NVIDIA spent fifteen years building CUDA before the market it was built for arrived. That head start, not any single chip, is what competitors have failed to overcome — an accelerator is only as useful as the software willing to target it.
On the hardware side the cadence has compressed to roughly a year: Hopper to Blackwell to Blackwell Ultra to Vera Rubin, each generation arriving with a rack-scale system rather than a part number.
NVIDIA is also the reason cross-vendor comparison is so difficult. It quotes sparse throughput by default, which roughly doubles the headline figure relative to a dense number — and most people repeating the figure do not mention which one it is.