Graphcore Patent Portfolio Analysis

The cheque came late to the file

The $222M Series E was announced on 1 December 2020. By then the first filing was three years old, dated 20 October 2017, and 103 families had already taken their first filing date: fifteen in 2017, twenty-nine in 2018, twenty-four in 2019, thirty-five in 2020. The money did not start the work. It landed on top of a record that was already more than half built.

2020 is also the heaviest filing year in the whole file. Twenty-nine followed in 2021, twenty-five in 2022. Then the count falls away: six, three, two. The most recent published filing is dated 30 May 2025, titled Efficient Error Detection Code. Whatever the raise paid for, the first-filing dates say the technical positions were being staked out well before it arrived.

The money did not start the work.

Almost nothing is still waiting to be decided

Of 168 families, 136 are granted. Thirty-two remain published and pending. That is 81 percent through examination, and the skew runs the unusual way for a company whose filing history only begins in late 2017. Most early-stage engineering records are a thin granted core with a large speculative tail behind it. This one is the opposite shape.

A high grant share tells you two things at once. Examiners have found the claims allowable across the offices involved, so these are tested rather than hopeful. And the pending band is small enough that the shape of the file is now largely fixed. What Graphcore claims is, for the most part, already settled rather than still in argument.

The biggest band is not the label

G06F, data processing, sits on 141 families. G06N, covering artificial intelligence and machine learning, sits on 41. H04L, digital communication, sits on 25. Then two small bands of seven each: H10W, and H01L for semiconductor devices. Families carry more than one classification, so the bands overlap, but the ranking is unambiguous and it does not match the sector tag.

Read the recent titles alongside the bands and the emphasis holds. Data processing in a machine learning computer. Shared memory device. Issuing of Chip-Configuration Requests to an On-Chip Configuration Control Bus. Efficient Error Detection Code. Memory, buses, configuration, error checking. The claims live where instructions and data move, not where models are described.

The claims live where instructions and data move.

Four names keep coming back to these documents

INTEL CORP cites these filings 77 times. GOOGLE LLC cites them 33 times. SAMBANOVA SYSTEMS INC cites them 25 times. MARVELL ASIA PTE LTD cites them 22 times. Citation counts of that size are not incidental contact. They mean the examiners and drafters working on those organisations' own applications have repeatedly had to reckon with what Bristol filed first.

Note who is absent from the heaviest end and who is not. The most frequent citer is also the largest filer in the same CPC groups over the last five years, with 603 families of its own. A record that a filer of that scale returns to seventy-seven times is a record that has been read closely, whatever the difference in size between the two files.

Ten jurisdictions, and three chip-making ones

The families appear in the United States, the European Patent Office, China, Japan, Korea, Germany, Great Britain, France, Canada and Taiwan. Ten offices, PCT set aside. Germany, Great Britain and France appear as national filings alongside the EP route, which is a deliberate and expensive choice rather than a default one.

Japan, Korea and Taiwan together are the part worth pausing on. A United Kingdom company with 168 families has paid to be examined and granted in all three. Nothing in the record says why. What it does say is that the coverage was not left to the two largest markets, and that someone decided the claims needed to hold in the places where this class of hardware is fabricated.

What the record was actually built around

Put the pieces in order. First filing October 2017. Heaviest year 2020, the same year as the raise, with the raise arriving in December after the filings. Eighty-one percent granted. G06F on 141 families against G06N on 41. Ten offices including Japan, Korea and Taiwan. Seventy-seven citations from the largest filer in the same groups.

This is not a machine learning patent file. It is a processor and memory architecture file that happens to be aimed at machine learning workloads, built early, examined hard, and granted across the jurisdictions where such things get made. The sector tag describes the application. The claims describe the machine. Those are different assets, and the second one is the durable one.