Measured hospital AI deployment expanded between 2022 and 2024, but did the geography of access converge, and did persistent access deserts overlap with greater health burden? Using two waves of the American Hospital Association (AHA) Annual Survey linked to 2020 Census block-group populations, we estimate contiguous-U.S. coverage for 329.3 million residents and full-frame transition profiles for 334.7 million residents. Measured AI-enabled status is identified using five binary workforce/workflow AI-use items in 2022 and fourteen ordinal clinical and operational AI implementation items in 2024 (primary 2024 threshold: expanding or fully integrated). The share of hospitals reporting active AI deployment rose from 18.3% to 28.6%. Contiguous-U.S. coverage within a 30-minute drive increased from 67.0% to 76.1%, yet spatial inequality grew: the population-weighted Gini coefficient of access distances rose from 0.739 to 0.767. In the full transition frame, 45.1 million people newly crossed the 30-minute threshold, while 67.9 million remained outside it in both waves: 37.2 million experienced no travel-time improvement and 30.7 million improved but still did not cross the threshold. These findings reveal a diffusion paradox: measured expansion coexists with persistent and, by some measures, rising inequality in who benefits. The communities left behind are not randomly distributed; they are more rural, lower-income, higher-poverty, older, more uninsured, and carry higher baseline premature-mortality burden than persistently served communities. A pre-diffusion Years of Potential Life Lost (YPLL) check showed that the mortality-burden gradient was already present before the 2022 to 2024 diffusion window, supporting a burden-overlap interpretation rather than causal mortality evidence.