Public vs Private Cloud Cost Comparison

Infrastructure-as-a-Service Breakdown

There’s an interesting post by Advanced Systems Group that provides the top-level costs they were able to derive from an internal study they did to compare public and private cloud implementations.  It’s the start of a good discussion (questions/comments I posted are still in moderation).

Their post compares three different cloud options from a cost perspective: 1) Public Cloud from Amazon EC2; 2) Purchased Private Cloud infrastructure; and 3) Leased private cloud infrastructure.

All three systems were based around the equivalent of (using AWS parlance) 10 Extra Large CPU instances, 60 Large CPU instances, and 30 Small CPU instances.

The cost calculations were run by ASG over 3 years assuming an increase of 15% in compute demand, and not accounting for depreciation.

When ASG noted in an answer to a comment that all instances are assumed to be running 24/7 my curiosity was piqued to drill in a little bit.  The problem is, solely comparing private and public based upon 100% utilization removes one of the most compelling reason for the cloud - a pay-as-you-use model.  It skews the conclusion from the start.  (How many data centers run at 100% utilization anyway?)

Also, from a rough calculation, we can see that their Amazon numbers appeared to be based upon the cost of On-Demand Instances.  24/7 utilization over a three year period is not a cost effective use model for the On-Demand pricing model.

Performing a straight hardware costs calculation, (I’m assuming that ASG were adding in some additional costs for storage and data transfers to get to their top line numbers, as I couldn’t get there from a simple calculation.  Their 3Yr cumulative cost for the Public Cloud scenario was slightly over $1M, where as mine below came in slightly under), I got the following top line numbers for an On-Demand instance scenario (the least economical).  For completeness, I also added in the two more sensible scenarios of either a rolling 1 Year Reserved Instance model or a one-time 3 Year Instance model.

Public vs Private Cloud Costs (100% Utilization)

Public vs Private Cloud Costs (100% Utilization)

And so, taking these basic calculations in isolation of all other considerations, we can see that from a purely hardware resource expenditure perspective, a Public Cloud implementation is roughly the same or even less than a private cloud implementation.

However, this is still only part of the story.  As mentioned, a key draw to public cloud computing is the elasticity of the model - Pay-as-you-Use.  And so, considering a selection of rough percentage utilizations, we get the following breakdown of the different hardware costs of a public cloud implementation.

AWS Public Cloud Costs: Time vs Model vs Utilization

AWS Public Cloud Costs: Time vs Model vs Utilization

The comparison relative to utilization becomes even more stark when charted.

Cloud Cost Comparison vs Percentage Utilization

Cloud Cost Comparison vs Percentage Utilization

Not only is AWS more economical over utilization rates over extended periods, but it provides flexibility.  For example, if the software used on the servers is upgraded (as is likely) and hardware requirements increase (equally likely, given history), moving an image (AMI) from a Small to a Large instance is straight forward (for an “IT guy”).  No hardware needs to be thrown out or shipped in.

Additionally, there is some flexibility around migrating across business models.  Of course, it is easier to upgrade than downgrade ;)

Posted under cloud

This post was written by James Colgan on March 21, 2011

Tags: , , , , , ,


5 Comments so far

  1. resrpt March 28, 2011 12:29 am

    Thanks for the insightful piece. However, is this the correct way to define utilization? May be in cloud, but normally utilization is defined and given at a moment how much the system assets are getting utilized, rather than based on the time they are operating. So a server may be running 24/7 yet have a utilization of 15-25% (this is the norm in enterprise datacenters). The gospel truth is we can develop whatever TCO models we want to based on the assumption. One model can be used to prove that something is more costly, and the same model can be used to prove that the same thing is less costly by changing some underlying assumptions.

  2. James Colgan March 28, 2011 9:00 am

    Thank you Resrpt. Indeed, in this case, I’m considering utilization as a measure of how much a unit is used over a period of time. As you point out, when it comes to a piece of data center hardware, the server is active/on and being paid for (or amortized) whether utilization is 100% or zero. However, in the cloud, when the server is not used it is “switched off” and is no longer being paid for (and there’s nothing to amortize).
    Agreed, when making a TCO comparison, it is crucial to understand and model the expected use-model.

  3. PDC February 14, 2012 11:07 am

    Thanks very much for this analysis. Really helpful!

  4. Dave Kovatch March 8, 2013 10:08 am

    James, I think it is a given that if you have an actively managed environment where someone actually takes the time to monitor images for lack of use and turns them off.

    But in large organizations, say 100 servers or more, where the server population is not dynamic, my experience has been that a “Private Cloud”, or even a dedicated converged infrastructure using virtualization, is cheaper than Public Cloud. Would you agree?

    Granted, the ideal situation is to be able to merge the two, and have the Private/Dedicated environment for the stagnant uses and augment it with Public Cloud services.



  5. James Colgan March 8, 2013 11:32 am

    Indeed Dave, the greatest factor here is the degree of fluctuation in demand. If servers are being used constantly, a large number of servers are needed for a length of time (longer than 3 months, say), and you have the CapEx budget to spend, then a Private cloud will likely be the best way to go (assuming also that you have the IT expertise in-hand). But if there’s a great deal of variability in demand, regardless of the other factors, then the flexibility of the cloud would be the more cost effective. One other point to consider, cloud instance pricing is decreasing rapidly. (I think I read recently that AWS have dropped their prices 21 times since launch….so I should re-post this report with new numbers.) And so the cross-over point should continue to move in. What if we got to the point where processing was free and we only had to pay for storage? That’s an interesting idea don’t you think? By the way, switching servers on/off automatically when not being used can be done - we do that on the Xuropa Platform.


Leave a Comment

Name (required)

Email (required)



More Blog Post