Tl;dr: “You know what’s cool? A trillion dollars.”
Previous retrospectives: 2016, 2017, 2018, 2019, 2020, 2021, 2022, 2023 plus earlier/other CAPEX musings.
When I first started to follow the cloud CAPEX (capital expenditures), it was a super-niche Internet otaku, as I obsessed over a single line on the cash flow statements of companies considered “asset-light”.
A decade later AI-driven cloud CAPEX is the epicenter of the world’s biggest investment boom, entwined with a coincident energy build-out catalyzed by data center demands (but also atoning for decades of underinvestment).
We’re all CAPEX obsessives now (for better or worse!).
Get Platformonomics Updates By Email
CAPEX offered a useful tell to separate the real clouds from the clowns, and still helps call out today’s AI clowns. Beyond partitioning off the poseurs, CAPEX both shapes and reflects the great debates of the AI Era. Will brute force scaling keep scaling? How far will algorithmic optimization go? What are the new economics of software, a domain once gloriously free of marginal costs? How will we buy incremental reasoning inference?
Deploying CAPEX requires a manufacturing capability. Turning tens of billions into cloud capacity efficiently and at hyperscale is an extremely complex, multi-year process. The supply chain is beset by component scarcity plus long lead times for permitting, construction, and grid interconnection. Power is the scarcest resource, highlighted by the creative and decade-away investments we’re seeing in nuclear power. This year’s CAPEX has been in process for years. You can’t ramp CAPEX to the big leagues overnight (or even in a couple years), however much you desire it so (Hello Oracle! Hello Stargate!).
Historically, this analysis has focused on the hyperclouds: Amazon, Google, and Microsoft. With AI driving CAPEX growth, we’re adding Meta to this annual review. They’re the smallest spender of the bunch and don’t offer infrastructure services, but have joined the top tier. Let’s call them the hyperCAPEX companies.
In past years, I’ve included a disclaimer that these numbers are the companies’ total CAPEX spending, not just data center infrastructure. Then I’d include a lengthy list of other CAPEX categories and a joke lamenting Google’s still missing-in-action space elevator. But today, except for Amazon, the CAPEX goes overwhelmingly to data center infrastructure. The numbers include finance leases for Amazon, Meta, and Microsoft (infrastructure, typically servers, acquired using debt that is repaid over the useful life of the hardware).
Summary
The four hyperCAPEX companies – Amazon, Google, Meta and Microsoft – collectively spent over $251 billion on CAPEX in 2024 (up 62% from 2023’s $155 billion). This is company-wide CAPEX, which is almost all data center infrastructure except for Amazon, whose spending on AWS infrastructure was 64% of their corporate total: $53 billion. For continuity with previous retrospectives, our three hypercloud companies spent $212 billion, up 67% from 2023’s $127 billion.
After two down years of post-pandemic infrastructure digestion, Amazon returned to CAPEX growth in 2024, boosting spending 57% to $83.9 billion (including both CAPEX and finance leases), hitting an all-time high. After a first-ever decline in 2023, AWS CAPEX rebounded by 114% to $53.27 billion (64% of their overall CAPEX), also an all-time high.
Google CAPEX grew 63% to $52.5 billion. An all-time high.
Meta grew CAPEX 40% to $39.23 billion. Yep, an all-time high.
Microsoft spent $75.6 billion on CAPEX, an 83% increase. Yet another all-time high.
Over the last few years the hyperCAPEX companies have gone from just cracking the top ten to dominating the ranks of the world’s biggest CAPEX spenders. Amazon and Microsoft spend more on CAPEX than any other (non-Chinese[1]) company in the world, with Google and Meta in the top seven. Biggest Tech outspends the largest auto, energy, semiconductor, and telecom companies.
FUN FACT: Amazon, Google and Microsoft each outspent all three US mobile operators (ATT, T-Mobile and Verizon) combined in 2024.
Our three hypercloud companies passed a TRILLION dollars in cumulative 21st century CAPEX in 2024. A TRILLION dollars! With a T. The four comma club. Both US Federal government deficit and Chinese trade surplus territory. Adding Meta, the four hyperCAPEX companies’ cumulative CAPEX spend since 2000 is over $1.19 TRILLION, with $406 billion of that spending in the last two years. Pretty soon we’re going to be talking about real money.
FUN FACT: Meta, the runt of the hyperCAPEX litter, spent more on CAPEX in 2024 than outspoken nanocloud promoter Oracle has in its nearly 50 year history.
Looking at CAPEX as a percentage of revenue, all our companies’ ratios increased, with Microsoft hitting a whopping 29%!
There is another line on the cash flow statement that is getting interesting: amortization and depreciation. Whee!!!
