>Give it another 6-12 months until the big AI shops shut those capabilities down and running local ones becomes cost prohibitive.

The local inference threshold has been crossed already. Its possible to achieve big model like performance through agent harnesses with local models.