How Amazon’s cloud business generates billions in profit

Every year since at least 2014, more than half of Amazon’s operating profit has come from the online retailer’s cloud division, Amazon Web Services, which provides online services and tools that software developers can stitch together to run websites and applications.

It’s an impressive business in absolute dollar terms, not just percentages: AWS ended 2020 with $13 billion in operating income, which helped Amazon report total net income of $21 billion for the year. No wonder Amazon chose the head of AWS for 15 years, Andy Jassy, as its CEO when Jeff Bezos stepped down earler this year.

Alibaba, Google, IBM, Microsoft and Oracle are all looking to challenge AWS. And unlike the Alibaba and Google clouds, AWS is profitable.

In the second quarter, Jassy’s final quarter as cloud chief, AWS had $4 billion in operating income, up about 25% year over year and ahead of analysts’ estimates. Revenue grew 37%. Adam Selipsky, who sold data analytics software company Tableau to Salesforce after 11 years at AWS, took Jassy’s spot in July.

AWS is like a buffet, serving developers and end users with hundreds of services, and it comes out with new ones each year. From a financial perspective, it’s a bit of a black box: Amazon doesn’t provide information about where AWS revenue comes from, nor does it call out the segment’s most profitable parts..

But AWS got its start as an outsourced provider of the most basic types of computing functions — computing power and data storage — and that’s likely where most of its money still comes from today.

Corey Quinn, who helps companies lower their AWS costs as chief cloud economist at the privately held Duckbill Group, estimated that more than 50% of AWS revenue comes from the EC2 computing service. It essentially lets customers rent virtual slices of physical computer servers in Amazon data centers.

Add in the Elastic Block Store and Simple Storage Service data-storage services, the Relational Database Service for storing and serving up data, and data-transfer fees, and that brings the total above 70% of revenue, Quinn said.

Nowadays, however, Amazon increasingly emphasizes more complicated, higher-level capabilities, which offer higher margins and can make it harder for customers to switch to other providers. In May, for example, Selipsky tweeted about a new service called App Runner that deploys applications in virtual containers, saying that it could help clients “focus on building their businesses, not managing #containers & #infrastructure.”

Analysts at Bernstein in June estimated that in 2015, 14% of AWS revenue came from higher-margin “platform as a service” products, while the rest coming from lower-level “infrastructure as a service” products. In the first half of 2020, the analysts estimated that the platform percentage had risen to 18%.

Revenue and profit are probably not evenly distributed among the millions of AWS customers. If 20% of AWS customers deliver 80% of revenue, then the widest margins are coming from the small and medium-sized customers in the remaining 80%, said Joe Kinsella, founder and former technology chief of CloudHealth, a start-up VMware acquired in 2018 whose software helps companies tune their cloud usage. Larger customers can access more meaningful discounts.

Bernstein’s analysts estimated that in 2019 AWS had a 61% gross margin overall.

Amazon declined to comment for this article.

EC2 is one of the oldest AWS services and it still stands out as the major driver of gross profit and operating profit, said Bhavan Suri, co-head of technology equity research at William Blair, which has the equivalent of a buy rating on Amazon stock.

The EC2 service might have a gross margin percentage in the mid-50s, Suri said. Amazon does not disclose margins for individual AWS services, although the AWS operating margin was around 31% in the first quarter. Gross margin doesn’t reflect the cost of research and development and marketing.

A smaller cloud infrastructure provider, DigitalOcean, reported having a 58% gross margin in the second quarter, and much of the company’s revenue comes from virtual machines that are similar in concept to EC2 instances. A DigitalOcean spokesperson said about 90% of its revenue derives from computing, storage and network services.

Customers can lower their bills by reserving EC2 instances for future use. The AWS website says Reserved Instances can be as much as 75% cheaper than on-demand instances. Another option is using EC2 spot instances, which are spare computing resources Amazon has available and can reclaim quickly when the need arises. AWS advertises Spot instances as being up to 90% less expensive than on-demand instances.

Organizations can leave their applications running as they are on AWS, or they can lower the cost of operating them by switching out the underlying components. When AWS introduced EC2 in 2006, there was just one instance type, known as m1.small.

That variety remains available — “we’re not a company that deprecates stuff,” said David Brown, the vice president in charge of EC2 — but there are now over 300 types of instances. It’s not easy to stay on top of the announcements about new instances, said Brooke McKim, co-founder and technology chief of start-up Vantage, which offers a tool for cloud cost monitoring.

AWS customers can use EC2 instances that draw on chips from AMD, Intel and Nvidia, as well as artificial-intelligence processors and power-sipping Arm-based chips that Amazon designed in house. AWS itself is using those Arm chips “all over the place across AWS where it makes sense” to deliver its services to customers, Brown said. That lowers Amazon’s own costs. Amazon’s’ closest competitor in the cloud, Microsoft, has not talked about doing anything similar.

One of the more well received product introductions in recent years is the Lambda service, which runs computing jobs in response to certain events, such as automatically creating a new thumbnail image when a person uploads a new photo to a social network. Developers call this type of tool “serverless” because cloud providers don’t require them to manage any server infrastructure personally. When customers use Lambda, AWS can wind up charging almost double the price of EC2 instances, analysts at William Blair wrote in an August research note.

Another early AWS service was the Simple Storage Service, or S3, which debuted in 2006. Customers use S3 to store objects, including photos, videos and other types of files.

One early user was Don Alvarez, who worked at start-up building collaboration software for filmmakers in the Seattle area. Back then, companies rented servers in colocation facilities to run websites, and they had to pay for enough equipment to meet their expected capacity needs. Start-ups always ended up buying more than they needed, Alvarez said. Amazon met with him and described the service it had come up with, then provided pre-release access. “I knew in that first moment, AWS gave me something that changed my life, and that I knew would change my life and the lives of all developers,” he said.

