At least with LLMs you have more options (Deepseek, Grok, offline models, etc.). It's still far from perfect, but it's not as bad as phones where you basically only have a choice of Android or iPhone if you don't want to have to live with major inconvenience (such as being unable to do online banking or pay for parking). It's also a lot easier to launch a competitor in the AI market: you just need capital. With phone OS's it's essentially impossible. The barriers to entry are too high.

> you just need capital

The capital itself isn't going to do anything sitting in the bank. It's used to procure a team of PhDs, an team of SWEs, DevOps, business people, HR, marketing, access to a GPU supercomputer (renting a couple of 5090s off Vast.ai ain't gonna cut it). For, say, $50 million, you could get the blueprints to an Android phone and port your choice of Linux userland and get drivers working, and then do a run of 20,000, sell them for $1100. Compared to training GPT5, $50 million is cheap. If we use an estimate of $1 billion for the whole thing, making a Linux phone running a hypervisor with an Android VM to run banking apps seems not-impossible. (Based on AVF.)