Now to the individual companies:
Amazon
The most common Amazon question is how much of their CAPEX goes to AWS infrastructure and how much supports the $500+ billion retail business. Or at least it should be a common question. Many new CAPEX obsessives (especially VCs on Twitter) attribute all of Amazon’s corporate CAPEX to AWS, which is a mistake. Amazon helpfully breaks this out annually in their 10-k:
In 2024, we see a big increase in AWS CAPEX, from $24.8 billion in 2023 to $53.3 billion. AWS also jumps to an all-time high as a percentage of Amazon’s overall CAPEX spend at 64%. It looks like AWS had trouble ramping AI-driven spend in 2023. They spent the whole year just three steps into a 10K race, and also had to come to grips with unfamiliar supplier dynamics (Hello NVIDIA!). It is yet further evidence CAPEX deployment is a multi-year process, even with a well-tuned, at scale machine. And after years of a steady decline in the capital-intensity of the AWS business, in 2024 it zoomed back over 50% of revenue.
Amazon’s guidance in October for total 2024 CAPEX spend was $75 billion, but ended the year at $83.9 billion! With that variance in mind, Amazon’s CAPEX guidance for 2025 is a “reasonably representative” $105 billion, and “The vast majority of that capex spend is on AI for AWS.” This comment is ambiguous whether it is about 2024 or 2025, but is certainly an exaggeration for either year or uses an extremely broad definition of what is AI vs. ordinary AWS infrastructure. They still need servers, storage, switches, etc. to support a $100+ billion business, and AI is a small fraction of overall revenue.
After years of all the hyperCAPEX companies extending the useful life of their servers (dropping billions to the bottom line through GAAP magic), Amazon has reversed direction: “we’re decreasing the useful life for a subset of our servers and networking equipment from six years to five years.” This is material, may or may not be AI-related, and it will be interesting to see if other companies follow:
We anticipate this will decrease full year 2025 operating income by approximately $700 million. In addition, we also early retired a subset of our servers and network equipment. We recorded a Q4 2024 expense of approximately $920 million from accelerated depreciation and related charges and expect this will also decrease full year 2025 operating income by approximately $600 million. Both of these server and network equipment useful life changes primarily impact our AWS segment.
Amazon’s use of finance leases, which they relied on heavily in the early days of AWS, have become di minimis. In 2024 they procured a mere $854 million of infrastructure through leases, just 1% of their overall CAPEX.
My oft-stated view is Google has been skimping on CAPEX in recent years, relative to their historical refresh cycles, as they prioritized genuflecting to the (peacetime) god of Wall Street.
But in a less idyllic world, Google is finally stirring, with CAPEX spending up 63% in 2024. Most of their infrastructure supports search and YouTube, but they are deploying lots of AI infrastructure that can serve both their first party businesses and Google Cloud customers.
Google’s CAPEX guidance for 2025 is $75 billion, which would be a 43% increase.
The fundamental dilemma for Google is they are chasing (and spending on) the new thing (AI), and the more they do so, the more they undermine their amazingly profitable franchise based on an index of the web. It is strongly reminiscent of Microsoft and the web a generation ago.
Meta
Meta’s CAPEX guidance for 2025 is $60-65 billion which would be a 53-66% increase.
My (meta) Meta question is what are they going to do with all this AI infrastructure? Is it all for internal use or do they have ambitions to offer infrastructure services to third parties? How much ad tracking (aka “innovation” as it is described when they’re complaining about Apple) does the company need?
Can Llama build a successful ecosystem without a thoroughly committed first party provider at scale? Does Meta need to offer Llama Cloud? It is a slippery slope as there are so many related services that must be provided and inevitably you end up chasing the “enterprise”. Meta’s history of developer rug pulls would be an obstacle to success in this business. I’ll be printing “fool me once, shame on you…” laptop stickers to remind developers to think hard about the track records of their dependencies (Hello Elon!).
Microsoft
CAPEX go up! Up up up!
Microsoft got a jump on the AI buildout by virtue of hosting ChatGPT (their numbers further illustrate the lag in translating CAPEX desires into data centers, as they probably stomped on the accelerator in late 2022). They’re NVIDIA’s biggest customer and have a lot more NVIDIA boxes than the people publicly beating their chests about their H100 collections. They’re even restarting Three Mile Island (the other reactor there, not the one that melted down).
Microsoft’s CAPEX guidance is to spend $80 billion in their fiscal year ending June 2025, with continued, if lower, growth in FY26.
In FY26, we expect to continue investing against strong demand signals including customer contracted backlog we need to deliver against across the entirety of our Microsoft Cloud. However, the growth rate will be lower than FY25 and the mix of spend will begin to shift back to short-lived assets which are more correlated to revenue growth.
These CAPEX numbers understate Microsoft’s AI infrastructure, as they are also spending billions to rent capacity from various boutique GPU hosters (including CoreWeave, Lambda Labs, and Oracle). That matrix multiplication machinery shows up in OPEX, not CAPEX.
The big question is how Microsoft’s strategy shifts will impact CAPEX. They’re moving away from just Open AI as their foundation model and restructured the relationship with Open AI. Microsoft seems to have disembarked from the Open AI CAPEX train to infinity (and beyond!) and given up their seat to that notoriously cautious steward of capital, Softbank. Their attention is moving up the stack to AI-powered experience for various customer segments. The models are increasingly a means, not an end unto themselves. Will their capital-intensity peak?
Keep up with Platformonomics analysis of CAPEX trends and other topics by subscribing:
Get Platformonomics Updates By Email
[1] Because I neither believe Chinese numbers nor want to look them up.