Thanks. This price reduction is fairly standard as newer generations are available. However, rental prices for AWS are still at a massive premium to neoclouds, which is the point I’m making.
Prices will come down over time, but the hyperscaler premium will remain. Neoclouds, to the extent they can offer more services will be able to lower the delta between them and hyperscalers, but most won’t.
In an earlier article I also showed how GPU rental economics still work for fully depreciated 5-8 year old chips at spot rental prices.
Thank you. I didn't realize the rental prices for AWS are at a massive premium to neoclouds. Weird. I don't really understand why, though? AMZN can price this below cost for an extended period of time to get rid of the competition. They have the staying power. They have done it numerous times. However, they may need to be working together with NVDA (and hence the neoclouds), otherwise they may not get the GPUs :)
Thank you for sharing. It's a well written TSMC post.
Well they sell a very different product. AWS, in addition to renting GPUs, also provide the ecosystem to use that compute, e.g. pre-configured software environments, storage solutions, networking, etc.
Neoclouds mostly just provide the compute and nothing else. AWS kinda doesn’t need to do anything because most of the smaller neoclouds are already getting rid of themselves through commoditized pricing as they don’t have a differentiated product.
AWS can accelerate the process, of course, but not much benefit for them.
Thank you. What about this on June 6th https://aws.amazon.com/blogs/aws/announcing-up-to-45-price-reduction-for-amazon-ec2-nvidia-gpu-accelerated-instances/ ?
Not sure if I have missed it in your post.
It's not a neocloud but a hyperscaller. How do you think about this?
Thanks. This price reduction is fairly standard as newer generations are available. However, rental prices for AWS are still at a massive premium to neoclouds, which is the point I’m making.
https://aws.amazon.com/ec2/capacityblocks/pricing/
^ H100s at $3.93 per accelerator for AWS
https://www.runpod.io/gpu-models/h100-pcie
^ $2.19 for neocloud
https://www.cudocompute.com/products/gpu-cloud/nvidia-h100
^ $2.25 for neocloud
You get the idea.
Prices will come down over time, but the hyperscaler premium will remain. Neoclouds, to the extent they can offer more services will be able to lower the delta between them and hyperscalers, but most won’t.
In an earlier article I also showed how GPU rental economics still work for fully depreciated 5-8 year old chips at spot rental prices.
Thank you. I didn't realize the rental prices for AWS are at a massive premium to neoclouds. Weird. I don't really understand why, though? AMZN can price this below cost for an extended period of time to get rid of the competition. They have the staying power. They have done it numerous times. However, they may need to be working together with NVDA (and hence the neoclouds), otherwise they may not get the GPUs :)
Thank you for sharing. It's a well written TSMC post.
Well they sell a very different product. AWS, in addition to renting GPUs, also provide the ecosystem to use that compute, e.g. pre-configured software environments, storage solutions, networking, etc.
Neoclouds mostly just provide the compute and nothing else. AWS kinda doesn’t need to do anything because most of the smaller neoclouds are already getting rid of themselves through commoditized pricing as they don’t have a differentiated product.
AWS can accelerate the process, of course, but not much benefit for them.