Tl;dr: Cloud CAPEX is once again both exciting and revealing!
Previous retrospectives: 2016, 2017, 2018, 2019, 2020, 2021, 2022 plus earlier/other CAPEX musings.
Capital expenditures (CAPEX) are old school. You acquire some abiding apparatus around which you build a business. It is the antithesis of the “asset light” avoidance of hard work or the financial shenanigans that characterize many trendy business models. Buy some (typically unglamorous and/or unwieldy) machinery and earn back your investment over time.
The CAPEX that begets cloud infrastructure has been mostly uneventful in recent years, ever since it helped us separate the clouds from the clowns (the cloud pretenders who talked a good game but didn’t put real money where their mouths were). But times change! Cloud CAPEX is again both exciting and revealing thanks to generative AI.
Cloud CAPEX is fascinating at multiple levels. The sheer magnitude of hypercloud CAPEX spending is breathtaking and glorious in and of itself. The hypercloud trio, who I quaintly think of as software companies, are marshaling matter across a variety of businesses (not just cloud infrastructure) on the same scale as the world’s very largest metal-benders.
Cloud infrastructure is a multi-hundred-billion-dollar industry, with the majority of IT spend still an opportunity ahead of them. Hypercloud CAPEX not only tells us what is happening in the market, but also what the providers think will happen, as they must invest ahead of revenue.
And now generative AI is disrupting the cloud infrastructure business, toppling architectures, customer priorities and industry pecking orders. Cloud CAPEX gives perspective on how AI platitudes are turning into GPU purchases.
There is lots of good analysis about who has the most GPUs or which silicon might challenge NVIDIA. As fun (and interesting) as that is, that is not our purpose here. We’ll stick to what the cloud infrastructure lens reveals about both cloud and AI investment. We’ll try to stick to the data and save the really speculative conclusions for later. And we’ll even resist (almost) making fun of the rise of GPUs as an asset class (because everything should be financialized!).
Standard disclaimer: Unless otherwise indicated, 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 still 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 (the leases are debt instruments used to finance specific CAPEX expenditures, namely servers and buildings).
Get Platformonomics Updates By Email
Now the numbers!
SUMMARY
The three hypercloud companies – Amazon, Google, and Microsoft – collectively spent over $127 billion on CAPEX in 2023, flat from 2022. That is company-wide CAPEX. Their combined cloud infrastructure CAPEX was on the order of $80 billion (we don’t get clean breakouts from Google or Microsoft).
Amazon’s (relative) CAPEX austerity continues, as the company spent a measly $53.7 billion, a decline of 20%, following a 9% decline in 2022. Unlike last year, it wasn’t just retail fulfillment retrenchment. AWS infrastructure spending also declined for the first time ever. Given how much Amazon talks about generative AI, it is a very odd year to cut AWS CAPEX.
Google eked out a 2% increase to $32.3 billion. Their much-anticipated 2022 server fleet refresh still hasn’t happened. There may be a glimmer of an AI infrastructure uptick in the fourth quarter, but otherwise nothing that screams “Code Red”.
Microsoft CAPEX grew 45% to $41.2 billion. Microsoft is patient zero in the mad scramble for GPUs with its OpenAI partnership. OpenAI is nothing without GPUs, and those GPUs are in Azure. Microsoft crossed the chasm from training to inference at scale, serving hundreds of millions of monthly users. Microsoft’s AI investments are impossible to miss.
Amazon still spends more than any other (non-Chinese1) company in the world, while Google and Microsoft stand amongst the biggest CAPEX spenders on the planet.
The three hypercloud companies’ cumulative CAPEX spend since 2000 is over $820 billion, with $250 billion of that spending in the last two years. Even in the event of continued sobriety at Amazon, that number should easily pass a trillion dollars invested in 2025.
Looking at CAPEX as a percentage of revenue shows Amazon’s decline back to 2015 levels, the fizzling out of Google’s quadrennial server refresh cycle, and Microsoft’s generative AI spike.
This year we’ll review the hypercloud companies in reverse alphabetic order.
Microsoft
Microsoft has gone from perennially the most boring to the most interesting of our CAPEX contestants. After being smoothly monotonic since 2016, and very consistent as a percentage of revenue, Microsoft’s CAPEX jumped by 45% (+$12.8 billion) in 2023 (ChatGPT was introduced November 30, 2022).
FUN FACT: Microsoft spent more on CAPEX in 2023 than Oracle has in its entire history.
The company said calendar fourth quarter CAPEX spending was actually “lower-than-expected due to delivery for a third-party capacity contract shifting from [fiscal year] Q2 to Q3” (NVIDIA or AMD?) and they expect “capital expenditures to increase materially” in 2024. In June the company sent the rarely-fielded CFO/CTO tag team to warn Wall Street of its coming CAPEX binge, introduce the concept of software COGS, and assure them it would all pay off.
