,
9 tweets,
3 min read
Read on Twitter
My post on @lyft / cloud blew up 🙏. I got a lot of people thanking me for the data I also got flack for throwing numbers w/o references. I've been in these trenches so lot of this was through experience but the engineer in me couldn't sit by without providing data, so here goes:
Criticism 1: why cloud vs build? That's extreme. Why not build vs rent/Colo vs buy? Fair. For 3-10 cabinets at 20-50kw total renting is 50% of building over 5yrs, so renting is same as building over 10 (minus tax benefits). For larger sizes, renting is ~20% cheaper, see below:
For a 3 cabinet, 30kw total capacity, renting over 5yrs is 230k vs 640k. Breakdowns below. 
For a slightly larger build, 8cabinet at 70kw, it's still much cheaper to rent: 560k over 5yrs vs 900k odd to build. Note that over 10 years this has diminishing returns vs owning.
For much larger builds, say 20 cabinets at 150kw, renting over 5 years is just 20% cheaper than building. If you grow beyond this, it's anyone's guess.
Conclusion: my take was on large businesses going on cloud vs building. Renting is an option that's not relevant here given their size.
Criticism 2: Netflix is not 30% of Internet data. You are correct. It's roughly 15% on avg. I missed noting that Netflix is 30% of Internet data at "peak". At @Google we build for peak, not 95%, so my memory is tied to that. Here's the data:
Criticism 3: networking is nowhere near as expensive. Wrong. Google spent 300M on an undersea cable. Large cos pay shit ton in secret handshakes with telcos. My favorite book on this recommended by @sunubunu who built AWS and Google interconnects: amazon.com/Tubes-Journey-…
Self criticism: I said Google has fancy robots that replace hot-swappable hw in seconds. I was wrong, they do have hot-swaps but robots to replace stuff are still WIP. datacenterknowledge.com/google-alphabe…
