3 ms·

The inference providers are running batch sizes much larger than 10