Phil Wainewright penned an interesting analysis of what I had been calling Salesforce.com’s “Green Crystals,” but more commonly known as multitenancy. In the past I’ve said that the exact technology involved there was not really the important thing, it was just an interesting side story told for marketing purposes. Put another way, there are lots of ways to skin the multitenancy cat, but that isn’t even the important cat to be after.
At one point in his post, Phil frames the discussion in a particularly interesting light:
So are Salesforce.com’s green crystals just a marketing ploy, or is it really something that customers can’t easily replicate in their own private clouds, or by adopting rival cloud platforms?
I like this statement from a couple of angles. First, it asks the legitimate question of whether the non-SaaS world can learn some interesting things from SaaS technology. Second, it ultimately leads into a discussion of some technology around query optimization that Salesforce has patented, and whether those patents have given them an interesting monopoly now that will be denied others.
But before we delve into all that patented query optimization, let me make another controversial statement:
It’s just more green crystals and it is not the source of Salesforce’s advantage.
Why do I keep bringing up this silly green crystals notion? It is human nature to want a reason to believe. Marketing has trained us to look for the “silver bullet advantage.” Green crystals marketing is a term I first heard coined at Borland years ago when I was VP Of Engineering. Our products had loads of advantages over the competition, but marketing wanted one single reason they could tell customers why we had all those advantages, because its impossible to make a compelling sale off hundreds of small advantages. The term “Green Crystals” comes from soap marketing. Why is my soap better than yours? Because it has Green Crystals, that’s why!
Likewise, SaaS has a myriad of advantages over other delivery vehicles, and one SaaS vendor may have a myriad of advantages over another. But it helps to have Green Crystals. In the past, it was enough just to be a SaaS vendor, but the new breed had to differentiate themselves from the old hosted ASP model. Hence Marc Benioff brilliantly coined the multitenant argument. Now there are getting to be a lot of SaaS vendors with multitenancy, so maybe it’s time to coin something new, “patented query optimization,” anyone?
Whether you believe they’re just Green Crystals or not, their only advantage to the customer is that they enable SaaS companies to deliver their service at a lower cost. I have argued in the past that most of the savings is not due to multitenancy (or patented query optimization), but rather automation of the people costs. I stand by that argument, but I want to drill down on it further with some real numbers.
Let’s start by considering the cost to deliver the service for a few SaaS companies. Here, Salesforce really does have a tremendous advantage over most of its brethren. Consider these numbers taken from the latest 10K filings for these SaaS companies:
- Salesforce delivers $1 in revenue at a cost of about 12 cents to deliver the service.
- Concur delivers $1 of revenue at a cost of about 32 cents–almost 3x what Salesforce spends.
- Companies like Success Factors (about 35 cents) and Omniture (36 cents) have a similar much higher cost than Salesforce.
You can go analyze all the public SaaS companies, but the last time I did so, I don’t remember anyone doing better than Salesforce. They know how to deliver their service cheaply!
Now the next thing to look into is what the cost components might be. The point of multitenancy (and virtualization for that matter), is to share infrastructure as needed, amortizing it across a lot of customers, so that costs are saved through a reduction in waste. All of Salesforce runs on only 1,000 machines.
That sounds really impressive. Perhaps the competition is simply running 3x as many servers because of their inefficient multi-tenant architectures. Is that the source of the cost differential? Well, let’s start by figuring out what 1,000 machines cost. I’m a Cloud guy, so I would buy them from Amazon. I will assume that Salesforce can run their own data center at least as cheaply, so the cost of 1,000 servers on Amazon running 24x7x52 ought to be a conservative number. It comes to about $4.7 million, which is less than 4% of Salesforce’s cost to deliver their service. The competition would have to be using more than 5x as much hardware as Salesforce to account for that kind of cost differential from hardware alone. Now you see why I have a hard time believing the Green Crystals are telling the whole story.
What else may be at work here? It’s the cost of the operations people that drives the Lion’s Share of what it costs to deliver a service. Consider a fairly low-paid ops person who is technical (DBA’s and the like) may have a loaded cost of perhaps $120K per year. That cost is equivalent to 25 of those servers, BTW. So how many ops people does the gap in the cost of the hardware versus the overall Salesforce service cost represent? About 1,000!
Now I am not trying to say all of the cost differential is due to the cost of such ops people, but it does bear thinking about. Certainly there are other costs involved too. I have a friend running a SaaS company who privately says his Oracle licenses are the biggest cost after the personnel and ahead of hardware. Most SaaS companies didn’t build on Open Source technology, so they are going to have all sorts of software license costs. Still, I assert most of their cost is people.
Salesforce must therefore be radically more efficient at automating what it is taking other companies a lot of warm bodies to deliver. So if Phil’s corporate clients want to even begin to approach the efficiency of a Salesforce (let alone another SaaS vendor), they have to focus on that kind of automation.
How good can it get? That’s an interesting question. Salesforce is using 1,000 servers to deliver a service to 1.5 million people, according to the Techcrunch article I linked to above. Consider these numbers for Facebook that are taken from a great posting over at Diamond Notes:
- Facebook at that time (early 2008) had 1800 MySQL servers, 10,000 web servers, and 800 Memcached servers. They had a total of 12,600 servers, in other words.
- We’re hearing Facebook has 200 million users today, let’s assume they had 100 million a year ago. May have ben less, but we need to make a guess. That means they’re housing nearly 8,000 users per server, versus Salesforce housing about 1,500.
- Even more interesting are the figures on how many MySQL DBA’s each company in that survey uses. Facebook has just 2 DBA’s keeping 1800 servers happy! That’s a couple orders of magnitude better than anyone else in the survey. Clearly Facebook has been very successful in automating a lot of operations work.
All of these factors will play a role. However, given that a single operations person is equivalent to 25 servers in cost, and given that we have the evidence of companies like Facebook, massive automation of the manual tasks is the big cost reduction opportunity moreso than making more efficient use of the hardware.
Imagine being able to keep Salesforce’s 1,000 servers humming with just 10 or 20 ops people. Can corporate IT ever get there? Sounds very hard to me when you consider all the advantages a company like Salesforce has over Corporate IT. If nothing else, they can change the software to facilitate the automation instead of having to go in after the fact with third party tools.