A Tax Company Out-Patents Google on Generative AI, Banks Out-File NVIDIA — and OpenAI Barely Files at All

Search two decades of US patent titles for the vocabulary of the current boom — "large language model," "generative AI," "diffusion model" — and the curve barely exists before 2023, then goes vertical: a 27× explosion in three years. But rank who is filing and the leaderboard is not the one you'd guess. There is exactly one GPU company on it, at #11. Ahead of it: a tax-preparation company, two banks, a CRM vendor, an IT-outsourcing firm. The companies patenting generative AI are not the ones building it — they're the ones wiring it into money.

The vertical part of the curve

GenAI-titled filings crawled along at a few dozen a year through 2022 — the term "generative model" has a long academic history — then the ChatGPT cohort hit the patent office. Applications publish 18 months after filing, so filings made in mid-2023 start appearing in late 2024: 439 published applications that year, 1,729 in 2025, and 1,551 in just the first eight months of 2026. Grants — the slower-moving series — are following the same shape two years behind, and 2026 already has more GenAI-titled grants than every prior year combined.

The leaderboard nobody would guess

Here are the top fifteen filers of GenAI-titled applications published since 2023. Microsoft at #1 is the expected answer. Almost everything after it isn't: Intuit — TurboTax — at #2 with 54, ahead of Google. JPMorgan Chase and Toronto-Dominion Bank both out-filing NVIDIA. Tata Consultancy Services and UiPath ahead of Samsung Electronics. Finance filers as a group out-file the twelve biggest chipmakers combined, four to one.

The pattern makes sense once you see what these patents claim. A chipmaker's generative-AI value lives in silicon and is patented as accelerators, packaging, and memory — categories these title searches don't touch. But a bank's or a tax company's value lives in the workflow: using an LLM to reconcile transactions, draft an audit response, detect fraud in a payment stream. Those are exactly the "apply known AI to domain X" claims that flooded the patent office the moment LLM APIs became a commodity — each one cheap to file, plausibly novel in its niche, and useful as a defensive stake in a land rush.

And the labs that built the models? Barely there.

The starkest absence: the companies whose models started the boom. OpenAI's entire published US patent portfolio — every application, on any topic, ever — is roughly 78 documents. Anthropic's is 7. Microsoft files more GenAI-titled applications per year than OpenAI has filed in its existence. The frontier labs are betting on speed, scale, and secrecy — weights are trade secrets, not patents — while the application layer above them and the hardware layer below them file as fast as they can. It's the NVIDIA story of the 2010s replaying one level up the stack: the winners of a platform shift rarely patent their way to the win — but everyone building on the platform patents furiously.