>I can't help but feel this is just Google trying to pull the ladder up behind then and make it more difficult for other companies to collect training data.
I can very easily see this as being Google's reasoning for these actions, but let's not pretend that clandestine residential proxies aren't used for nefarious things. The vast majority of social media networks will ban - or more generally and insiously - shadow ban accounts/IPs that use known proxy IPs. This means that they are gating access to their platforms behind residential IPs (on top of their other various blackboxes and heuristics like fingerprinting). Operators of bot networks thus rely on residential proxy services to engage in their work, which ranges from mundane things like engagement farming to outright dangerous things like political astroturfing, sentiment manipulation, and propaganda dissemination.
LLMs and generative image and video models have made the creation of biased and convincing content trivial and cheap, if not free. The days of "troll farms" is over, and now the greatest expense for a bad actor wishing to influence the world with fake engagement and biased opinions is their access to platforms, which means accounts and internet connections that aren't blacklisted or shadow banned. Account maturity and reputation farming is also feeling a massive boon due to these tools, but as an independent market it also similarly requires internet connections that aren't blacklisted or shadow banned. Residential proxies are the bottleneck for the vast majority of bad actors.