Last week we interviewed a Former Finance Manager at Amazon, who was partly responsible for collating country and retail product category financials into one consolidated P&L for Brian Olsavsky, AMZN’s CFO. 

Our aim was to deconstruct Amazon’s Retail P&L from the consolidated accounts to better understand its e-commerce unit economics. Members can read the full analysis and how we estimate the gross, contribution, and long-run EBIT margins of AMZN retail here. 

This analysis aims to explore AWS’ potential long run and current normalised FCF margin.

Each quarter, AMZN reports only three core KPI’s for AWS: backlog, revenue, operating income. For a business generating over $80bn in revenue, it’s incredible how AMZN gets away with such poor disclosure. We have no details on revenue segments, opex lines, or working capital movements for the core asset arguably driving most of AMZN’s enterprise value today. 

Using last week’s estimates of AMZN Retail’s opex lines, we decouple AWS’ operating expenses:

AWS’ tech and content line includes infrastructure costs, and all D&A and SBC associated with the line item. Given the rapid growth in data center square footage and a temporarily elevated D&A expense as a percentage of revenue, a 64% gross margin for AWS seems conservative. 

Digital Ocean, a smaller public cloud provider for SME’s, has a 65% gross margin at 1% of AWS’ size. 

The challenge forecasting AWS’ long-run EBIT margin is that EC2 and S3, estimated to be ~75% of AWS’ revenue, is a different margin profile than PaaS services like Redshift, Amazon’s data warehouse product. Revenue mix heavily impacts margins. 

For example, what happens to AWS margins when Redshift loses revenue to Snowflake?

If we assume Redshift has SaaS-like 75% gross margins, AWS will be losing high-margin revenue to Snowflake, but gaining more compute and storage revenue. Greater EC2 and S3 revenue fills data center capacity and will drive higher gross margins. The net margin impact is unclear. 

We recently discussed how AMZN’s physical assets are potentially the real moat. And maybe AWS doesn’t mind losing Redshift revenue as long as it runs SNOW’s compute. 

The implicit assumption is that owning the rails for cloud computing is a more durable business than the software offered on top. After all, software tends to move much faster than hardware. 

For example, Snowflake rapidly took market share from Redshift, and now Databricks is ‘melting the snow’ with a best-in-class lakehouse. We recently asked a Databricks VP what the company’s biggest risk is. 

His answer: open-source table formats like Apache Iceberg and Delta. Software moves rapidly. 

I would say I think the table format wars, Iceberg and Delta, is extremely important...a way for you to leverage cheap object store, like S3 or Google Cloud Storage; it's a file destination where you can store structured, unstructured, and semi-structured data. The way for it to act like a data warehouse or to have performance is to put it into a table format. The most commonly used table formats are Delta Lake, Apache Iceberg, and Apache Hudi. Our motion is that we are very open-source first and friendly, so we have several open-source projects that we've released that have gone mainstream and wild. It's great for us..That's been our special sauce. The table format wars are substantial, and Iceberg is up there. In benchmarking done by external companies, they've consistently seen the Delta Lake file format be more performant across the board. - VP, Databricks

There are three major risks for AWS’s long-run margins: losing high-margin software revenue, heavy price competition with Azure / GCP, and EC2/S3 being displaced as a cloud computing platform. 

Rapid innovation, typically from open-source projects, puts AMZN’s microservices on top of EC2 and S3 at risk of obsolescence. AWS’ ‘good enough, me-too’ services will always have a market, but losing high-margin revenue to best-in-class ISV’s poses a long-term risk to AWS’s margin. If AMZN can partner with the best ISV’s and host the service revenue it may lose, the net impact may be negligible. 

Another risk is that AWS is that the cloud compute pricing converges to the cost of capital due to heavy competition between the three major players. This makes the assumption cloud computing is a commodity and even in an oligopoly the fixed asset base cannot earn an economic profit. This seems unlikely to us due to the scale and expertise required to run the infrastructure and the cost and time savings to customers. 

What seems more likely is that AWS loses significant market share to Azure. MSFT has a long tail of legacy Windows customers and a huge, experienced sales team to help win the incremental dollar moving to the cloud. 

Another under-discussed risk to AWS is the positioning of Github. MSFT now owns the default code hosting platform and leading social network for developers. If MSFT can find a way to allow developers to simply push code directly from Github to Azure, it will save a lot of time and effort for developers. This would also take share from all PaaS that use AWS and could limit Amazon’s share of startup infrastructure spend going forward. 

