Now is the season we gather to be awestruck by the immense CAPEX spending of the hypercloud companies (across all their businesses). Previous installments available for 2016, 2017, 2018, 2019, 2020 plus earlier/other CAPEX musings.
The three companies with hyperclouds – Amazon, Google, and Microsoft – collectively spent over $124 billion on CAPEX in 2021, up 28% from 2020’s almost $97 billion.
As a semi-useful comparison that doesn’t make the common error of conflating a stock with a flow, that spend is nearly the size of Morocco’s GDP, the 60th largest economy in the world (on a combined revenue basis, these three would be the 19th largest economy, just behind the Netherlands).
Amazon’s and Microsoft’s CAPEX spending again hit new highs (double bonkers new highs in Amazon’s case), while Google’s spending is still below their most recent 2018 peak.
In 2020, Amazon spent more on CAPEX than any other company I checked. For 2021 I’m not even going to look, as Amazon’s CAPEX was up 37% to $73.7 billion (note the majority of Amazon’s spend is things other than cloud infrastructure).
It is nice the semiconductor guys are rediscovering the old time CAPEX religion, but they’re still pikers compared to Amazon (and both Google and Microsoft will likely still outspend very-proud-of-their-increased-spending Intel this year).
Microsoft was up 26% to $25.9 billion while Google increased 11% to $24.6 billion. 2021 is the first year Microsoft has outspent Google on CAPEX since 2009.
Standard caveat: The reported numbers are the companies’ total CAPEX spend, not just cloud infrastructure, so includes land, office buildings, campus redevelopments, warehouses, panopticonvenience stores, manufacturing tooling, self-driving cars, delivery vehicles, flying machines, flying delivery machines, satellite constellations, hardware that both is and is not required for quantum computing, and – what should be the absolute top priority for Congressional hearings – the missing-in-action Google space elevator. The numbers include finance leases for both Amazon and Microsoft, as well as build-to-suit leases for Amazon (these are debt instruments used to finance specific CAPEX expenditures, namely servers and buildings).
The three hypercloud companies’ cumulative CAPEX spend since 2000 is over $586 billion, with over $221 billion of that occurring in the last two years. In 2024 these three companies could collectively surpass a trillion dollars in cumulative CAPEX spend. As the saying goes, a trillion here, a trillion there, pretty soon you’re talking about real money.
As a percentage of revenue, Amazon CAPEX hit a new high of 16%, Microsoft hit a new high of 14% and Google logged their lowest ratio since 2012 at 10% (but I am expecting big CAPEX things from them this year as discussed below).
I described Amazon’s 2020 CAPEX spend as “#bonkers” (which I am glad to see has caught on as the adjective of choice for vigorous CAPEX spending), last year accelerating 69%, ahead of the company’s 38% pandemic tailwind revenue growth.
The theory their CAPEX might slow down in 2021 to digest a little did not pan out. At all. CAPEX again grew at almost double the rate of revenue (37% vs 22%), even as retail revenue growth almost halved. So we’ll go with #bonkersbonkers this year. You can really see the step function after 2019 in the chart above.
It gets more and more tenuous to compare Amazon’s overall CAPEX spending to Google and Microsoft, because Amazon spends massively on their retail fulfillment infrastructure. The retail business is very likely the majority of Amazon’s CAPEX spending. Google has told us that 60-70% of their CAPEX goes to data center infrastructure, varying based on their current trophy real estate mood. Microsoft is likely comparable. Even if AWS is twice the size of Azure, it is hard to believe AWS’s data center infrastructure (and spending) is even twice the size of Microsoft’s, given Microsoft has other large services living in its data centers.
So let’s speculate about how much of Amazon’s CAPEX goes to cloud infrastructure. Amazon gives us a few indirect clues in their SEC filings, in which they reveal square footage by business segment (North America vs. International vs. AWS) and type of facility (office space vs. physical stores vs. “fulfillment, data centers and other”). So they conflate AWS office space with data center space and obscure data centers amidst fulfillment warehouses (and toss in the ominous “other”). And the balance sheet breaks out the value of land and buildings vs. equipment (with some depreciation deprivations going on which CAPEX purists prefer to ignore, caring as we do only about the raw expansion of infrastructure). We can extract a couple things from the filings: AWS accounts for 5% of Amazon’s square footage (but growing a little faster than the rest of the company). That is both office space and data centers. Office space is under 10% of the company’s square footage, and physical stores under 4%. “Fulfillment, data centers and other” are almost 86% of Amazon’s collective 550-600 million square feet (the different breakouts don’t agree, but who among us hasn’t misplaced a few dozen million square feet of real estate?). Land and buildings are about a third of the depreciated assets on the balance sheet while equipment is 55-60% depending on the year (that pesky depreciation is a bigger factor for equipment). Amazon added 125-130 million square feet in 2021 at a cost of almost $24 billion. Allocating 5% of this to AWS yields a little over a billion-dollar investment (data center floor space could be more or less expensive than warehouses, but let’s assume on average about the same). Beyond that it gets purely speculative. Partitioning 2021’s approximately $31 billion in equipment spend between data centers and fulfillment is very difficult. You can play with models of the relative cost per square foot to kit out a data center versus a warehouse (which still have forklifts, robots, conveyors, scanners, etc.). Its hard to make the numbers work if the data center equipment is upwards of two orders of magnitude more costly per square foot, but that still leaves enormous leeway. And you’ve also got network costs decoupled from square footage as well as new and replacement hardware going into existing data centers. In the absence of a defensible model, I’m just going to guess that roughly half the equipment budget – say $15 billion – goes to racking out data centers. (Needless to say, I’m very interested in more rigorous models). Adding in the billion for real estate gets us to a guess of $16 billion in AWS spend out of a total of a total company-wide spend of almost $74 billion (so 22% of the total). This totally wet-finger-raised-in-the-air stab in the dark seems like it is in the right ballpark as 60-70% of both Microsoft and Google’s CAPEX spending gets you a range of $15-18 billion. Given both Google and Microsoft have a bunch of other services, this AWS estimate could be high, so consider it a ceiling. Amazon employees are of course welcome to educate us about their data center frugality.
