[Twirrim]

> most of AWS revenue is servers and storage

The way that cloud businesses work, you sell the servers for about as cheap as you possibly can do. Instance prices are all a race to the bottom among the providers, because servers are largely commodity hardware that's easy to get from any number of providers, and it's one of the first prices customers see, and often plays a big role in their choice of provider.

So that's not where you make your profits. So you're right, lots of revenue, but crucially, there is no real profit. There never will be. That makes it a boring product, not worth focusing a lot on from a marketing perspective etc. Same tends to go for all of what you might think of as the basic building blocks of the cloud. e.g. object storage prices are often really close to what it actually costs to provide the service.

You make your profits on what you sell that runs on the cloud. All those additional things like databases, streaming services, kubernetes bits, functions etc. Those are where you make your actual profits. GenAI is a big potential profit driver for AWS, so that's where they're pushing. A couple of years ago it was "$foo, but on Kubernetes". Before that it was "$foo, but Serverless". They're just pushing where the profit and interest is, and pretty much always have done.

sort of side-note: Gartner's evaluation of cloud providers got really absurd around kubernetes stuff. Because one cloud would do it, you'd miss out on points if you didn't also add it, even if being on kubernetes literally added zero benefit, or arguably was worse. Same for "Serverless". It didn't matter if customers were actually using it, or wanted it, if AWS/Azure/GCP launched it, you'd better have it too.

[chipdart]

> A couple of years ago it was "$foo, but on Kubernetes". Before that it was "$foo, but Serverless". They're just pushing where the profit and interest is, and pretty much always have done.

I don't think that's it. Those who migrated their EC2 apps to ECS/EKS/Fargate/App Runner have already migrated, so there is diminished returns in pushing those technologies.

The same goes for serverless. The whole world already adopted this to it's full extent. Those who want/can use these services, are already running these services, and AWS is already getting better utilization rates from their idle computational resources from this.

These are not fads. They are already infrastructure.

What we are seeing is additional high-level services being released to meet customer demand. There's now a massive need for training and running your custom private LLMs. There is absolutely no justification to skip the revenue you can generate by serving these markets.

[mtnGoat]

“The whole world adopted that” … I think only 30% of the worlds internet accessible compute is in the cloud. Hyperbole much? lol.

[chipdart]

You're commenting on services provided by a cloud provider. I think it's obvious this is about which services are being used by clients of cloud providers, and not how your grandma stopped using a kettle to instead make tea by invoking a lambda through API Gateway.

[dilyevsky]

Storage can be considered commodity too yet big three charges massive premium for it

[Twirrim]

Having worked for storage services for the cloud, it's not a massive premium. Things are kept pretty close to costs. We have to, it's a similar race to the bottom case. Each time one of the clouds drop storage costs, the others rapidly follow even to the point of it losing them money while they figure out how they might return to profitability.

The biggest cost for cloud storage services is not disks. Never has been. They're pretty cheap all things told. It's the per rack operational costs that dwarf them. The cost of electricity, cooling etc. You're also paying for the durability and accessibility that is built in to the software. It's why you see, e.g. backblaze obsessing about their server specifications and how many disks they can cram in to the server. Everyone is trying to maximise the rack data density, and trying to tow a really fine line on having just enough compute power in the individual server.

[dilyevsky]

Durability and accessibility is solved problem using erasure coding with automatic recovery. Nothing interesting in this space going on for like 10 years unless you’re pushing planetscale storage.

> It's the per rack operational costs that dwarf them. The cost of electricity, cooling etc

Ok lets do some simple math - lets say you have a few racks with 50PB in there, colocated. Power + floor space + remote hands will not cost you more to than 250k/y and that’s being very generous. Dividing by 50PB thats .5c per gig PER YEAR. Can you explain why GCS/S3 charges 50x that not including egress which is also ridiculous?

[mtnGoat]

Have you seen their private rate cards for this level of buying? I bet if you had your calculus would change here.

Also, it’s easy to beat most costs if you get consumer grade hardware and don’t refresh it more than once a decade. I see that often when folks compare costs, but that’s dishonest and just kind of shoots their credibility in the foot.

[dilyevsky]

Note that we haven’t even started discussing hw costs - tp correctly noted that amortized capex is less than the opex. It doesn’t change the math that much. I had pulled the relevant quotes to make a google sheets model and even with top line hardware i couldn’t stretch the break-even point beyond 2 years.

Private rates are cool except they come with strings attached and even discounted 90% its still not even close to break even on decent sized commits. Purely on costs the cloud cant win whichever way you slice it.

[mtnGoat]

So you’ve never actually had to acquire hardware or run systems at this scale? You’re just doing this on some Google sheets?

[sciencesama]

And securely provision it and access it

[jen20]

Metered egress would like a word, too!