I'm interested to see how this scales cost effectively, at all the search engines.
DuckDuckGo is less profitable I think than Google/Bing, and are paying for API access. Bing appears to be running the models directly on hardware rather than paying for marked-up API access, but it's still expensive. Google seems to be in a good place with their custom chips, but who knows. It'll all be fascinating to see evolve.
What are the custom chips that Google has?
This feels like it’s not on my radar.
These have been huge for things like Home Assistant and object detection in real time video processing via Frigate and Google Coral TPU Accelerator. It sure cut down on the false positives I was getting for detecting cars and people coming up my driveway and not deer or my dog walking across it. They're only about $100 and come in M.2 and USB form factors.