The AI chip stack: who researches it, who patents it

Sign up for Personal to download the dataset

Loading…

From power plant to chip

Every discipline the AI build-out needs. Click a part to select its layer, more parts to add them.
POWER IN Generation Grid Storage AI DATA CENTER Conversion Power devices Cooling Networking Photonics AI & compute INSIDE A COMPUTE TRAY Packaging HBM memory GPU CPU Chip cooling Interconnect Transistors SerDes CHIP MANUFACTURING Lithography Fab process Metrology EDA

The stack

Measure
Region
Years
Companies

Share of papers over time

Click a country on the map or in the list to add it · click a chip to remove it

US patents by company

Click a company in the list or a logo on the map to add it

Who researches, who patents

Research papers by author country against US patents by owner country, per layer · click a layer to select it

How to read this

  • Layers run from power in (generation, grid, storage, conversion, power devices) through making the chip (lithography, fab processes, metrology), what's on it (memory, interconnect, analog, design tools) and around it (packaging, chip cooling, photonics) to the system (AI compute, networking, data-center cooling). Chip cooling is heat spreaders, interface films and cold plates on the package; data-center cooling is racks, coolant units and chillers, split by patent class. Click a layer in The stack to show it alone, click more to add them, or click a group name.
  • The stack: each layer's US patents in the selected year, split by the type of company that owns them (the leading patent holders in each layer; the rest are other owners). Mix shows each layer's split; Count scales bars to the largest layer (the largest layers dwarf the rest). The copper square and percentage are the share of the layer's patents naming an element.
  • Who researches, who patents: for each layer, the split of papers by author country against the split of US patents by owner country. Unknown countries are left out of both.
  • Papers: research papers in the materials and chemistry corpus Elements Live tracks whose title or abstract matches the layer's terms. Each paper's credit is split equally among its authors, and each author's share goes to their institution's country, or the country in their address. A paper can match more than one layer, so the whole stack counts some papers twice.
  • Research intensity (RI): the same papers weighted by citations and how recent they are, as on the research atlas: each paper's country credit times its impact score. Share of RI is a country's part of it; RIS is its RI against the leading country's (1.00). Recent years are provisional, since young papers have had little time to be cited.
  • US patents: patents granted by the USPTO since 2005 whose main patent class (CPC) is specific to a layer, credited to the owner's country. Only patents that apply count: switching and routing count for networking, phone radio and video don't; processors, caches , logic transistors and machine learning count for compute, user interfaces and databases don't; fibre and lasers count for photonics, displays and LEDs don't. Classes that mix the stack with other uses (gas turbines for power plants and jet engines, heat exchangers, circuit boards) are decided patent by patent: a 27-billion-parameter language model, run on our own hardware, read the title and abstract of 700 sampled patents per layer and judged whether each applies, and a text classifier trained on those judgements decided the remaining 370,000. On held-out patents the classifier agrees with the model 78–94% of the time depending on the layer, about as often as a second language model does; 163,000 of the mixed-class patents were kept. Power devices are the transistors whose text names silicon carbide, gallium nitride, IGBTs or high voltage. A company counts only its patents that apply, never its whole portfolio. Chinese companies file mostly at home, so their share of US grants understates their patenting; a US grant is the one that matters for selling into the US.
  • Patents ÷ papers: a country's share of the layer's US patents divided by its share of the papers. Above 1 it owns more than it publishes; below 1 it publishes more than it owns. Countries under 10 papers or 10 patents in a year are grey.
  • Naming an element: the share of patents whose text names a chemical element or compound, our proxy for patents that depend on a material. The rest are methods and designs.
  • Companies come from the patents: the 15 largest holders of counted patents in each layer since 2005, plus the 20 largest since 2021, so rising names show. A company counts only in the layers where it makes one of those cuts; its patents in other layers stay with its country but never carry it in on its total. Its role follows the layer, so Hitachi is equipment in metrology and a memory maker in memory. Universities and government labs are left out of the companies (they still count for their country). Owner names are folded into one company with its subsidiaries and former names (Kioxia includes Toshiba's memory patents, Tesla includes SolarCity and Maxwell, NVIDIA includes Mellanox, Hitachi includes Hitachi High-Tech). Each sits at its headquarters; a logo's size follows its US patents in the selected layers. Logos are the companies' trademarks, shown to identify them.
  • 3-year average smooths small layers; switch to single years for the raw counts. 2026* runs to , so its counts are partial; shares are not affected.