Friday, May 23, 2014

Bezos' Law

Greg O'Connor had a piece at Gigaom entitled Moore's Law Gives Way To Bezos' Law sparked by the recent price cuts by Google and Amazon in which he claimed that:
The latest cuts make it clear there’s a new business model driving cloud that is every bit as exponential in growth — with order of magnitude improvements to pricing — as Moore’s Law has been to computing.
If you need a refresher, Moore’s Law is “the observation that, over the history of computing hardware, the number of transistors on integrated circuits doubles approximately every two years.” I propose my own version, Bezos’s law. Named for Amazon CEO Jeff Bezos, I define it as the observation that, over the history of cloud, a unit of computing power price is reduced by 50 percent approximately every three years.
Both Moore's and Kryder's laws held for multiple decades. Below the fold I ask whether a putative Bezos' law could be equally long-lived?

The article follows Google's announcement:
Google was first to announce “deep” cuts in on-demand instance pricing across the board. To make the point that cloud pricing has been long overdue, Google’s Urs Hölzle showed in March just how much cloud pricing hasn’t followed Moore’s Law: Over the past five years, hardware costs decreased by 20 to 30 percent annually, but public cloud prices fell by just 8 percent annually:
Google's graph shows the cost of hardware dropping rapidly over the past 8 years, and much less rapidly over the next 8, until early in the next decade it is essentially flat. This is realistic; while there may still be economies to be wrung out of system architecture, data center design and operations it is clear that the major drivers of cost decreases, Moore's and Kryder's laws, have flattened out. O'Connor makes the casual assumption, belied by the graph he references, that because costs have dropped exponentially in the past they will continue to drop at the same exponential rate in the future.

Patrick Sinz on Dave Farber's IP list provides the appropriately skeptical response:
But nevertheless the numbers in the article tend to demonstrate that the cloud providers margins have a good tendency to grow (per compute unit) and this would tend to make them LESS attractive not more ....

I do agree though that we'll migrate "en masse" to the cloud, but not because it's a "good" solution, but because it's "almost good enough", and that the real advantage of cloud provider is that they provide "prime network real estate". If you want to distribute data and service globally you want to have compute units "close to your users" (in terms of ping latency) that are globally distributed.

But the sad reality for the customers will be that in exchange for a lower initial investment we will end up paying more, and have less choice and no control over our infrastructure.

4 comments:

David. said...

Wolf Richter comments on impact of "the cloud" on IBM. Tip of the hat to Yves Smith.

GregO said...

David:

I fail to see what it is a "casual assumption". The data is there that show how the numbers are generated.

t that cloud pricing has been long overdue, Google’s Urs Hölzle showed in March just how much cloud pricing hasn’t followed Moore’s Law: Over the past five years, hardware costs decreased by 20 to 30 percent annually, but public cloud prices fell by just 8 percent annually.

He does not show data. If you did deeper this is the price for a given machine at amazon and does not account for the period that the compute increased 65%. The metics are for different things. Mine is clear and there is data. Google is marketing and the compute power is not fixed for the comparison period.

David. said...

GregO, as I wrote in the post: "O'Connor makes the casual assumption ... that because costs have dropped exponentially in the past they will continue to drop at the same exponential rate in the future."

For "Bezos Law" to hold for decades as Moore's and Kryder's laws have, which is what I was discussing, the underlying technology costs need to continue to drop exponentially for decades. Chip and disk industry experts believe that the rate at which both compute costs and storage costs will drop over the next 5 years or more will be much lower than it has been in the past. There are very strong reasons for believing this; related to the costs of manufacturing chips and disks. So Bezos' Law is unlikely to hold for anything like one decade, let alone the multiple decades of Moore's and Kryder's laws.

That is the argument of my post. Refuting it would require some argument as to why compute and storage costs will, against industry projections, continue to drop as fast in the future as they have in the past. The GigaOm piece does not provide any such argument, it just assumes that Moore's law continues.

David. said...

The wonderful Charlie Stross explains, in a different context, why doing business with Amazon is a trap. Charlie points out, as he has done before, that Amazon's business model in other areas such as books has been to subsidize sales to drive competitors out of business and then to raise prices.

There is no reason to believe that this isn't their model in the cloud. If it is, the rush into the cloud in search of cost savings will turn out to be another example of the pervasive short-termism identified by Andrew Haldane and Richard Davies of the Bank of England.