the only pushback I have for this summary is on point #2. If by local models you mean Kimi K3. Only the likes of fireworks can run that beast. Even the quantized versions are massive and require a cluster of high performance GPUs to run. It's going to be awhile before Local (ie: people with consumer GPUs) models will rival frontier at the level everyone is imagining.

There's also this gap between what you need a large model to be. Large frontier models are generalists and stuck into harnesses that allow general interaction. Only now are they starting to differentiate the harnesses, and as a rule of thumb, narrower use cases don't need as massive/general of models. So that acts as a pressure, but it's not a linear pressure and based on many factors about consumer harness based apps that become popular.

Could there one day be something as versatile as claude code running locally? Sure, but we need hardware and software breakthroughs to get there, and they are pretty big breakthroughs... hardware breakthroughs have happened, but for instance optical chips are starting to ship, though there will be no optical consumer level chips for years, not because of scaling issues, but because the datacenters have reserved those chips for themselves over the next couple of years.

It literally turns into a race between them trying to make datacenters acceptable to the general public hoarding all the hardware that could be used to run local models and, what... people and companies that want to run locally?

I agree though that something has got to give.