model is visual creation software—is getting out-innovated by a social media company.
ByteDance isn’t filing patents for bragging rights. Look at what they’re protecting:
This is application-layer IP. This is the stuff that turns research into products people actually use. And they’re filing at a velocity that suggests they’re building something much bigger than anyone realizes.
The strategic read is straightforward: ByteDance has distribution, compute infrastructure, and a fast-growing visual-generation patent position. That combination points toward a full-stack generative AI platform, not a side feature inside TikTok.
Here’s a number that should wake you up: 5,863 patents filed for world models, spatial reasoning, and environment simulation technologies since 2023.
That’s not a typo. While everyone watches ChatGPT demos, companies across wildly different industries are quietly building systems that understand 3D space, predict physical interactions, and model complex real-world environments.
And the leaders aren’t who you think.
Deere & Company (yes, the tractor company): 146 patents. They’re not making tractors. They’re making autonomous agricultural robots that understand crop health, soil conditions, weather patterns, and optimal harvesting strategies. World models for farming at scale.
Capital One (the credit card company): 88 patents. They’re building predictive models for financial fraud, risk assessment, and customer behavior that operate like world models for financial systems.
Koninklijke Philips (healthcare tech): 80 patents. Medical imaging, patient monitoring, surgical robotics—all powered by world models that understand human anatomy, physiology, and disease progression in 3D space.
Freeport Minerals (mining company): 60 patents. Autonomous mining equipment that navigates underground tunnels, predicts geological structures, and optimizes resource extraction using spatial reasoning AI.
World models aren’t just for self-driving cars. They’re the next general-purpose technology—the AI equivalent of what databases were to the internet era.
Every industry that deals with physical space, complex systems, or temporal prediction needs world models. And the companies building domain-specific world model IP today will own their verticals for the next decade.
The strategic implication is that Zoox and Waymo may be building infrastructure worth more than the robotaxi business alone. World-model IP can extend into delivery drones, warehouse robotics, fulfillment automation, and other physical systems that need spatial reasoning.
Now here’s where it gets really interesting.
OpenAI and Anthropic—the two companies everyone associates with cutting-edge AI—have filed a combined 138 patents since 2023. That sounds impressive until you realize:
This isn’t a bug. It’s a feature. OpenAI and Anthropic have read the room: model architectures commoditize fast, but application-layer IP compounds.
Look at Anthropic’s most recent filing: “Artificial Intelligence Agents for User Interface Task Workflow Automation” (WO-2025199330-A1). They’re not patenting the model. They’re patenting how the model helps users get work done.
That’s the app layer strategy. And it’s the right one.
Moonshot AI just proved that frontier-level model performance costs $4.6 million, not $100 million+. When training costs drop 95%, the training moat disappears.
So OpenAI and Anthropic are doing the smart thing: protecting the workflows, interfaces, and integration patterns that turn models into indispensable tools. Because when anyone can train a Claude-competitor for under $5M, the defensible IP shifts to:
Microsoft gets this too. They’ve filed 34 app-layer patents vs. just 8 multimodal model patents. They’re betting that owning the enterprise integration layer (Office, Teams, Azure) matters more than owning the model weights.
Here’s the framework that makes sense of all this data:
This is where user behavior gets captured, workflows get optimized, and moats get built through data network effects. Every interaction teaches the system more about how humans work.
Leaders: Google, Microsoft, OpenAI, Anthropic, Salesforce
Model IP can decay in months, while application-layer IP can compound over years as products learn user preferences, industry-specific workflows, and domain expertise.
Chips, training systems, simulation environments, orchestration tools. This layer is capital-intensive and tends toward oligopoly. A few players dominate.
Leaders: NVIDIA (chips), Deere (simulation), IBM (enterprise infrastructure), Waymo/Zoox (world model infrastructure)
Infrastructure is defensible when it is tied to vertical integration or domain-specific advantages. General-purpose infrastructure commoditizes quickly.
Architectures, training methods, parameter optimizations. This layer is growing the fastest (312%) but retaining value the slowest.
Leaders: ByteDance ecosystem (327), Google (144), Adobe (142), NVIDIA (75)
Open source and cheap training costs are driving model IP toward commodity status. The winners will be those who own distribution, like ByteDance and Google, or unique data access, like Adobe.
Based on patent velocity, filing focus, and strategic positioning, here are the companies most likely to surprise everyone in the next 12-24 months:
Why: 327 visual generation patents, 2 billion user distribution, vertical integration from model training to consumer app. They’re building a full-stack AI platform behind the scenes.
Watch for: TikTok-native AI creation tools that make Midjourney/Runway look like desktop software in a mobile world.
Why: World model IP that extends far beyond autonomous vehicles. Amazon needs this for delivery drones, warehouse automation, and fulfillment robotics. Alphabet needs it for everything.
Watch for: Licensing deals where auto manufacturers pay to use Waymo’s world model infrastructure instead of building their own.
Why: 14 conversational agent patents, deep enterprise integration, and a customer base that desperately needs AI workflow automation.
Watch for: Einstein AI becoming the default enterprise AI agent layer—the “Slack for AI workflows” that every company needs.
Why: 9 patents, all filed in 2025. That’s not research. That’s productization. And when a research lab starts filing patents at velocity, they’re preparing for commercialization.
Watch for: Standalone DeepMind products that compete directly with OpenAI/Anthropic, possibly focused on scientific computing and medical AI where their domain expertise is unmatched.
If you’re building in AI, the patent data gives you a playbook:
Unless you have a genuine architecture breakthrough (rare), model IP won’t hold value. Competitors will train around it, open source will release equivalents, and your IP decays faster than you can enforce it.
How users interact with your AI. How your system integrates into existing software. How you capture and utilize user behavior data. This IP compounds in value as you build network effects.
If you control a vertical (like Deere in agriculture or Philips in healthcare), world model infrastructure for your domain is massively defensible. General-purpose infrastructure? Let NVIDIA handle it.
Companies with strong app-layer IP but weak distribution are prime acquisition candidates. Microsoft didn’t buy OpenAI’s model weights. They bought the ChatGPT interface and user relationship.
If your industry involves physical space, complex systems, or temporal prediction, there’s still white space in world model IP. But that window is closing fast—5,863 patents in three years means the gold rush is already underway.
The AI race everyone’s watching—who has the best LLM, who raised the most money, who has the most parameters—that’s not the race that matters.
The real race is:
Patent data doesn’t lie. It shows where companies are placing their bets with money, lawyers, and multi-year R&D commitments.
And right now, the smartest companies in AI are betting that:
The question isn’t whether these trends are real. The patent filings prove they are.
The question is: Where’s your IP strategy positioned?
Model layer (commoditizing), infrastructure layer (concentrated), or app layer (compounding)?
Because in 12 months, that question will determine whether you’re leading the next AI wave or licensing from the companies that are.
Need to understand what this means for your portfolio? ipCapital Group helps leadership teams turn patent landscapes into practical decisions about filing strategy, competitive diligence, M&A readiness, and monetization.
Start a conversation with ipCapital Group or download the IP Strategy Playbook.
Work with ipCapital Group
From invention to monetization, our team has guided 2,000+ engagements across the full IP lifecycle. Start with a free 30-minute discovery call.
Written by
Seth Cronin