A few things, including:

1. Price-performance has struggled to stay competitive. There’s some supply-demand forces at play, but the top companies consistently seem to strike better deals elsewhere.

2. The way AWS is architected, especially on networking, isn’t ideal for AI. They’ve dug their heels on in their own networking protocols despite struggling to compete on performance. I personally know of several workloads that left AWS because they couldn’t compete on networking performance.

3. Struggling on the managed side. On paper a service like Bedrock should be great but in practice it’s been a hot mess. I’d love to use Anthropic via Bedrock, but it’s just much more reliable when going direct. AWS has never been great at these sort of managed services at scale and they’re again struggling here.