Grow or Die
For its whole life, Amazon faced a classic capacity planning problem that everyone has. You need to have many machines at the busy time of the year, but they will lie idle the rest of the year.
While growing rapidly, your answer is easy: arrange to buy the needed machines as late as possible, so they came on-line just before Black Friday. The rest of the year they don’t exist, so they don’t cost you anything.
Something changed
But what if you’re not growing at 30% a year? Now you do have idle machines, for 11 out of 12 months a year.
You could arrange to start growing again. Amazon did just that, when they went from running a few data centres just for themselves and started AWS, Amazon web services. Now they were getting business from tons of new customers, and jumped back to their previous solution.
In it, you add as many machines as you can afford, until you are up to the predicted number you needed for the next “Black Friday”. You’re growing like crazy again. You can buy machines and entire data centres again … until business starts to flatten out.
What now?
You’re back to not having to buy extra every year, but you still have to pay for running everything you bought. And your profits have to keep paying back the capital investment you made when you bought those machines.
For a little while you can just coast, profit-taking on your investments.
But that won’t last. Your smaller competitors are still growing as fast as they can get customers., including by taking them from you. When that stops, you’re all in the same boat. Microsoft, Google, Alibaba and all the little guys.
Choose any Two
There are three classic approaches to staying alive in business: maximize growth, maximize revenue, or minimize cost. You did the first, now you need to do the others.
While you’re doing that, let’s pretend you’re Dr. Evil so your solutions don’t have to be ethical (:-))
Start with cost minimization: extend the effective life of your servers. That works because Moore’s Law is coming to an end, and clock speeds have stagnated. That in turn means a three-year-old machine is no longer hugely slower that a new machine. You can keep using your old machines without a cost-benefit crash. Amazon did just that, extending the life of its servers to six years. They claimed a $900m benefit in the first 90 days [Cantrill, 2024].
Then take advantage of some of the smaller advantages:
Wright’s law [Wright, 2025] gives you the benefit of a roughly 15% improvement when you double the volume of your purchases. Keep your old servers for longer, but replace them in big batches to get the best possible price.Sell services on a per-core basis. The CPU speed of new machines isn’t rocketing upwards, but the number of CPU cores is still improving at a reasonable rate. Each time you get a server with twice as many cores, you double your profit from it.Reduce your hardware costs. All the big cloud vendors now assemble their own machines, which they call “hyperscalars”. These have the maximum number of CPU cards that will fit a largish rack, but far fewer non-profit-making parts. A normal rack of 32 high-performance 1U machines will have 32 sockets, but 64 power supplies, at least 128 fans and 32 cases. None of the latter produce profit, but they cost money. In the case of power supplies and fans, they suck electricity. A good hyperscaler can save almost half the power and cooling that a rack of normal machines uses, over and above requiring you buy fewer cases, fans and power supplies.Charge for extras. If you’re VMware and selling a virtualisation engine, charge one price for a version that makes you shut down a service before moving it, and more for a version that will move the service to a new physical server without interruptions.Make it hard to go elsewhere, until moving is seen as difficult and unusual. 37signals exited the cloud back in 2022, and at the time that was seen as something only brave experts could do.Finally, keep creeping the price upwards. Your cost may be slowly decreasing, but you should always be increasing your prices, now you’ve locked in your customers.
Does this Sound Familiar?
Yup!
Cory Doctorow defines enshittification as
- starting with high-quality offerings to attract users, then
- optionally degrading those offerings to better serve business customers, such as advertisers, and finally
- degrading their services to all their users and customers, to maximize profits for the shareholders.
That’s a good tactic for Dr. Evil. But maybe not for everyone.
What does this do to me?
If you’re in the same pickle as Amazon, leave the Dr. Evil solutions out, and see how many you can apply to your version of the problem.
If your year-end business isn’t hugely greater that your regular business, the capacity planning problem is easier. A customer I know was once was in just that situation. They were able to use the longer life of recent servers and the price advantage of large purchases to solve their problem.
–dave
[Cantrill 2024] Bryan Cantrill, Moore’s Scofflaws, Oxide blog, https://oxide.computer/blog/moores-scofflaws
[Wright 2025] Wikipedia, Experience curve effects, https://en.wikipedia.org/wiki/Experience_curve_effects
#amazon #business #capacityPlanning #DrEvil #technology
Sanjay K Mohindroo
Sanjay Mohindroo
This Blog has No Leaves
zartom
sadedoes
Prashanth Mohan
Slim Bill (He/Him)
Maciej Lasyk