PatentBrief

Patent Landscape

Patent Landscape:
Natural Language Processing

Google published “Attention Is All You Need” in 2017 with no patent. Six years later, the patent landscape around large language models is the most contested AI battleground.

The NLP patent landscape splits cleanly into two eras. Pre-transformer methods — Word2Vec embeddings, early attention mechanisms, encoder-decoder architectures — are heavily held by Google and Microsoft Research from the 2014–2017 deep learning wave. Post-transformer methods are increasingly filed by OpenAI, Anthropic, Meta, and Google, and cover the techniques that turned raw foundation models into commercial products: RLHF, instruction tuning, MoE routing, retrieval-augmented generation, and the inference-time methods that make LLM serving economical.

The strategic question of the decade: can a patent on a foundation model architecture actually be enforced when the underlying methods are widely published in academic papers, openly distributed as weights by Meta and Mistral, and the courts have not yet ruled on AI training infringement? The most valuable LLM patents may end up being not the architecture patents but the product-layer patents — RLHF, in-context learning, agent orchestration — where the methods sit closer to commercial behavior and further from open-research norms.

Key Patents

US10,452,9782019

Attention-Based Sequence Transduction Neural Networks

Google

The original Transformer patent. Covers the multi-head self-attention mechanism that powers every modern LLM from GPT-4 to Claude to Gemini. Google chose not to enforce this against the industry, but its existence shapes every licensing conversation in foundation AI.

US10,963,6522021

Reinforcement Learning from Human Feedback for Language Models

OpenAI

RLHF was the breakthrough that turned GPT-3 into ChatGPT. This patent covers the reward modeling and PPO-based fine-tuning loop that produces aligned assistants — one of the most commercially valuable methods in AI.

US11,521,0712022

Mixture-of-Experts Language Model Routing

Google

Switch Transformer and GShard introduced sparse routing that lets trillion-parameter models train efficiently. Google's MoE patents underpin Gemini's architecture and OpenAI's rumored GPT-4 design.

US11,222,1932022

Few-Shot In-Context Learning Method

OpenAI

Covers the prompting-as-programming paradigm: providing examples in context rather than fine-tuning. This is the patent behind the entire prompt engineering industry and most ChatGPT/Claude API usage.

US11,138,3922021

BERT Bidirectional Transformer for Language Understanding

Google

BERT was the dominant pre-trained language model from 2018 to 2020. Google's BERT patents cover the masked-language-modeling objective that became foundational for retrieval, search ranking, and downstream NLP.

US11,580,1482023

Retrieval-Augmented Generation for Large Language Models

Meta

RAG is now the dominant pattern for grounding LLM responses in private knowledge. Meta's original RAG patent covers the dense retrieval + generator architecture used by every enterprise AI product today.

Key Players

Google

Holds the foundation patents most modern LLMs depend on — the Transformer, BERT, and the Mixture-of-Experts routing that underpins trillion-parameter models. Google's patent strategy is notably non-aggressive: it published the transformer paper openly and has not enforced the core IP against the industry. The patents remain as defensive leverage and licensing currency rather than active litigation tools.

OpenAI

The commercial leader in LLM deployment, filing patents on training methods and product UX rather than fundamental architecture. RLHF, in-context learning, and ChatGPT-style assistant patterns form the core of OpenAI's IP — methods that turn raw foundation models into products. The patent portfolio is small relative to Google's but commercially focused on the layer where users interact with LLMs.

Meta

The author of RAG and LLaMA, with a deliberately open-weight strategy that complicates traditional patent enforcement. Meta's NLP patents focus on retrieval, multilingual systems, and recommendation-oriented language models. The open-weight release of LLaMA reshaped the competitive dynamics — and the legal question of how patents apply to publicly released model weights remains unresolved.

Microsoft

Holds its own NLP patent portfolio (Turing models, semantic search, Copilot UX) while licensing OpenAI's technology via the deepest commercial AI partnership in the industry. This dual position — patent owner and licensee — creates the most complex IP overlap in foundation AI. Microsoft's enterprise reach makes its patent portfolio the most valuable for defensive cross-licensing.

What to Watch

01

Training Data & Output Litigation

NYT v. OpenAI and the cluster of related copyright cases will define whether training large language models on scraped data, and whether their outputs, trigger copyright and patent exposure. The outcomes will determine the cost basis for every foundation model going forward and may force the industry toward licensed-data-only training — a fundamental restructuring of LLM economics.

02

Agent & Tool-Use Patents

The next product layer above raw chat is agents: LLMs that plan, call functions, use tools, and coordinate with other agents. The methods — function calling APIs, planning loops, multi-agent orchestration, and memory architectures — are being patented right now by OpenAI, Anthropic, Google, and a wave of startups. The IP filed in 2024 and 2025 will define agent infrastructure for the next decade.

03

Inference Efficiency Patents

The unglamorous but most commercially valuable LLM patents cover inference: speculative decoding, KV-cache compression, quantization methods, and continuous batching. Each represents a multiplier on serving margins for any LLM provider. The patents are being filed by both the foundation labs and the specialized inference companies (Groq, SambaNova, Together) — and will determine the unit economics of generative AI at scale.

From PatentBrief

Explore NLP patents on PatentBrief →

Search transformer, LLM, and language model patents. Read any patent in plain English and understand the claims that define modern language AI.

Search NLP patentsAll patent landscapes →