Microsoft is reputed to be the largest customer for both NVIDIA (Q4 2022, Q2 2023, 2H 2023) and AMD (plus is doing its own AI silicon). The abrupt 4% bump in CAPEX as a percentage of revenue, after a steady 13-14% for years prior, is the best proxy for the incremental AI spending in 2023. That suggests an incremental AI-driven spend of about $9 billion, or 22% of overall CAPEX.
But this number understates Microsoft’s mad scramble for GPUs. The company is also spending billions to rent capacity from various boutique GPU clouds (including CoreWeave, Lambda Labs, and Oracle). That matrix multiplication machinery shows up as OPEX, not CAPEX.
Microsoft even identifies GPU availability as a risk factor in its 10-K:
Our datacenters depend on the availability of permitted and buildable land, predictable energy, networking supplies, and servers, including graphics processing units (“GPUs”) and other components.
Now they just have to turn all this investment into profitable business. And profit margin expectations at Microsoft are very high…
Google has been a disappointment to CAPEX obsessives for the last two years. We were hoping for their fifth quadrennial server upgrade cycle to hit in 2022 with CAPEX spiking back to at least 17% of revenue, thereby boosting annual CAPEX to over $48 billion.
Instead, we got an anemic bump from 2021’s nine-year low of 10% of revenue all the way up to 11% in 2022 (it was still 28% dollar growth to $31.5 billion in CAPEX given their ever-growing revenue base). In 2023 Google CAPEX nudged up 2% to $32.3 billion.
We remain hopeful that despite the bean counters seemingly winning the battle between depreciation schedules and a withering Moore’s Law, that fifth refresh cycle is merely delayed as the useful accounting life of servers has been extended from four years to six years.
It is amusing that on the list of “discipline” and “efficiency” initiatives Google felt worth surfacing to Wall Street is improved “machine utilization”. Presumably that means there are material savings to be had (they also announced they were eliminating staplers2). One can easily imagine individual development teams having (and having forgotten about) tens of thousands of servers in their couch cushions3. So there is a scenario where Google’s infrastructure was (vastly?) overbuilt and some of it can be recycled to meet growth demands.
FUN FACT: Google spent more on CAPEX in 2023 than IBM has over the last decade.
On the AI front, Google faces the most interesting strategic quandary of the hyperclouds. Google invented transformers and has long led in AI research. Letting others capitalize on their inventions would (merely) merit a Fumbling the Future sequel. But to the degree generative AI poses a real and disruptive threat to Google’s search franchise (one of the best businesses ever), their response is both incredibly important and incredibly closely scrutinized.
That response, as Google plays strategic defense for the first time, has been muddled. The CEO dedicated time to denying he invoked a “Code Red” response to ChatGPT (he certainly should have invoked a “Code Red”, so why deny it?)4. They are (appropriately) trying to downplay any disruption to search, but a competitive version of Gemini is late. Questions are coming from inside the house about management’s ability to transition the company from peacetime to wartime.
Infrastructure should be an immense advantage for Google as they respond to the “Code Red”. But that 2023 bump in CAPEX of 2% doesn’t even keep up with inflation (though IBM has been triumphant about that level of “growth”). So there was no infrastructure “Code Red” in 2023, and the contrast with Microsoft is stark.
But if we drill down to the quarterly trend, Google’s CAPEX did grow sequentially through 2023, with over a third of their annual spend coming in the fourth quarter. We expect that will continue in 2024, with Google’s guidance of “notably larger” CAPEX. The sleeping giant is at least stirring, but they’re still at least a year behind Microsoft.
In terms of the implications of AI’s rise for the cloud computing business, I’ve argued that Google Cloud is a hobby for Google, well down the CEO’s list of priorities. If Google’s core franchise is at risk, I think hobby status becomes even more pronounced, as (strategic) attention is all you have5.
Amazon
The biggest cloud CAPEX news from 2023 is the first ever decline in AWS infrastructure investment. Amazon’s overall CAPEX spending fell 20% to $53.7 billion (still the biggest corporate CAPEX spend ex-China), while AWS spending fell 10% to $24.8 billion. AWS was outspent on infrastructure this year by Microsoft and perhaps also Google.
FUN FACT: Amazon in austerity mode still spent more on CAPEX in 2023 than all three US mobile operators combined (ATT, T-Mobile, Verizon)
There are two broad dynamics in play here. The first is regime change (both financial and in leadership) at Amazon, combined with a lot of previous excess (some called it #bonkers but “CAP-EXcess” would have been good too) that needed to be “optimized”.
For a long time (call it the Bezos Doctrine), Amazon tried to reinvest every single penny back into the business and minimize accounting profits. By 2015, they couldn’t reinvest cash flow fast enough and the company began to show growing profits and free cash flow (despite their very best intentions).
The pandemic interrupts this trend, as Amazon revenues soared with people stuck at home with nothing to do but shop, while the company simultaneously mounted one last throwback #bonkersbonkers retail infrastructure investment blowout in 2021 (they invested $73.7 billion in CAPEX to double the size of their fulfillment network and build their own UPS-sized transportation network, overbuilding just as the pandemic ended, a binge from which they are still cleaning up “optimizing”).
