Tl;dr CAPEX Go Up! Up Up Up!
Previous retrospectives: 2016, 2017, 2018, 2019, 2020, 2021, 2022, 2023, 2024 plus earlier/other CAPEX musings.
This is my tenth annual CAPEX (capital expenditures) retrospective. The fascination began earlier, as an asset-oblivious software guy confronting an alien lifeform of spending on physical property, plant and equipment.
CAPEX separated the clouds from the clowns. A lull followed. Then AI, expressed through CAPEX, became the global economy’s main engine—dragging entire sectors in its wake and polarizing geopolitics.
CAPEX both shapes and reflects the fundamental AI debates—scaling and CAPEX are inexorably intertwined. Is there any chance of keeping up with token demand? How many more ways are there to throw scaled compute at the problem? Could algorithmic breakthroughs break the curve? Have we internalized the implications of software having marginal costs? Can dystopian economic scenarios plausibly pair ever accelerating CAPEX investment with mass unemployment? Will AI infrastructure become the center ring of the political circus? So many interesting questions revolve around CAPEX.
Today nearly everyone is a fellow CAPEX obsessive. I no longer need to explain it like you’re five (or like you came from software), so we can get straight to it. Beyond reviewing the numbers, I’ll share my big questions and flag some of the most common CAPEX mistakes I see.
The figures herein include investments in property, plant and equipment plus the infrastructure acquired using finance leases by Amazon, Meta, and Microsoft (primarily servers financed with debt and repaid over the useful life of the hardware).
Summary
The four hyperCAPEX companies —Amazon, Google, Meta and Microsoft—collectively spent over $416 billion on CAPEX in 2025 (up 66% from 2024’s $251 billion). This is company-wide CAPEX, which is overwhelmingly data center infrastructure with one exception: Amazon, which also has a retail business with a few trucks and warehouses (plus a satellite constellation).
I focus on those four because no one else is close in their spending. Just as there are clouds and clowns, there are hyperscalers and mere scalers of hype. Amazon’s and Microsoft’s incremental CAPEX in 2025 is each roughly on par with hype scaler Oracle’s lifetime CAPEX.
Every hustler wants the hyperscaler label. But if everyone is hyper, no one is. Abuse is devaluing the term. It’s embarrassing to see Morgan Stanley identify 11 “hyperscalers” (including IBM, the financial engineering company whose only even remotely hyper- worthy feat was reporting negative CAPEX) while Synergy Research stretches the list to 21. Hence our hyperCAPEX label.
Amazon CAPEX grew 61% to $134.7 billion, an all-time high (abbreviated ATH here for reasons that will soon become obvious).
AWS CAPEX which Amazon kindly breaks out jumped 81% to $96.5 billion (ATH), accounting for 72% of total Amazon CAPEX (ATH).
Google CAPEX rose 74% to $91.5 billion (ATH).
Meta CAPEX surged 84% to $72.2 billion (ATH).
Microsoft CAPEX increased 56% to $118 billion (ATH).
ATH. ATH. ATH.
I used to benchmark hyperCAPEX spending to traditionally capital-intensive industries—energy, telecom, utilities and semiconductor manufacturing—and marvel that tech was approaching them. That comparison is no longer fruitful. The hyperCAPEX companies are the biggest spenders, period. Even Meta—the runt of our litter—outspends Saudi Aramco, TSMC and all three US telcos combined by tens of billions.
Get Platformonomics Updates By Email
Our hyperCAPEX cohort took 25 years to pass a trillion dollars in cumulative spending. One year later, they hit $1.6 trillion, and are on pace to blow past $2 trillion in 2026.
We’re just two sticks short of a hockey team here. All four hyperCAPEX companies spent a record share of revenue on CAPEX in 2025, led by Meta (44%) and Microsoft (39%).
This chart is going to be an absolute banger next year. Free cash flow —operating cash flow minus CAPEX —has already inflected downward at Amazon and Meta, and all four companies are guiding to further increases in CAPEX.
| 2025 Operating Cash Flow | 2025 Actual CAPEX | 2026 Guidance CAPEX | |
| Amazon | $139.5 billion | $134.7 billion | $200 billion |
| $164.7 billion | $91.5 billion | $180 billion | |
| Meta | $115.8 billion | $72.2 billion | $125 billion |
| Microsoft | $160.5 billion | $118 billion | $140+ billion * |
* Microsoft doesn’t provide guidance for the calendar year
Amazon’s long-time ethos was a penny of free cash flow (i.e. not reinvested) is a penny wasted. But our other three companies have historically gushed free cash flow. In 2026, all but Microsoft could be burning cash.
