I made a big mistake a few months ago. I bought a few servers for us to use in our office - you know, it's important to have a few machines at your disposal so you can run intensive tasks on them or run all kind of different tests you couldn't otherwise do on your single development machine.
Makes sense right? Well, not in term of ROI and here is why:
I bought a fairly good machine with 4GB of RAM and an AMD 6400+ X2 CPU - that cost me about $1000. Let's assume the machine is utilized for about 8 hours a day (in reality I guess most lab machines are used much less). Assuming 250 work days a year that's 2000 CPU hours a year. Let's say the life expectancy of this machine is 3 years so for 6000 hours I paid $1000. That's about $0.16 a machine hour.
Now let's compare this to a machine I can rent from Amazon EC2 (prices). Assuming Moore's law will hold true for the next three years (limits of Moore's law) today's EC2 large instance (m1.large) will be equivalent to two of the machines I bought for my office (I'm assuming that Amazon are nice enough to lower prices for a given hardware configuration as it becomes older). Add to my in-house investment the electricity / cooling / management (as in personnel) and it will make my (already) bad investment even worse. Oh, and for my in-house machine I had to put the money upfront and not "pay as I go".
Now what happens If I need more machines at one time (for isntance, to do some tests before a release) and none at another time (like every night). Having these machines hosted on EC2 gives me the flexibility to use as many machines as I need and then don't pay for the time they sit idle.
The bottom line - If you're a small startup - don't buy lab machine. If you're a big company you could probably save thousands of dollars every month by hosting your lab on EC2.