That is interesting. I know nothing about server loads in the Cloud, but I'm guessing that they're more hammered than many on-premise company servers. Does that have a real impact on hardware life ?

I mean, sure, the disks will need replacing every now and then, but that depends on the disk, not on an arbitrarily-determined lifespan record. I've just bought a few 8TB NAS discs for my Synology. They are guaranteed for the EU-standard 2 years. They will likely have a much longer useful life, but they're spinning rust. I will use them until they fail, not until the warranty runs out.

So five years for a server, why ? CPU and architecture progress has slowed to a crawl, and I seriously doubt the newest USB protocol is a requirement for a server.

So, if you replace the disks when they fail, why throw out a motherboard if it's still working ? It's not like a CPU slows down over time, its cycles don't wear out like batteries.

So why set an arbitrary limit of five years for kit that could well do double that ?

When we compared their prices to on prem we could see that they're quite high... so my guess would be that they make so much money it's just not worth their while to bother repairing anything.

The accounting rules say that you divide the capital cost of the server by its expected lifetime in years, and you put that cost in your accounts each year.

So you have to know how long the server will last.

Amazon announced a change to how long the server lasts *for accounting purposes*. While this has to be based on reality, it may be based on changes to replacement cycles that have happened gradually.

There are a bunch of reasons to replace servers:

Harder to get hold of replacements. No-one's making 5 year old CPUs as a rule (although I daresay AWS may have contracts with Intel/AMD etc to supply CPUs for a longer period) so maintaining a big enough fleet of server type A becomes too hard and you have to retire them and replace with server type B.

Technical debt - keeping your stack running on old kit means one more piece of tin to support so you have to retire it as new kit comes in and replaces it.

Marketing - if your cloud is using 5 year old CPUs and competitor is using 1 year old CPUs, which cloud are you going to use?

Power/cooling. New CPUs are often more efficient than older ones, so there are cases it's worth replacing old CPU models with newer CPUs. They probably don't have a much higher clock speed (if at all), but probably have higher core count so you can gain significant savings in footprint, power and cooling with newer CPUs.

As for impact on life of running in the cloud, I'd imagine they're mostly running hotter all day than on-premises kit; I don't know if that impacts their running life more than the warmup/cooldown cycles of on-premises.

A long time ago Sun Microsystems' datacentre-in-a-container had "fail in place" designed in. You didn't replace failed components, you just blacklisted them and carried on working around them.

If I were running AWS I'd expect something similar - when components fail you take them out of service forever and operate the rig round them. If at 5 years you have 10% of failed components then it might well be worth operating them another year.

You're right about the useful life being potentially quite long now, but the five years 'book life' may not be the actual true lifespan for installed hardware which may vary by model, we're binning kit at between four & eight years so I'm guessing that our 'book life' is four.

Given the footprint AWS servers will occupy and the daily harware refresh rate required, I doubt they do anything other than shutdown individual servers if anything fails and just wait for the entire rack to either reach a overall failure level or (far more likely) just hit max age and be replaced en-masse.

All they need is a few operatives (not expensive techies) and a few mins to wheel in a prebuilt rack, connect the power & data cables, job done.

We budget for, and depreciate, our servers based on a five year life (desktops and laptops three, tablets two) but don't necessarily replace them that often. The last couple I replaced ran for over seven years, in the end the need for more storage meant we needed a new SAN anyway and support for the motherboards was limited so I replaced them.

At $31Bn >>>is now big enough that it deserves its own disclosure and discussion<<<

So to make the financial summary sheet these days the business needs to be heading towards $100 million revenue per day, quickly scribbles some numbers $31bn works out at around $59k per minute - substantially more than the $51.5k US median annual salary*

*using the U.S. Bureau of Labor Statistics numbers for Q2 2021 (Yes I know, govt. statistics) and if you don't know the difference between median and average you're in the wrong forum.

I'm still getting my head around the utterly bizarre costs to make a video...

$465 million to make the show, on top of the $250 million for the rights

No wonder they've had to put the price of Prime up. No doubt I will be subject to every increasing attempts to get me to sign up...

At which point in the future does it make economic sense for AWS, Google, Microsoft, etc

to develop and/or operate there own mini, small nuclear power plants? Are there rumors of any plans ?

The amount of energy these players consume is crazy and still growing fast just as the rest of the market..

and actual supply has to be shared with other industries and end-users ...

and we are at the beginning of the EV-cars energy needs .

15 years ago non of these cloud-players where into developing and making there complete hardware stack form CPU to network gear ...

and now there the biggest server builders on the planet.

Not a stupid question, If one decides to build a server farm in the high Arctic I'm sure the SMR designs will be considered to power the months of darkness.

For the rest of the world, I'm guessing they will just build enormous solar & battery farms as that provides a far more granular up-tick in capacity.