R0ML's Ratio

I found this article somewhat confusing to read, so I asked for a synthesis of it, shared below for anyone else who might run into the same.

https://share.cleanshot.com/7zq42cMg

I once worked for a company that went from 100k annually to over $1B in under a year. I was 18 and they gave me a corporate card and almost no particular restrictions on its use. I thought they were a bozo. Now I’m not so sure.

I see this regularly in large enterprise. My favourite example is getting a 20-30% discount on rack-mount server hardware purchased up-front for the next 3 to 5 years of growth requirements. Invariably, the executives that signed this "great deal" treat themselves to a champagne lunch to celebrate the savings.

Never mind that most of that capacity will sit around unused for years. The real problem is that by the time they get around to filling the last half of the capacity, the exponential march of Moore's law -- or the equivalents for storage and networking -- will have discounted the cost of that capacity by 50% or more anyway.

Similarly, I saw a corporation "lock in" a sweet discount for the WAN network provider for 7 years where they got an upgrade from 1 Mbps to 2 Mbps and a discount of 40%. At the same time, residential broadband was already 25 Mbps minimum and business-grade fibre had three-digit Mbps speeds within months. The sales rep that got that deal signed probably laughed his ass off.

It's a fundamental misunderstanding that it's easier or better to purchase things in big stair-step increments when following an exponential curve. The waste will be exponentially higher the longer the steps. Ideally, one would want to purchase hardware in the smallest possible time-steps, following product release cycles exactly. That's one of the benefits of the public cloud, you can switch CPUs (or whatever) with a button-press.

The example of this that's burned into my mind is that a vendor (Dell or HPE, I can't remember) convinced a government department to buy a 48-core AMD EPYC server for bioinformatics. This is one of those problems where you want the biggest possible single box because of the way the algorithms scale ("up", not "out"). They're stuck with that box for the next five or so years. Meanwhile the cloud public cloud is making available these monsters: https://techcommunity.microsoft.com/blog/azurehighperformanc...

Check out that beautiful exponential curve! If you "lock in" just 2 or 3 generations back, you're missing out on 90% of what you could be using.