Hardware for local coding models is still affordable. For how long?