The AI Chip Race Peaked in 2022. The Race That Replaced It Is About Packaging, Power, and Plumbing

Ask the patent record where the generative-AI hardware race is happening and the answer has already changed once. The first land-grab — the neural accelerator chip itself — peaked in 2022 and is receding. What's rising instead is everything around the chip: the advanced packaging that stitches dies together, the memory that computes in place, and — fastest of all — the plumbing that keeps the whole thing from melting. Ten years of USPTO grant data trace a bottleneck moving down the stack, from silicon to package to radiator.

The accelerator wave has already crested

Patents dedicated to neural-network hardware — the USPTO's classification for circuits purpose-built to run neural nets, the category NPUs and tensor cores land in — grew twentyfold from 2016 to a 2023 peak. But the forward- looking series, published applications, tells you what companies are filing right now: it peaked in 2022 at 363 and has fallen three years straight, to less than half that. Grants echo applications with a roughly two-year lag, and they turned down right on schedule in 2024.

That's not the AI hardware race ending — it's the easy architectural territory being claimed. The matrix-multiply engine at the heart of an AI accelerator is, by 2025, settled ground: Samsung, IBM, and Intel hold the biggest positions, and the marginal patent there protects less and less. The filings moved to where the constraint moved.

Where the constraint went: the package, the power, the plumbing

Three physical categories tell the story of AI's industrial phase. Advanced packaging — the H01L23/25 classifications that cover die stacking, interposers, and the CoWoS-style assembly every AI GPU depends on — is the steadiest riser in the entire hardware corpus: up in nine of the last ten years, +33% since 2016, and it never stopped growing through the downturns that dented every other semiconductor category. Power conversion patents tell a demand story: they slid for years to a 2021 trough, then turned around exactly as datacenter power became a boardroom topic. And rack-level cooling sat flat at ~500 grants a year for six years — then inflected in 2022 and hit 903 by 2025.

The steep small curves are the next decade

Below the big buckets, four young categories are compounding from a small base — and their slopes, not their absolute counts, are the signal. Liquid and immersion cooling grew 4.8× and is accelerating (47 → 224 grants/yr, with 2024–25 the two biggest jumps on record). Quantum computing went from a classification that didn't exist to 341 grants a year in seven years. In-memory compute — doing the math inside the memory array instead of shuttling data to a processor, the direct heir of the neuromorphic wave — is up from single digits to 90. And chiplet patents, near zero before 2021, are doubling roughly every two years as the industry decomposes big dies into linked tiles.

Each layer also has a different cast. The accelerator category belongs to Samsung, IBM, and Intel. In-memory compute is led by SK hynix — a memory maker betting the category is memory's future. Chiplets are Intel and AMD plus a wave of startups (d-Matrix, Eliyan, Kepler Computing). And the cooling leaderboard has no chip designers on it at all: Dell, Baidu, Cooler Master, Quanta — the people who build the boxes, not the silicon. The bottleneck's owners change as it moves.

The full scoreboard

Read top to bottom, the table is a roadmap. The mature categories at the top move a few percent a year and belong to incumbents. The young ones at the bottom double and re-double, and their ownership is still up for grabs. If the pattern of the last decade holds — applications peak first, grants follow two years later, then the frontier moves down the stack again — the next categories to crest are packaging's successors: co-packaged optics and the thermal problem that never really gets solved, just relocated.