Interview with a current $AMZN AWS employee on AI model usage, switching costs, and when to use what model: 1. He explains the pros and cons of using models via a service like Bedrock. The pros are that it's already native in your AWS, your data residency is all good there, you have all the HIPAA and standards that come with your AWS agreement, you get one single invoice, you can use AWS usage commitments, and you have one gateway for multimodal providers. The cons are there is sometimes a feature lag when comparing to direct API access at one of the model providers, and there is usually a thin extra layer of latency; usually it's sub-100 milliseconds. From a cost perspective, he doesn't believe there is a difference. 2. For AWS & the hyperscalers, the play is more in the clients using
View on X ↗