Tl;dr: AMZN goes ex-#bonkers
Since we separated the clouds from the clowns, cloud capital expenditures (CAPEX) haven’t been a particularly interesting tell. Huge numbers but few omens. But ironically just as money gets more expensive with the end of the zero-interest rate-fueled bubble, cloud CAPEX (and CAPEX more generally) is getting more interesting again.
New questions include how cloud CAPEX will adapt to rising inflation and interest rates, a tougher economic environment, and the “tech pullback”, as well as how the AI enthusiasm shows up in infrastructure spending (spoiler: no answers in this data).
And beyond technology, the wider world has rediscovered the glories of CAPEX. Supply chains need to be expanded and repositioned. Putin precipitated a reappraisal of “energy security” after years of underinvestment in both production and distribution (I suggest the cloud lead the way and go nuclear). And governments are throwing huge amounts of money at battery and semiconductor manufacturing as every region seeks to be the dominant (or at least self-sufficient) producer.
On to the 2022 numbers!
SUBSCRIBE TO GET THESE POSTS DELIVERED BY EMAIL (see box ->)
The three hypercloud companies – Amazon, Google, and Microsoft – collectively spent over $127 billion on CAPEX in 2022, up a mere 2.5% from 2021’s $124 billion (that is company-wide CAPEX, not just cloud infrastructure).
Google and Microsoft continued to grow CAPEX by double digits, but Amazon’s once exponential spending curve now looks more parabolic.
Amazon spent $67.5 billion, a decline of 9% from 2021. The retrenchment is in the retail business after its absolutely bonkers pandemic build-out. Perhaps to ensure AWS isn’t tainted by that belt-tightening, Amazon started calling out its CAPEX spending for AWS, which remains comfortably up and to the right (Google and Microsoft, feel free to make similar disclosures).
Google increased CAPEX 28% to $31.5 billion while Microsoft was up 10% to $28.4 billion (Microsoft had outspent Google for the first time post 2009 in 2021). But Google’s spending was still low relative to what we’d expect for its historic four-year infrastructure refresh cycle.
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).
Amazon still spends more on CAPEX than any (non-Chinese) company in the world, and Google and Microsoft still spend more than all but the very biggest energy and semiconductor companies (sorry Intel, but you’re not yet back in the “very biggest” category). Double-digit increases were the rule in the high CAPEX precincts for 2022:
Aramco: $40-50B est. +25% est. (not reported 2022 yet, 1H under estimates)
ATT: $19.6B +19%
Exxon: $18.4B +52% (less than forecast)
Intel: $25.1B +23% (less than forecast)
Meta: $32.0B +67% (a big u-turn is a coming…)
Samsung: $43.5B +28% (more than forecast)
TSMC: $36.6B +21% (less than forecast)
Verizon: $23.1B +14%
The three hypercloud companies’ cumulative CAPEX spend since 2000 is just shy of $700 billion, with over $250 billion of that spending occurring in the last two years. Amazon’s slowdown probably means we won’t see a trillion dollars in cumulative spend until 2025. Alas.
As a percentage of revenue, Amazon’s spend dropped from their all-time (pandemic) high of 16% down to 13%, Google saw a small jump to 11% (more on that below), while Microsoft was steady at 14%.
After several years of #bonkers CAPEX, Amazon has gone ex-#bonkers. Given they spent nearly half their lifetime CAPEX in 2020 and 2021, some indigestion might be expected. With the transition from Jeff “Just Reinvest Every Single Penny of Free Cash Flow on Something, Anything” Bezos to Andy “We Spent How Much on Alexa?” Jassy, they are sorting through the fulfillment, logistics, and retail stores overbuild.
But we don’t really care about e-commerce warehouses and can instead dig into AWS CAPEX numbers. As they previously did with their use of leases, Amazon proactively called attention to numbers they have been disclosing for a while. Buried deep in SEC filings, and opaque to a search for the keyword “CAPEX”, lies a breakout of AWS-specific CAPEX.
The capital intensity of AWS is quite stunning. They are still investing 35% of AWS revenue in infrastructure, and that number has come down from over 60%.
if we do a quick and dirty proxy that says 70% of Google and Microsoft’s CAPEX goes to data centers, we can compare their spending with AWS (note those data centers are serving more than just cloud infrastructure services). This is more apples-to-apples than company-level CAPEX.
Interestingly, AWS doesn’t surpass Google spending until 2020 (but also remember Google search and YouTube use far more infrastructure than Google Cloud). But by this estimate, all three companies are pretty close on their data center spend.
We can also observe AWS’ use of finance leases, which were funding upwards of 80% of AWS infrastructure as recently as 2019, have dropped to almost zero. Having $70 billion in cash and short-term securities, as well as rising interest rates, may make leases less attractive.
The biggest news on the Google front is the failure of their fifth hardware refresh cycle to materialize. Google had been consistently upgrading their fleet of servers every four years since 2006. The guidance last year was to expect a “meaningful increase in CapEx” (that is their less enthusiastic casing as opposed to the full-throated “CAPEX” we prefer), but this year’s actual increase in investment was relatively anemic ($31.5 billion +28%).
In the past, we’ve seen CAPEX spent go up by 111 to 173% from peak to peak in the fourth year of the cycle (a 150% increase would have been $63 billion, almost exactly twice what they did spend). Returning to a modest 16% of revenue would have resulted in a $45 billion spend. CAPEX budgets presumably got cut along with everything else over the last year. And their guidance for 2023 is for flat spending on CAPEX. Disappointing, needless to say.
Google also continues to extend its depreciation period for servers on annual basis, suggesting they now have a useful life of six years (Amazon is at five, Microsoft six). Whether this is true or just gives accountants some comfort in cutting CAPEX budgets (it will drop $3.4 billion to their bottom line) is not clear. But it is yet another sign that the once bold Google has become timid.
As Google faces real economic constraints and unprecedented competitive threats to its core business (“code red”), we will be closely monitoring the fate of their cloud computing hobby.
Microsoft has always been the most boring and monotonic of the hyperclouds when it comes to CAPEX. They have gotten to a very consistent spend relative to revenue over the last five years. A 10% increase in 2022 revenue was coupled with a 10% increase in CAPEX.
But things have gotten more exciting with Azure hosting ChatGPT and the widely publicized capacity shortfalls. Beyond an announcement with NVIDIA for tens of thousands of high-end GPUs, we don’t have much insight into how Microsoft’s infrastructure spend is shifting to support AI applications. This is on the list of CAPEX questions to watch (and let me know if you have any data or insights).