authors:
- Juan Camilo Castillo, Assistant Professor of Economics, University of Pennsylvania
- Amit Gandhi, Professor of Economics, University of Pennsylvania
This blog article is derived from the authors’ paper titled Competition in Cloud Computing, a project of the Economics of Digital Services (EODS) research initiative led by Penn’s Center for Technology, Innovation and Competition (CTIC) and The Warren Center for Network & Data Services. CTIC and The Warren Center are grateful to the John S. and James L. Knight Foundation for its major support of the EODS initiative.
__________
Businesses used to have a hard time making decisions about IT investment. Companies had to buy and install large servers, which involved long lead times, and they had to pay upfront for expensive software licenses. For that reason, they had to anticipate how their businesses would evolve years in advance to make the right decisions.
Cloud computing has completely changed how companies think about IT. Since Amazon Web Services launched in 2006, businesses can simply rent IT infrastructure. If companies grow suddenly, they can expand their computing capacity with only a few clicks. If they want to experiment with new technologies, they can do so at a small cost without having to set up expensive infrastructure. Businesses ranging from technology startups to well-established blue chip companies have thus been avid adopters of cloud computing, which is already a $270 billion market—accounting for over 10% of global IT spending—and has been growing over 20% per year.
Given how important cloud computing has become to all kinds of businesses, it is important to ensure that the cloud market delivers as much value to customers as possible. One reason why it might be failing to fulfil its potential is a high degree of concentration. There is one dominant player (Amazon Web Services) that has around 48% market share, and a few secondary players have 16%, 8%, and 4% market shares (Microsoft Azure, Alibaba Cloud, and Google Cloud Platform, respectively)—similar to other technology markets like search (Google), social networks (Facebook), or e-commerce (Amazon). However, this is a market where the main good that is transacted is in essence a commodity: all cloud providers rent out products produced by the same semiconductor companies like Intel or AMD. This would suggest that it would be hard for dominant players in the industry to charge high prices because competitors would be able to undercut them to capture customers.
Our research found that prices do not behave as they would in a perfectly competitive market. Why is that the case when competitors offer almost identical products? One reason might be that they do not only sell basic cloud infrastructure but a full bundle that includes software, personalized customer support, and several tools to optimize the customer experience.
Methodology
To get a sense of how much of a problem concentration is in cloud markets, we started by collecting data on the prices of Amazon Web Services (AWS). By querying APIs in their website, we were able to construct a dataset of all historical prices between late 2015 and today. We then did a few statistical analyses to see whether the price patterns we observe resemble what one would expect if, in fact, cloud computing behaves like a commodity market. In that case prices would drop over time as the prices of semiconductors decrease according to Moore’s law, allowing other competitors to lower prices.
The first pattern that we observed in the data is that price changes are very infrequent. Most of the products in our data have no price changes over the whole five-year period we analyzed. To see this more clearly, we constructed three price indices, one for each of the main categories of cloud services offered by AWS: virtual machines (computers that lives in the cloud), disk units that can be used by virtual machines, and storage, a service that closely resembles Dropbox or iCloud. Figure 1 shows that those price indices drop somewhat over time but nowhere nearly as much as they would if they followed the same trend of processor prices.
Figure 1: Price Indices for AWS. Reductions are relative to the levels in December 2015. The three series correspond to virtual machines (EC2), disks (EBS), and storage (S3). The dashed black line shows to the trend prices would follow if they decreased at the same rate as processor prices
We also looked into the price trends of individual products to get a better sense of the patterns behind Figure 1. Figure 2 focuses on the prices that a customer would get for products in North Virginia. There is nothing particular about this region; other regions show almost identical patterns. Each subfigure focuses on all possible products within a family of products. Subfigure (a), for instance, shows prices for all general purpose virtual machines. m1 is the oldest generation of general-purpose virtual machines offered by AWS, whereas m6g and m6i use more advanced processors that were not launched until the last two years.
As we already mentioned, there are very few price changes. In fact, the only product that had price changes was m4. As AWS introduces newer generations, it offers them at lower prices than older products, and, except for m4, it never changes the original price level. Subfigures (b), (c), and (d) show a very similar pattern for other families of products. All these price drops, however, are much smaller than the kind of price drops observed in the processor market.
Figure 2: Log price behavior over time for a few select products. These figures show the price evolution for AWS virtual machines’ key products in the region US East (N. Virginia). The dashed black lines represent a trend decreasing at an annual rate of 43%, the rate at which processor prices drop.
Conclusion
In summary, these figures show that prices do not behave as they would in a perfectly competitive market. Why is that the case when competitors offer identical products? One reason might be that they do not only sell basic cloud infrastructure but a full bundle that includes software, personalized customer support, and several tools to optimize the customer experience. Once a company is used to MS Azure’s products, for instance, it is hard for it to switch to AWS.
Future steps
We will further explore these mechanisms and to what extent they might hurt consumers. With that goal in mind, we are expanding our dataset to include historic prices for MS Azure and Google Cloud Platform. Furthermore, we will get access to data about MS Azure customers’ usage patterns, which will vastly expand the scope of the statistical analysis we can perform. We will build a model of the cloud market that includes customers choosing among different cloud products as well as providers that choose how to set prices. We will then be able to simulate how the market would behave under different circumstances, such as with more or less competition, and we will be able to measure to what extent that might hurt or benefit customers. This will provide key inputs for regulators that wish to understand to what extent concentration is an issue in cloud computing.