UPDATE: while I was playing with the READENTRAILS() function in Excel, Amazon was actually disclosing how their CAPEX breaks down:
Let’s talk a little bit about capital expenditures. And I’m going to do this with the inclusion of equipment finance leases, which is the residual that we sometimes lease on our infrastructure assets. We’re doing less of it now, but we still do some and have done it historically. So, when you look at those numbers and how they’ve grown over the last few years, I’ll give you the proportions, which I’m not sure we’ve initially shown before is about 40% — just under 40% of that CapEx is going into infrastructure, most of it’s feeding AWS, but also certainly, Amazon is a large customer of that as well as we build and structure for ourselves directly or through AWS.
About just under 30% is fulfillment capacity building warehouses — warehouse only, not transportation. And then just under 25% is transportation capacity and building out our AMZL network, principally globally. The remaining 5% or so is small things like offices and stores and other capital areas. But those are the three main areas.
If I look to the future, we’re still working through some of our plans 2022, but it’s coming into focus a bit. We see the CapEx for infrastructure going up. We still have a very fast-growing business this growing globally, and we’re adding regions and capacity to handle usage that still exceeds revenue growth in that business. So, we feel good about making those investments.
On the fulfillment center side, that’s about 30% of the spend in the last two years. We see that moderating and that will probably now match growth of our underlying businesses. I think there’s always things that can tick up that growth rate, things like expansion of our FBA business, expansion of cube that maybe not be different than the square footage. So, there’s — we want to have capacity to have a healthy retail and FBA business because that fuels prime in 1-day delivery and 2-day delivery and same-day delivery. So, that’s very important. But we see the FC piece likely moderating this year. And then the third piece is transportation. We still see additional levels of investment in that in 2022.
So, if you wrap that up, we expect CapEx including equipment finance leases to increase year-over-year.
So if we add up standard CAPEX plus finance leases (but not build-to-suit leases), Amazon spent $68.1 billion in 2021. Breaking it down yields:
- Infrastructure: <$27.2 billion. So more than my stab in the dark, with the caveat that some of this is for Amazon, not AWS (whatever that means). How much of AWS’s business comes from the rest of Amazon is an interesting question.
- Fulfillment warehouses: <$20.4 billion
- Logistics: <$17 billion. For comparison, Fedex spent $6.2 billion and UPS spent $4.2 billion in the last year…
- Offices/Stores/Other: $3.4 billion
The use of finance leases continued to decline (to just over half the level of 2019) while build-to-suit leases hit an all-time high (physical stores?). My guess is Amazon is just less capital constrained than in the past so can pay cash for what they used to buy with finance leases (aka servers). But, as another data point for our equipment triangulation, finance leases peaked at $13.7 billion in 2019, which is the right order of magnitude for data center equipment spend.
Possibly also relevant to the decline in finance leases, Amazon again bumped the useful life of their servers to five years, after extending it to four years just two years ago. This doesn’t change raw CAPEX spending, but is material to accounting earnings: “based on servers and networking equipment that are included in “Property and equipment, net” as of December 31, 2021, will have an anticipated impact to our 2022 operating income of $3.1 billion.” We’ll see if Google and Microsoft follow, as they’re both at four years.
Google (if Larry Page is done with Alphabet, so am I) has been on a four-year CAPEX cycle since 2006, and 2022 is the next peak year. The company confirmed this in their most recent quarterly earnings call: “In 2022, we expect a meaningful increase in CapEx. In Technical Infrastructure, servers will again be the largest driver of spend.” If they spend the historic cycle peak of at least 16% of revenue in 2022, on ~$300 billion of revenue, they could push $50 billion in CAPEX. Every little bit helps.
Consistent with its overall strategy to stay nondescript and leave the (antitrust) spotlight to others, Microsoft’s CAPEX spending remains boring, monotonic and inscrutable. Spend has been almost perfectly linear up and to the right over the last decade. They have ticked up one percentage point to 14% of revenue, a new high for the company. I again implore them to do something dramatic on behalf of CAPEX color commentators everywhere.