He realized that insurance generated enormous carry, and if you were able to better utilize that (say, by buying other companies and improving them) then you had a pretty winning strategy.

Interestingly the Bank of England's Chief Economist has today attacked that principle of company law: http://www.bbc.co.uk/news/business-33660426

To quote from the article:

The Bank of England's chief economist has expressed concern that shareholder power is leading to slower growth.

Andy Haldane told BBC Newsnight that business investment had been lower than was "desirable" for years.

One reason was that a high proportion of corporate profits were being paid out to shareholders rather than reinvested in the company.

He said that in 1970, £10 out of each £100 of profits were typically paid to shareholders through dividends.

Today, however, that figure was between £60 and £70. Mr Haldane argued that left far less cash available for growth-boosting investment and that firms risked "eating themselves".

Corporate short-termism - a focus on immediate gains rather than long-term prospects - was a rising problem for companies and pre-dated the financial crisis, he said.

If so, then it's natural that stock price increases when revenue increases, albeit at a discount to actual profit depending on what your expected revenue:profit conversion ratio is.

The interesting part is the next step. Someone buys a stock expecting to benefit from the exchange (sell at a higher price that one purchased it for). Thus, if the stock price is correlated with revenue, and Amazon can increase revenue even without increasing profit, stockholders will still be happy.* Stockholders are looking for increased value for their stock, not any particular metric thereof.

*This is assuming Amazon finances the revenue increases through free cash flow (as mentioned elsewhere) rather than diluting stock

I'd say the optimal strategy is what Facebook and Google do with the double Irish Dutch sandwich http://conversableeconomist.blogspot.com/2014/07/double-iris...

There are all sorts of strategies for optimization. Many involve lobbying Congress for special carve-outs.

You can certainly find more expensive options than AWS if you look, but similarly paying less than for AWS is easy too unless you start relying on specific AWS features.

Someone also has to know how to negotiate with DC's (because power circuits are often the biggest cost factor) and how to separately get good deals on fiber (which is usually far cheaper than the DC's charges because most customers don't know better). DC's often nickel and dime for simple tasks if you don't do it yourself (or have hands that can do it for you). (RackSpace used to charge something $75 every time a box need to be reimaged. DC's will often take forever and charge several arms and legs to rack & stack & cable in your cage, but you will still be at their mercy for freight & delivery and pulling outside fiber, electrical circuits, cooling & rack leveling.)

Overall though, it's still cheaper for web-ops shops to do bare metal in (Google, Yahoo, Apple, etc. in their own DC's and other large/medium shops with predictable baseline loads in someone else's DC's) otherwise everyone would be using AWS. And that's clearly not the case for anyone who's ever been in a real enterprise shop.

Look at Wikia or most any other enterprise shop for baseline loads: mostly bare metal, because it's far cheaper for people whom know what they're doing. Also, BackBlaze... pretty much no way to undercut with them apart from buying an HDD manufacturer and going multi-datacenter-scale.