NVIDIA Won the Generative AI Boom With the Emptiest Patent Vault in the Room — Then Started Re-Arming
In 2020, the year every AI lab started queuing for GPUs, NVIDIA was granted 155 US patents — fewer than AMD, one-eighteenth of Intel. Twenty years of USPTO data show a company that wound its patent machine down to almost nothing, won the generative AI boom anyway, and then quietly wound it back up 7× — starting well before ChatGPT.
In 2020 — the year GPT-3 landed and every AI lab on earth started queuing for GPUs — NVIDIA was granted 155 US patents. AMD got 204. Intel got 2,872. Among the ten biggest names in chips, the company about to win the generative AI boom held the emptiest patent vault in the room, and it wasn't close. Twenty years of USPTO grant data show something stranger than a company that patents little: a company that deliberately wound its patent machine down to almost nothing, won anyway, and then — starting well before ChatGPT — quietly wound it back up 7×.
The drought: NVIDIA's grants halved while it was winning
NVIDIA's granted-patent curve looks nothing like a rising tech giant's. It climbed steadily to a 352 peak in 2016 — the year the deep-learning datacenter business took off — then fell for four straight years, bottoming at 155 in 2020, an 11-year low. For two of those years, 2019 and 2020, AMD — a company then one-tenth NVIDIA's market value — was granted more US patents than NVIDIA. Hover the chart for exact values.
A grant arrives two to three years after its filing, so the 2020 trough traces back to a filing decision made around 2016–17 — and the filing data (next chart) shows exactly that decision. What makes the trough remarkable is the timing: those are precisely the years CUDA became the default substrate of machine learning. NVIDIA's moat in that period wasn't patents at all — it was software lock-in, developer mindshare, and shipping faster than anyone could copy. The patent numbers say the company knew it: one of history's most valuable technology franchises was defended, through its entire ascent, by one of the industry's thinnest patent walls.
The re-arm started before ChatGPT
Counting distinct new applications by filing year strips out the grant-lag blur, and the story sharpens. NVIDIA filed roughly 300–400 applications a year through the late 2000s, spiked to 584 in 2013 during its brief patent-licensing era (it launched a GPU licensing program in 2013 and sued Samsung and Qualcomm in 2014; the campaign fizzled and was settled in 2016) — and then all but stopped: 77 new filings in 2016, and 77 again in 2017, for a company already worth $100B.
Then the machine restarted: 166 filings in 2018, 301 in 2019, 394, 460, and 541 by 2022 — a 7× ramp that began four years before ChatGPT, right as the datacenter segment became NVIDIA's growth engine. This wasn't a reaction to the generative-AI frenzy; by the time the frenzy arrived, the re-arming was already in year five. The 2023–25 grant surge — 370, then 420, then a record 567 — is simply that filing wave clearing the patent office. And the new filings are more secretive than the old ones: 15.8% of NVIDIA's 2019-filed grants were never published before grant (kept confidential until issuance, versus an 8.4% norm across the whole corpus), a filing posture typically chosen when you'd rather competitors not read your applications while they're pending.
Meanwhile, the rest of the room never stopped filing
The contrast with the companies NVIDIA depends on — and competes with — could not be sharper. TSMC, which manufactures every flagship NVIDIA GPU, grew its US grants from 434 to 4,170 over the same twenty years, passing Intel around 2020 and never looking back. Intel, the patent superpower of the 2005 chip world, peaked at ~3,000 grants a year in 2017–19 and has slid toward 2,000 since.
Put NVIDIA in that room and the asymmetry is stark: over 2016–2025 the generative AI boom's biggest winner collected fewer total US patents than Samsung Electronics collects in five months. Every other company on this table out-patented NVIDIA over the decade — most of them by 5× or more.
| # | Assignee | Trend | Total | ’16 | ’17 | ’18 | ’19 | ’20 | ’21 | ’22 | ’23 | ’24 | ’25 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Samsung Electronics | 62,942 | 5,531 | 5,828 | 5,883 | 6,522 | 6,494 | 6,460 | 6,346 | 6,207 | 6,513 | 7,158 | |
| 2 | TSMC | 29,487 | 2,250 | 2,389 | 2,423 | 2,298 | 2,808 | 2,769 | 3,007 | 3,655 | 3,718 | 4,170 | |
| 3 | Qualcomm | 28,622 | 3,030 | 2,694 | 2,323 | 2,366 | 2,292 | 2,152 | 2,646 | 3,879 | 3,452 | 3,788 | |
| 4 | Intel | 25,647 | 2,786 | 3,036 | 2,740 | 3,030 | 2,872 | 2,613 | 2,428 | 2,147 | 1,928 | 2,067 | |
| 5 | Apple | 24,936 | 2,109 | 2,224 | 2,159 | 2,483 | 2,781 | 2,536 | 2,290 | 2,536 | 3,084 | 2,734 | |
| 6 | 21,042 | 3,216 | 2,709 | 2,203 | 2,173 | 1,866 | 1,527 | 1,569 | 1,876 | 2,091 | 1,812 | ||
| 7 | Micron | 14,532 | 864 | 802 | 922 | 1,274 | 1,534 | 1,791 | 1,926 | 2,232 | 1,796 | 1,391 | |
| 8 | SK hynix | 9,309 | 1,138 | 967 | 797 | 807 | 936 | 968 | 1,124 | 893 | 767 | 912 | |
| 9 | NVIDIA | 2,924 | 352 | 280 | 197 | 166 | 155 | 205 | 212 | 370 | 420 | 567 | |
| 10 | AMD | 2,345 | 211 | 135 | 151 | 205 | 204 | 199 | 239 | 303 | 356 | 342 | |
| All US grants (every assignee) | 3.2M | 304k | 319k | 308k | 355k | 353k | 329k | 324k | 313k | 325k | 325k |
Whether the re-arm changes anything depends on what the new filings protect — but the shape of the curve already says something: the era in which NVIDIA could defend its position with software gravity alone is, by its own filing behavior, over. When a company that famously didn't bother with patents starts filing 500+ a year and keeping an unusual share of them secret until grant, it is preparing for a different kind of competition.
1 Granted patents are counted from the grant document itself (USPTO kind codes B1/B2), by grant year — the earlier pre-grant publication of the same application is excluded so nothing is double-counted. New filings are counted as distinct application numbers by filing date, seen via either the application's publication or its grant. Filing-year counts for 2024–25 are omitted: applications normally stay unpublished for 18 months, so recent years are incomplete by construction.
2 "NVIDIA" merges assignee-string variants of NVIDIA Corporation (including misspellings in the source XML) and excludes separately-named subsidiaries (NVIDIA International, NVIDIA Technology UK, acquired companies such as Mellanox that file under their own names). Per this site's convention, each table row is a single filing entity — Samsung Electronics excludes Samsung Display/SDI, Google excludes other Alphabet entities, and so on.
3 "Never published before grant" = granted as kind B1 (no pre-grant A-publication). The usual route is a nonpublication request, available only for US-only filings; a small share are applications that simply granted within the 18-month window.
4 Source: 13.7M USPTO patent documents, analyzed by essentia.ai — Patent Analysis AI; figures pulled August 2026.