Patent Landscape:
Artificial Intelligence

PageRank — Hyperlink-Based Document Ranking

Stanford University / Google

The foundational algorithm behind Google Search, PageRank introduced probabilistic graph-based ranking that became the conceptual ancestor of modern embedding and attention-based ranking in AI. This patent expired in 2018, opening the technique to all.

Reinforcement Learning with Deep Neural Networks

Google DeepMind

Covers the deep Q-network (DQN) architecture that enabled AI agents to learn Atari games at superhuman levels — the breakthrough that ignited the modern deep reinforcement learning era. Foundation for AlphaGo and subsequent game-playing AI systems.

Transformer-Based Language Model Training Methods

Microsoft

One of Microsoft's derivative filings covering transformer architecture applications. The original transformer paper (Vaswani et al., 2017) was not patented by Google, making the surrounding ecosystem of application patents significant for enterprise AI licensing.

Facial Recognition via Neural Network Feature Extraction

Apple

The neural network underlying Face ID — Apple's 3D facial recognition system. Covers the specific method of generating a mathematical face model using infrared dot projection and a neural network, powering biometric authentication for over a billion devices.

Intent Recognition for Voice-Based AI Assistants

Amazon

Covers Alexa's intent recognition pipeline — parsing spoken utterances into structured commands using a combination of acoustic models and natural language understanding. Fundamental to any voice assistant that maps speech to device actions.

Natural Language Processing for Structured Data Extraction

IBM (Watson)

One of IBM's Watson NLP portfolio patents covering entity extraction and semantic relationship identification in unstructured text. IBM has filed over 9,000 AI-related patents — the largest portfolio of any company — making Watson's NLP IP a critical reference for enterprise AI developers.

Convolutional Neural Network Architecture for Image Classification

Google

Covers specific CNN layer configurations and training methods used in Google's image recognition systems. These architectural patents define how modern vision models are structured, relevant to anyone building image classification or object detection systems.

Speech Recognition via Recurrent Neural Networks

Baidu

Baidu's Deep Speech patents cover end-to-end speech recognition using RNNs trained on large datasets — the approach that displaced Hidden Markov Models and became the basis for modern ASR systems. Critical prior art for any speech-to-text system.

Google DeepMind

Publishes extensively while filing patents on specific architectural implementations. Leads in reinforcement learning, AlphaFold protein-folding IP, and large language model training methods. Strategic tension between open research culture and commercial IP protection.

IBM Research

World's largest AI patent portfolio — over 9,000 filings. IBM's strategy centers on enterprise AI: data governance, explainability, and bias detection. Aggressively licenses its portfolio to enterprise customers while using patents to maintain enterprise relevance as cloud competitors rise.

Microsoft

Leverages its OpenAI partnership strategically — OpenAI deliberately limits its patent filings (to prevent forks from being blocked), while Microsoft patents the commercial applications. Azure AI services, GitHub Copilot, and enterprise integration are the core of Microsoft's AI IP strategy.

Nvidia

Dominates AI chip architecture patents. CUDA parallel computing IP, tensor core architecture, and NVLink interconnects are the infrastructure patents underpinning every major LLM training cluster. Nvidia's hardware patents create a deeper moat than software AI patents.