$AMZN AWS CEO just shared some important details on AI usage on the a16z podcast. The key point is that AWS is more skewed toward inference and has lower exposure to frontier labs (so a healthier, more diversified workload base). Some of his comments: 1. Compared to some other hyperscalers and neoclouds, AWS's exposure to a single frontier model provider is only in the single digits %, while some others are in the 30-60% range. 2. Most of the usage on AWS today is core compute, storage, and inference, which they believe is durable (so less training than others) 3. AWS says YES to around 60% of compute requests and is planning to buy 2M GPUs in the next couple of years. 4. I found it interesting as he mentioned during the talk that because of AWS Bedrock and the OAI partnership: "That i
在 X 查看 ↗