On March 27th, Jeff Bar posted on the official blog for Amazon AWS that the company would be offering dedicated – as opposed to time-sharing – hardware through the EC2 cloud computing service. With 26 articles logged and counting on Google News, the story has been well covered, yet one of the more interesting details has been all but passed over. A few months ago, UBS came out with a study on the business potential for Amazon AWS. The study made headlines because it concluded that Amazon’s cloud possibly already produced $500M USD in revenue in 2010. The company’s announcement of dedicated instances is notable, because it allows back-of-the-envelope calculations to determine whether that number is actually feasible given the technology and pricing models used by Amazon – along with a whole bunch of other goodies.
The real juice of the blog post was in the last sentence of the second to last paragraph (more is added for context):
When you launch a Dedicated Instance, we can't use the remaining "slots" on the hardware to run instances for other AWS users. Therefore, we incur an opportunity cost when you launch a single Dedicated Instance. Put another way, if you run one Dedicated Instance on a machine that can support 10 instances, 9/10ths of the potential revenue from that machine is lost to us.
In order to keep things simple (and to keep you from wasting your time trying to figure out how many instances can run on a single piece of hardware), we add a $10/hour charge whenever you have at least one Dedicated Instance running in a Region. When figured as a per-instance cost, this charge will asymptotically approach $0 (per instance) for customers that run hundreds or thousands of instances in a Region.
Thus, Amazon charges two premiums for a dedicated instance – the $10 per hour flat fee incurred whether there are 10 or 100 instances, and the difference in price from the regular per-hour instance cost. To highlight a very simple example with a one-hour time frame, the difference between a small dedicated Linux instance in Virginia, and a small flexible (non-dedicated) Linux instance in the same region is:
Notice the difference in the per hour pricing, fractions in bold. As the blog stated, the $10 fee becomes trivial as the number of servers becomes very large. The annual per- instance premium paid and the annual "asymptotic" premium is then ("n" is the number of servers):
OK, that might be a bit hard on the eyes, but the lesson is that the yearly premium for dedicated servers converges on a constant price as the number of servers moves to infinity (see graph above).
As shown in the table below, this constant premium can be as high as $1,400 for extra large instances. When the number of servers grows, the extra $10 eventually disappears in the total cost of each instance (yearly cost is one instance running 24x7x365, or a duty cycle of 100%).
Note: click to enlarge table
It is well documented that many Amazon EC2 instances run on Intel Xeon E5345 processors. There are only a few servers on the market that are equipped with that type of processor, including Dell PowerEdge 1955 blades. The Large and High Memory Extra Large Instance are worth highlighting, because by comparing the RPE2 values of those instances to the RPE2 value of the Dell PowerEdge 1955 blade server, IDEAS determined that about one instance can fit on each box (this calculation is not straight forward -- see previous blogs for some analysis on one of many issues that can appear).
To make the previous point especially clear, if an end-user bought every Amazon EC2 server with the dedicated pricing model, they would be paying either $3,679 per box for a Large instance, or $5,431 per box, for a High Memory Extra Large instance. The revenue generated per server is clearly not a maximum either. For example, if each end-user only purchased 10 dedicated instances, then the cost per server would be much larger (see graph again). Those two prices then, are probably close to a minimum. Also, the networking, storage, support, and other charges would further add to the end-user cost, and Amazon’s revenue. Provided that only a very simple analysis is possible, if the total price tag of the cloud is three times the $5,431 High Memory Extra Large instance cost, and if Amazon had as many servers as Facebook (30,000) the revenue generated would be roughly $489M (USD).
That’s pretty close to that $500M figure, which enters the realm of credibility when all the "ifs" are satisfied.
However, this scenario assumes that end-users would be paying for all of these instances on a 24x7x365 basis, which would clearly work in Amazon’s favor, but in fact is likely to over-estimate the revenue. As mentioned earlier, the instance premium price is more likely the minimum amount that Amazon can collect for a server, which lowers the revenue estimate considerably. The 30,000 server number would require 500 fully-maximized racks of 60 PowerEdge 1955 blades, 60 per 42U rack. Another way to think about it is 10 rows of 10 racks, i.e. 100 racks per data center. The reader can decide on the feasibility of that issue.
All of this assumes that the total cost of the cloud to the end-user would be 3 times the compute cost, which would favor Amazon’s business model, but could be an over-estimate. Playing with these "ifs" can change the potential revenue to "iffy," yet even when the cards are stacked against Amazon, in a worst case scenario with the end-user only paying for half the year, 1/3 the number of dedicated servers, and a total cost multiplier of 2 (reasonable with storage), the revenue still stands at $53M (USD). That is not half-a-billion, but it is not what the public cloud naysayers were expecting either.
Comments