That assessment proved to be correct. S3 has reached massive scale, now storing more than 100 trillion objects.

“We have individual customers who have hundreds of petabytes, in some cases an exabyte of storage or more,” Mai-Lan Tomsen Bukovec, the vice president in charge of object and block storage, told CNBC in an interview.

Companies can end up storing more data than necessary in S3. Kinsella recalled that one CloudHealth customer discovered that it was indefinitely depositing log files in S3 by accident. When it fixed the problem, it ended up saving $1.2 million a year Last year AWS itself introduced a tool called S3 Storage Lens that can help organizations uncover unnecessary spending. The service was adopted very quickly, Tomsen Bukovec said.

Other cloud providers sometimes offer pricing incentives to attract business, but customers are often much smarter than that, she said. For example, she talked about the ability to run a computing job right as a customer submits a request to retrieve an object, without having to deal with the underlying infrastructure.

One competing cloud storage provider is privately held Wasabi. Rather than offering a huge range of services like AWS, Wasabi focuses on only data storage, and it charges about one-fifth the price of Amazon S3, said David Friend, Wasabi’s CEO. But he doesn’t think Amazon’s cost of running the business is significantly higher — in other words, he believes S3 is a cash cow for Amazon.

“I’d be surprised if their S3 gross margin were lower than 70%,” Friend said. Other estimates are more modest: William Blair’s Suri estimated that it’s in the low-50% range.

Moving data into Amazon S3 is free. But transferring a terabyte of data out of S3 can cost three times more than the monthly cost of storing that terabyte in S3. These charges can make AWS bills unpredictable, said Friend from Wasabi (which does not charge for data transfer). It also creates a strong incentive to stay with AWS instead of moving data out to competitors.

“Almost everybody knows how much data they have. They have no idea how often they touch it,” he said. One customer, he added, spent $6 million in egress fees to move data from S3 to Wasabi.

Many AWS customers also rely on the Elastic Block Store, which can be attached to EC2 computing instances and hold data for them. Storage volumes from this service can remain available even after customers delete the EC2 instances. AWS offers the EBS service on hard disk drives for as low as 4.5 cents per gigabyte each month.

The EBS gross margin could be between 60% and 70%, Suri said.

The AWS product lineup includes several databases to choose from.

The most traditional approach is to use a relational database, which organizes data into tables with columns and rows. Programs can store information in databases and then request information from them. AWS’s Relational Database Service offers a few options: Proprietary but popular database systems from other companies, such as Microsoft’s SQL Server and Oracle’s flagship database software; open-source flavors such as MariaDB, MySQL and PostgreSQL; and Amazon’s homegrown Aurora, which is compatible with MySQL and PostgreSQL.

Using Aurora through the Relational Database Service might yield a gross margin percentage in the ballpark of high 60s or the low 70s, Suri said.

The start-up MariaDB, backed by the likes of Alibaba and Intel, offers a cloud version of its commercial version of the open-source MariaDB database. The start-up’s cloud service is half the price of the AWS Relational Database Service in MariaDB mode, said Michael Howard, the company’s CEO, in an interview.

“When you have absolute power, which Amazon does in…the infrastructure space, you are going to leverage it,” he said.

AWS doesn’t have a dedicated service for transferring data across its infrastructure or out of its infrastructure. Instead, it charges a wide range of fees for different types and methods of data transfer, and these fees can add up.

DigitalOcean today charges less than AWS for outward data transfers. McKim worked there before starting Vantage earlier this year, and he knows how much the company and its rivals were paying for network bandwidth. “Usually with bandwidth, the more you buy, the cheaper it is,” he said. “They [AWS] would be able to offer it even cheaper than DigitalOcean was. But they weren’t offering it cheaper.”

These network charges can be a major source of unpredictability.

For instance, NASA started a five-year, $65 million contract with AWS in 2018, but the Inspector General reported that the agency had not come up with cost predictions for network egress, which happens when people try to download files stored in S3. The Inspector General warned that it’s possible that scientific information could become less readily available if NASA puts limits on data egress to manage costs.

In 2018 Cloudflare, which provides fast access to data using infrastructure located around the world, announced the Bandwidth Alliance, a group of cloud and hosting providers that were willing to lower or eliminate data-transfer costs. Amazon has not elected to participate, said Cloudflare CEO Matthew Prince.

The cost of bandwidth has fallen each year since the 2018 announcement, and yet AWS has kept charging the same amount for data transfer in the U.S. and in Europe, where the price is perhaps 80 times the underlying cost, Prince said.

“The markup for bandwidth at AWS is unlike anything that exists anywhere else at Amazon,” he said. He guessed that the service has at least a 99% gross margin.

That sounded too high to Suri, but he said he wouldn’t be surprised if the gross margin were above 80%.

Increasingly, companies with large AWS commitments are going beyond spending that money by building on basic elements such as EC2 and S3, and they’re also buying software from other companies through AWS’ Marketplace. That benefits AWS. Analysts at UBS estimated in May that the average listing fee Amazon charges is around 5%. Microsoft lowered its marketplace fee from 20% to 3% in July.

Mobile app stores charge higher fees. Google’s operating margin for running its Play Store, which takes 30% of revenue from purchases, was above 62% in 2019, Reuters reported in August, citing an unredacted court filing.

The operating margin for AWS Marketplace “might be well above the overall AWS level of ~30%,” wrote the analysts, who have a buy rating on Amazon stock.

Marketplace is still small compared with EC2 and S3, though. The analysts said Marketplace could be delivering $1 billion to $2 billion in revenue, which would work out to 3% of AWS total 2020 revenue.

WATCH: Amazon Web Services, Dish team up for cloud-based 5G network