...in May, about two years after MemSQL was founded, [Eric] Frenkiel and company came down from the Amazon cloud, moving most of their operation onto a fleet of good old fashioned computers they could actually put their hands on. They had reached the point where physical machines were cheaper — much, much cheaper — than the virtual machines available from Amazon. “I’m not a big believer in the public cloud,” Frenkiel says. “It’s just not effective in the long run.”
The post goes on to describe a number of other companies that have reached the same conclusion: At scale, it simply is not economically viable to layer your SaaS or PaaS on top of someone else's rented infrastructure. Oddly, the disproportionate cost of cloud infrastructure is well known to all the startups that I know who are using it. Everyone is moving off of Amazon as far as I can tell. If you run the numbers, it is also pretty clear that AWS is likely reaping quite a fat profit from their operations.
In my case, I was forced to run the numbers because no one believed me. Not a problem. I love math. Here's what they looked like based on pricing available mid June of this year: I could get managed physical server hardware, colocated in several large data centers that would run 20 AWS Linux instances of my specification, at a third the cost. That included the VMware license. Even better, if you wanted to buy the hardware yourself and just rent the rack space, power, and connectivity, it got even cheaper. Based on my math, AWS only makes sense for a small number of compute instances, or when running for a short period of time. One more thing: The two approaches are headcount-neutral. There is no meaningful difference in the number of people and internal operating expenses that go with one way versus the other.
When you step back from this, you realize the big obvious truth: Nowhere in nature does renting cost less than buying: Not with leasing cars. Not with apartments and real estate. Certainly not with compute resources. Renting has always been a way to avoid major capital expenditures - either because the capital does not exist, or the need is short term. The longer and larger your commitment, the more advantageous it is to use capex if you have it available to you. In the case of the SaaS providers, the name of the game is to reduce service costs. Why let your service providers expropriate your profits if you don't need to?
What to rent? Buying a giant data center in a large metro is insanely expensive. It makes sense to rent space in someone else's. Renting someone else's PowerEdge R720 maybe does not.
Anyway, I feel better now that I don't feel alone...
There are 2 other interrelated considerations - CapEx vs. OpEx and available capital.
ReplyDeleteMany startups - including both Zynga and MemSQL mentioned in the article - bootstrap using AWS because it takes very little available capital, which is a very precious commodity.
Conversely, in the enterprise, CapEx and OpEx come from different budgets, and have different approval hierarchies. Depending on the organization, one may be easier to spend than the other, which has a major influence on the financing method chosen.
The same applies to cars and houses. When one have limited capital, renting is the best option, even if it costs more in the long run.
In my case, the discussion was framed around how to spend a series A investment. Hence conservation of capital was paramount. Cash flow becomes more important in a growing business. I guess that means that AWS may be a barometer for the health of the startup economy?
Delete