Pre-launch reporting widely claimed 2nm; a Google vice-president rebutted this publicly after launch, confirming 3nm. The chip targets always-running proactive assistance rather than discrete on-demand features, and the claimed 3.5x speed-and-energy improvement is the largest single-generation on-device AI claim Google has made.
Google Tensor G6
A 50 per cent larger TPU on the same 3nm class, so the gain is architectural rather than a node shrink.