[pbw]

Early on, the cloud's entire point was to use "commodity hardware," but now we have hyper-specialized hardware for individual services. AWS has Graviton, Inferentia and Tranium chips. Google has TPUs and Titan security cards, Azure uses FPGA's and Sphere for security. This trend will continue.

[0xbadcafebee]

The cloud's point was to be a computing utility, not to use commodity hardware. They may have been using commodity hardware but it was just a means to an end. That was also the era that "commodity hardware" was a literal game changer for all forms of large-scale computing for businesses, as before then you'd be beholden to an overpriced, over-valued proprietary computing vendor with crap support and no right to fix or modify it yourself. But anyway, the businesses you're referencing are literally the largest tech companies in the world; custom hardware is pretty normal at that size.

[jeffbee]

You must be talking about very early on because it would only have taken a short time spent on practical cloud building to begin realizing that much or even most of what is in "commodity hardware" is there to serve uses cases that cloud providers don't have. Why do servers have redundant power supplies? What is the BMC good for? Who cares about these LEDs? Why would anyone use SAS? Is it very important that rack posts are 19 inches between centers, or was that a totally arbitrary decision made by AT&T 90 years ago? What's the point of 12V or 5V intermediate power rails? Is there a benefit from AC power to the host?

[wmf]

You're not wrong but I would make a distinction between removing features (which requires little or no R&D and saves money immediately) and designing custom ASICs (which requires large upfront R&D and pay off only over time and at large scale).

[re-thc]

> realizing that much or even most of what is in "commodity hardware" is there to serve uses cases that cloud providers don't have.

Why wouldn't they have?

E.g.

> Why do servers have redundant power supplies?

So if you lose power you don't lose the whole server? It's even more important for cloud providers that have huge servers with high density. You connect the different power supplies each to an independent power feed so 1 can go down. Would you rather have 2x the capacity instead?

[BikiniPrince]

It was more of an opportunity initially. Bezos saw a chance to purchase extremely discounted systems. He had the foresight to move redundancy from a single host to clusters and then eventually data center level redundancy. This decision reshaped how services were implemented and scaled. Specialized devices returned because they offer a better value or in some cases enabled products that were never possible before. At the end of the day a lot of decisions are based around cost. Can you do it with a toaster or an oven cheaper? Can we cut all of our food into pieces and cook it in a Beowulf cluster of toasters? I think you get the idea.

[re-thc]

> hyper-specialized hardware for individual services > AWS has Graviton

This is commodity hardware.

It pretty much works off ARM spec. Besides AWS owning it and offering it cheaper it's not in any way faster / better than Intel / AMD / something else.

We've had custom motherboards, server cases, etc long ago even before clouds.

If Apple silicon happens in the cloud then maybe...

[benlivengood]

The trend was to begin to use distributed software in computing clusters that allowed lower cost with commodity machines with lower individual reliability.

Now the focus is on cost-efficient machines in clusters, which is why we have the specialization.