Microsoft extends life of cloud servers from four to six years Microsoft has extended the life of the machines powering its cloud by two years and will bank billions as a result. "We are extending the depreciable useful life for server and network equipment assets in our cloud infrastructure from four to six years," chief financial officer Amy Hood told investors last week during …
“I Like Money”
I know my jibe is not really relevant to the story, or maybe it is, but I have witnessed too often, poorly written processes which eat up electricity either because of inefficiency or worse, they are in an insane state.
Question: how many power stations are required to power machines all over the world running windows in the usual “stuck” state of “completing” updates?
One day, oomans might discover what true value means, I can give some clues, it ain’t money
it's a case of MS having vast amounts of concrete data on reliability of its fleet, which is custom hardware. They know way more about its performance and reliability than the vendors themselves do and indeed have negotiated custom warranties with the components. When you run at planetary scale you can try things like "what if we run a DC at a slightly higher temp, can we see any difference in hardware reliability? Can we save energy on cooling?" or "what if we extend hardware life, do we see any increase in storage errors?"
It's foolish to replace functional kit when you don't need to based on an arbitary lifetime. Do it based on changing hardware requirements, or improvements in perf or reliability. A fixed "best before" date doesn't make a lot of sense, particularly when you have billions of cores...
I agree with most of your points, but I'm not sure that much of their fleet is truly custom. Or certainly the stuff that drives data centre replacements ....
CPU - stock (Intel, AMD and Ampere) ... possibly custom sku
DRAM - stock
PMEM - stock
HDD - stock - may have custom firmware
SSD - may have custom firmware, but commodity NAND
PSU/PDU- maybe custom
motherboard - custom
switches - ??
Indeed, all good points and I am sure that is exactly what they are doing.
But, the largest public clouds are not really selling hardware. They are selling a software stack. And the margins on that trounces the margins on hardware and the marginal gain realized by extending their useful life. So, the real financial benefit for MS is less performant hardware consumes more software -- and customers will pay-up for that because the (software) service itself is sticky.
Looking at it from a revenue management perspective, slowing down kit upgrades suggests they don't expect to be power or space constrained -- and will not be "spilling" demand (i.e., lost revenue capture opportunities). So extending hardware lifespan might also suggest their internal estimates of future demand are slowing down and they see no operational need to upgrade hardware at the same pace -- because the service is sticky enough now.
Legacy outsourcers like IBM used the same playbook for the last 40 years. The real cash flow was --- and continues to be -- in the software stack.
There was a time when Google used medium tier off-the-.shelf stuff for their datacentres and made public some rough figures on how it made sense financially at the time. The core difference being that Google were masters of their own workloads, whereas Microsoft rent the compute and storage capacity out. But on their scale careful analysis of usage profile of existing stuff may be even more valuable than running their own stuff exclusively.
So I woud say perhaps more "sensible" with the money, rather then "ruthless".
It's probably a bit of both of those and a lot to do with Microsoft being unable to source kit in the volumes they'd like despite their bottomless money pit. Regions around the world have been plagued by instance shortages and capacity-linked outages for months on end now. Extending hardware life is a sensible response to that.
I was going to say, I didn't really understand the article.
Are they extending the *physical* life of the servers?
Are they extending the depreciation period? (I assume you're actually allowed to do that wherever it is that they file taxes)
Are they doing both?
Sorry if it's in the article and I missed it. It's getting late here.
You're right: once an asset has been depreciated it has been "paid for" by tax offsets.
What they're actually doing is assuming that they can extend the physical life of devices beyond the depreciation period and, therefore, profit from the window between refresh cycles. If necessary, ie. if it suits them, they'll get depreciation rules adjusted as well.
maybe Microsofts spreadsheet needs to factor in this week's meter readings ....
Servers are expensive... around the price of an automobile. Their power consumption is relatively moderate these days. The purchase price of the server can far outstrip the cost of electricity to operate it, particularly if you can choose to locate your data centre somewhere with inexpensive electricity and moderate cooling needs.
Dell estimates 3MWH/yr on a heavy workload for their fully kitted-out R740 servers:
* https://corporate.delltechnologies.com/content/dam/digitalassets/active/en/unauth/data-sheets/products/servers/Full_LCA_Dell_R740.pdf
Even using the UK average of £0.28 per kWh, that would be just £840/yr. At 5 years, that's £4200. You'll find that a fully populated new server costs considerably more than that, and that's not even accounting for the much lower electrical rate Microsoft pays. Locating close to cheap electricity is a trick Aluminum smelters have been doing for decades.
you need to pay to cool that 3MWH/yr also
With traditional refrigeration air conditioning systems, that only adds 1/3 to the power budget. But big cloud providers don't use traditional air conditioning, they run at high temperatures (so often need to HEAT-UP outside air they pull in), and often use evaporative cooling instead of refrigeration where needed.
See: https://www.youtube.com/watch?v=JZUX3n2yAzY
The data they have looks back on some of the most stagnant years for servers, since essentially all of the data concerns Skylake and its minor revisions over the past what, 6 years? A few Broadwell too, which were in the same process. So, this is an era of pretty much constant clock speeds, constant memory footprint, and unchanging core functionality. Did not even get AVX-512 off the ground. The recent surge in AMD will have little impact on the data the accountants look at, and anyway Rome and Milan brought essentially no new functionality. Pricing is dominated by DRAM which has also been stagnant around $3 per GB for years.
So, look ahead and what do we see? A torrent of truly new chips. Made on new processes. With new accelerators on chiplets, and new functionalities like security. With memory perhaps finally shifting to lower costs via disaggregation, which will also apply (probably even more effectively) to "local" SSD. And don't forget the rise of in-house processors like Graviton-3 putting the squeeze on.
It will be very interesting to see if any current machines are salable in 4 years, much less 6. The road ahead looks like much more change in function and price than the road behind.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