Beyond losing high margin software revenue and IaaS market share, the biggest terminal value risk is that EC2 and S3 itself is disrupted. 

Could more effective storage primitives or new ways to run compute on the edge neutralise AWS’ advantage?

For example, our engineering team has been experimenting with Fly.io, a service that hosts applications on top of physical dedicated servers they run globally. A new type of public cloud. Fly owns the hardware and aims to deliver applications all over the world, wherever the customer may be. It’s a cheaper and more effective way to deliver compute relative to EC2. Fly recently raised $25m led by Andreessen Horowitz and is small fry now, but is an example of innovation that may threaten AWS’s core service. 

With AWS scale, Graviton’s progress, and Amazon’s experience, it seems harder to truly disrupt the storage and compute services compared to any software service on top. But this is a major risk we’re investigating over the next few quarters.  

Putting aside longer term competitive risks, in AWS’ current form, what is the normalised FCF margin?

One variable that is somewhat overlooked is the working capital profile of AWS. For AMZN’s first 20 years of existence, it enjoyed favourable negative working capital where retail customers pay well in advance of vendors receiving payment. 

The chart below shows how AMZN’s negative cash conversion cycle has reduced from ~30 days of cash on hand to only 17 days in FY21. We believe this is largely due to the growth in AWS and the nature of selling on-demand instances at discounted prices to large enterprises. 

We estimate AWS’ eats 5% of sales as it grows in working capital due to longer receivables terms from enterprise customers. 

For simplicity, we include 50% of principal lease repayments and in AWS capex, exclude any spend on build-to-suit leases, and exclude SBC (we estimate 60% of the consolidated SBC is for AWS) to find AWS's current FCF margin in high-growth mode today. We’ve excluded tax given AMZN pays tax at corporate level for both retail and AWS. 

It’s also interesting to note that SBC as a % of total AMZN sales is <4%, but ~15% for AWS. Not crazy by any SaaS benchmark, but still a significant cost worth considering. 

To find normalised FCF margins, we have to net out the growth capex. 

We recently interviewed a Former AWS Director who ran 33% of its hyperscale data center footprint and estimates long-run server life is closer to 7 years than its current 5 years. With networking equipment, buildings, and other longer-lifetime equipment in AWS PPE, we estimate the normalised capex based on a 7-year asset lifetime. 

Again, we deduct lease principal repayments and SBC to find a normalised FCF margin for AWS at ~30% over the last 2 years. 

If we assume $80bn revenue grows at 25% for 2 years, AMZN is currently trading at ~42x normalised AWS FY24 FCF. 

The 7-year ROIIC is 31.4% and, depending on how much net working capital AWS employs, it’s earning 25-30% return on its capital. 

This is a fine business. But as we discussed, the longer-term competitive questions for EC2 / S3 and all the services on top remain. It’s our mission to share insight into such risks over the following quarters. 

One last interesting question worth considering: what % of AWS revenue is generated by Amazon.com? And what are the internal transfer pricing arrangements?

I would assume 5% to 10% of AWS’s total revenue is attributable to the consumer business. Everything is run on it; all the servers. AWS was built by the consumer business because it needed those services. - Former Finance Manager, AMZN

Our first thought was that Amazon.com would be getting compute at cost. If 5-10% of AWS revenue is from Amazon.com at cost, this compresses short term EBIT margins and long-run operating margin could be north of 35%.

But there is no breathing room for Amazon.com - if Amazon.com gets charged at cost, it doesn’t encourage a relentless push to be the most efficient retailer globally! 

I would probably not assume that [AWS] is charged at cost…because of frugality. It’s all about applying pressure. But at the same time, it is so beautiful because you can really learn how to run a business. - Former Finance Manager, AMZN

AWS: Long-Run and Normalised FCF Margin Estimates

In Practise Analysis

inpractise.com/articles/aws-normalised-fcf-margins

Related Content

AWS Series: Hyperscale Data Centers, Physical Moats, & Server Lifetime

AWS: Containers, Kubernetes, & EC2 Stickiness

Former Senior Manager at Amazon Web Services

Amazon Web Services: Data Center Economics

Former Director at Amazon Web Services

AWS Series: Kubernetes & The Portability Illusion