“Before receiving its silicon, Etched emulated its full chip on a cluster of over 700 FPGAs and ran a dozen different models using its complete inference software stack.”