This search returns 358 published records covering the intersection of a parallelism strategy (model, pipeline or data parallelism) and an efficiency mechanism (pipeline bubble reduction, gradient accumulation, activation recomputation, collective communication, device mesh, or communication overlap). This is a bounded search string, not the entire universe of AI-training patents, so a broader or narrower query would return a different count. Treat 358 as the scope for this specific technical intersection rather than the whole distributed-training field.

The ranking returned by the data endpoint covers 86 companies counted by record, with the leader holding 36 records against 18 for the fifth-ranked assignee and 9 for the tenth. The top 5 assignees combined hold 35.8% of all 358 records, and the top 10 combined hold 55.0% — concentrated, but with a long tail of single- and few-filing entrants making up the rest of the ranking. Momentum figures for individual assignees should be read against publication lag, since the most recent year is always understated.

Filings grew from 10 in 2021 to 79 in 2024, a 690% increase over that three-year span, and 2024 is the most recent year that can be treated as a complete data point. The 2025 figure of 96 looks like a peak but is still filling in as later publications clear the roughly 18-month lag between filing and publication. Nothing in the data supports reading the field as slowing; the honest read is that the curve was still climbing through the last complete year.