But now Amazon is inexorably reverting to that pre-pandemic profit curve. Amazon is now “a profit deal”, which means costs, including CAPEX, get actual scrutiny.
New CEO Andy Jassy has embraced financial optimization with gusto (broad layoffs, asking why Alexa exists much less is losing $5 billion a year, adding “rinky dink” ads to Prime Video, etc.), and has not exempted his baby AWS. This regime change probably marks the calcification of Amazon’s footprint, as once you become a profit optimizer, it is very hard to deviate from that path to make big, bold, new investments. But beyond whacking the money-losing projects, they still need to keep the franchise retail and cloud businesses growing and profitable. And both of those require massive CAPEX.
Beyond the regime level changes, 2023’s material decline in AWS CAPEX is quite alarming for what it says about both the existing cloud business and how they are faring in the AI race.
AWS has been reinvesting ~35% of revenue in recent years in cloud infrastructure. If you squint at the chart below, you’ll see the rate of investment roughly turning into a similar level of revenue growth the following year. But that relationship has broken down in the last two years, as revenue growth has slowed to 12-13%. That suggests that AWS is overbuilt relative to recent revenue growth.
But a surplus of older gear doesn’t help with generative AI. One observation about Amazon generally, and AWS in particular, is they don’t tend to talk about things unless they’re behind. And they’ve been running their mouth non-stop about generative AI, so cutting billions in CAPEX spend alongside that rhetorical frenzy is quite a surprise.
Amazon and AWS face a host of challenges around generative AI. Despite all the rhetoric, they’re way behind, and this is a space where catching up is especially hard. Their Titan LLM is missing in action (and on the wrong side of their own “there is no compression algorithm for experience” quip). They had to license Anthropic for Amazon’s own internal use, though the announcement was dressed up to look like an endorsement of AWS’s generative AI infrastructure.
But worse, their existing infrastructure may actually be a hindrance. It has been argued AWS’s infrastructure is “poorly adapted” for the generative AI world. And they clearly got sideways with NVIDIA, which means their H100 shipments are “significantly lower than their share of the public cloud”.
It looks like Amazon tried to push NVIDIA around, as they’re accustomed to doing, but lost. The power dynamics have changed and NVIDIA now has the upper hand6. Amazon resisted for months, but eventually capitulated, and had to host NVIDIA’s DGX Cloud (and even say nice things about it despite really not wanting to do it) to get more GPUs.
But beyond product and vendor travails, slashing CAPEX by 10% in the middle of the AI race is an enormous red flag (and different than a Code Red). It makes Amazon’s incessant claims to generative AI leadership all the harder to stomach. As a result, we are adding AWS to the negative watch list, putting them at risk of being added to the “AI clown” list (where they would join initial inductee IBM). If Google is at least a year behind Microsoft, AWS is even further behind. And their guidance doesn’t suggest they have a clear CAPEX plan for 2024:
CapEx will go up in 2024. I’m not giving a number today, but we do — we’re still working through plans for the year, but we do expect CapEx to rise as we add capacity in AWS for region expansions, but primarily the work we’re doing with generative AI projects.
As a final observation for the most hard core of CAPEX obsessives, AWS’s use of finance and built-to-suit leases have dropped to almost nothing (less than half a percent of AWS CAPEX in 2023). This is notable as they were ~80% as recently as 2019. It was also interesting to hear the CFO say, “We define our capital investments as a combination of CapEx plus equipment finance leases.” So no respect for build-to-suit leases.
What’s Next?
There are a couple things on the CAPEX agenda:
- Triangulate the cloud CAPEX numbers against NVIDIA’s numbers when they report in a couple weeks.
- Add Meta to the CAPEX club. They’re not a hypercloud (and aren’t likely to be given their developer DNA, or lack thereof), but they spend almost as much as the hyperclouds on CAPEX ($28.1 billion in 2023). And they’re gobbling GPUs. How much of that CAPEX supports Facebook/Instagram/Whatsapp vs metaverse/hardware vs. new things? Are they building generative AI-powered search?
- Check in on the clown car race amongst the CAPEX pretenders (after Oracle announces earnings).
Please comment below (or contact me privately) with what I got wrong here and what else I should know about!
- Because I neither believe nor want to look up the Chinese CAPEX numbers ↩︎
- Perhaps Google, with all its AI experience, considers staples a bigger risk than the more traditional existential risk of paper clips? ↩︎
- My bet is there are individual hypercloud dev team environments with more servers than IBM Cloud. ↩︎
- “Schrödinger’s Code Red” would be a good application for Google’s quantum computer. ↩︎
- It is probably a bad sign when a joke needs an explanatory footnote, but that was an “Attention is All You Need” reference. I also considered mashing that up with “You Can’t Always Get What You Want“. ↩︎
- To the point where NVIDIA are trying to revive the vertically integrated computing model of yore, spanning chips, systems, software and cloud services. But that is a topic for another day. ↩︎