Amazon is building substations, laying fiber and keeping cloud costs down
gigaom.com
gigaom.com
Prime exists to get shipping volumes up far enough to pull things like trucking items daily from Vegas to Southern California, directly into the USPS system at the sorting facilities - and eventually direct deliveries.
The other thing - which lead to AWS, FBA, etc, was that each Amazon department had to make a globally usable API, even for (then!) inside facing stuff.
Combining those and you get a giant. Hopefully Bezos will stick around for a while yet - I can't see Amazon continuing to not maximize profits in exchange for volume very long after he's gone.
Edit: I was angling more at the fact that Amazon has been very dominant, and that's not going away soon. So if it starts getting bad, it's not like there will be something just as good overnight. Look at how long it took to erode MS's dominance from the 90's.
What I'd like to see is hard-drives that don't need to be screwed in and don't need sleds. Instead they are machined with grooves on the tops, bottoms and sides to be "slide in, slide out" and have the grooves be multipurpose in the sense that they are designed for airflow. i.e. the top of one hard drive mates with the bottom of an identical hard drive to create channels through which air can be forced through and cool both drives maximally, simultaneously.
Basically take the level of machining Apple puts into designing the the Macbook Pros and apply it to hard drives. If designed correctly, you should be able to pack them together, side by side in a way that they lock (like a jigsaw puzzle or the Heineken World Bottle[0]) in a stable "lattice" of hard drives that only permit sliding out to one side and permit efficient cooling.
Screws and sleds are so yesteryear.
[0] http://www.fastcodesign.com/1672684/heineken-s-lost-plan-to-...
[1] I hope that HN comments like these constitute prior art in case someone decides to go out and do this and patent the idea.
but they could probably make some side grooves, and little cutouts for spring loaded latches to catch on.
add some cutouts on the front that something could grab onto, and you might be able to avoid the need for the sled and screws, without building up an entire latticework out of drives.
there probably needs to be structure to hold the back plane(s) and line them up with the drives, anyways.
So what happens when you have, say, five drives like that and the one in the middle fails? How do you get it out?
Here's one: http://www.icydock.com/goods.php?id=163
I suspect they are prone to vibration issues that would make them inappropriate for enterprise use, but here's a rackable enclosure too:
I could give you a dozen examples, but one will suffice.
From a hosting provider I've used for many years, WebNX. Here are two example configurations from them:
E5-1650v2, 64gb ram, 4x 450gb SAS 15,000 rpm enterprise drives and hardware raid 10 with battery backup, 20tb of bandwidth - $289 / month, no setup fee
Dual 2690v2, 256gb ram, 6x 600GB 15k sas raid 10 w/ 2x SSD Cache - $950 / month, no setup fee
The bandwidth alone would cost you $2,300 at Amazon. And 20tb is nothing these days, countless quality providers offer 100mbps standard with $150 machines.
The AWS cr1.8xlarge with just 20tb of bandwidth is $5,134 per month. A one year locked in contract gets it down to $2,815 ($7,200 up front) - three times more expensive. And there are far cheaper providers than WebNX.
AWS is almost interesting, if you never need any bandwidth.
For example at AWS you can get a fully bidirectional 10gbps link, regardless of what other tenants are doing on the network. Nobody else that I'm aware of can match that. And you'll quickly find that 20TB of bandwidth at those network speeds are nothing, you can cross into petabytes without breaking a sweat (1 petabyte at 50% of bandwidth 24x7 for a month for just one server.) At that kind of network volume bandwidth costs dominate everything (about $65K for an petabyte of bandwidth at Amazon.) I challenge you to beat that anywhere.
Also if you have bursty traffic patterns, AWS comes out a lot cheaper because it's designed around quickly scaling up and down capacity.
In the end it depends on matching a provider to your particular needs because they vary wildy in strengths and weaknesses and costs.
Just because someone says it does not mean it's true.
The Internet says SDE comp at amazon is $90-130K. Seems safe to estimate total cost per employee to $125K on average. They're building enough office space for an additional 12,000 employees in Seattle. Assuming there are already 8,000 SDEs that's $1B dollars. I wouldn't call that "not much."
Installing a dense rack of servers is in the neighborhood of $150K. How many racks per year would you have to buy to make $1B "round to zero"?
What's actually important, and not being explained, is the relationship between the different types of costs and revenue. Developing & operating a service is roughly flat. It doesn't take 10x as much human resources to go from 10,000 requests per second or 100,000. However you will need roughly 10x the hardware resources to serve those requests.
The other important property of infrastructure, which is mentioned, is scale. The cost to service each request can actually decrease due to increased utilization and efficiencies. The 10x requests might only require 8x infrastructure, for example.
Interesting. For reference; a Backblaze pod is 4U and weighs 145 pounds, so 10 of them in a standard 42U rack would be 1450 pounds = 0.715 (short) tons (or ~648kg).
For compute clusters, power is often the limiting factor; for storage clusters, weight is often the limiting factor. If you put these clusters in a general-purpose colocation facility, you'll end up with half-full racks. Density on a per-rack-unit basis is not necessarily important for most of us.
Of course, if you're Amazon's size, you build your own facilities, and you can make sure to support the weight and power requirements of whatever custom hardware you can dream up!
Just my two cents. I don't expect to change anybody's mind. :)
Seriously. I'm not sure how well compensated AWS engineers are, however if the price engineering is just noise, why not pay the best in the industry?
The only metrics people are looking at is revenue, profits, growth, bandwidth and storage. But how about we start measuring pollution, waste, and environmental destruction for every company, right alongside the traditional metrics?
It's great when the stock price doubles. But does anyone think about the fact that this probably also means the amount of plastic trash in our oceans, and poisonous gas in our atmosphere doubles?
So, they get the server manufacturer to do the work instead of building it themselves. I guess the scale goes:
low volume: buy -> build own -> custom outsource :high volume
I would love to see AWS moving a bit further down the customer chain, providing one-click installations to non-savvy customers (I'm thinking on-click wordpress, magento, docker, etc.). Given the recent acquisition of MediaTemple by SlowDaddy I'm confident many GridServer customers would of moved from MT to AWS.
AWS provides CloudFormation templates that setup whole environments in one (or so) clicks. Bitnami and others provide pre-built, operational Wordpress and other stuff. There's a ton of stuff available.
tip: search for wordpress, magento
The biggest drawback of AWS is that you miss features not because they don't have them, but because they have so many of them. :)