Depreciation spreads the cost of capital goods over their useful life. The more you invest, the more cost to be allocated (including when GPUs burn up early in their depreciation schedule). The particular simulation we live in is governed by double entry accounting, so CAPEX first hits cash flow, then boomerangs back via depreciation to dent earnings.
Whee!
On to the individual companies…
Amazon
Amazon provides a unique window into data center economics by generously breaking out both AWS-specific CAPEX and revenue. After an unprecedented decline in 2023, AWS CAPEX doubled in 2024 and nearly doubled again in 2025. With company-wide guidance of $200 billion, AWS spend could easily exceed $150 billion in 2026.
With visibility into both revenue and CAPEX, AWS offers the clearest view of capital intensity amongst the hyperCAPEX companies. CAPEX as a percentage of revenue drifts down from 2015 to 2023, then zooms upwards to over 75%. It is like they started DoorDash-ing GPUs!
We can benchmark the capital intensity of AWS—an experienced global data center operator—against some of the public nanoneoclouds. How much investment does it take to generate a dollar of revenue (never mind profitability)? Notice that Oracle, despite claiming to be “asset pretty light”, spends 3.7 times as much as AWS on infrastructure per dollar of revenue.
Most Common CAPEX Error: Attributing all of Amazon’s CAPEX to AWS. That erroneously shifts almost $40 billion of retail CAPEX to data center infrastructure, inflating Amazon’s infrastructure spending in the average drive-by AI CAPEX analysis.
What I’m Pondering: The AWS AI strategy appears to channel Stalin’s “Quantity has a quality all its own” mantra, with Trainium as the T-34 of AI chips. How dependent is AWS on scarcity everywhere else? Do they also need a competitive AI stack on top of all that hardware?
Fun Fact: Amazon disclosed more than $10 billion of infrastructure acquired but not yet paid for at the end of 2025. AWS Accounts Payable alone is bigger than most neoclouds.
After skimping on CAPEX for a few years—flat to even down and skipping their historical refresh cycle—the giant awakened in 2024 (+ 63%), accelerated in 2025 (+ 74%), and issued guidance to double their CAPEX in 2026. It is time to accept we’re not going to get a space elevator from Google.
Google’s CAPEX spending is the cleanest of the hyperCAPEX quartet. It’s the only one that doesn’t use finance leases, though it has recently issued corporate debt, including a 100-year bond that was ten-times oversubscribed—a notable statement of confidence by both Google and investors (and maybe even more so for the European currencies the bonds were denominated in).
But Google is the most opaque about where the spend is going, as they have multiple infrastructure-hungry businesses. Search and YouTube made Google the biggest CAPEX spender before the AI inflection. Google Cloud has been ramping, and now the company needs infrastructure for lots of first- and third-party AI workloads. (There was a time when Google bristled at the kind of Gemini bundling it’s now doing everywhere).
Google—like Amazon—is a reminder of how long CAPEX ramps can take, even for companies with global data center fleets and well-oiled procurement pipelines. It took more than a year after the ChatGPT moment (and internal “Code Red”) for infrastructure deployment to materially accelerate.
Most Common CAPEX Error: Attributing Google’s spend to any particular business. Google’s spend is spread across Search, YouTube, Google Cloud, AI training and inference.
What I’m Pondering: Having lived through Microsoft’s encounter with the web, it’s hard not to see parallels with Google. Companies can win the initial battle and still lose the war. Google still faces existential risks to its search business—arguably the best business in history. Adverse outcomes range from higher COGS impairing search economics to search being displaced altogether by LLMs.
Meta
Like AWS, Meta cut CAPEX in the first year after the ChatGPT moment as it revectored (yuk!) to AI. Spending then jumped 40% in 2024 and 84% in 2025. Even with 2026 guidance implying another 73% increase, Meta will still be our smallest spender. But the stated ambition is to eclipse all, with Zuckerberg saying (in english, not latin) Meta will “have industry-leading levels of compute”.
Keeping up with the hyperCAPEX Joneses has pushed Meta into off-balance-sheet machinations (Hello Enron!). Its Louisiana data centers (Project Hyperion) sit in a special purpose vehicle. Meta contributed assets, owns 20% of the $27 billion project, is committed to pay to use the data centers (booked as OPEX, CAPEX’s far less attractive sibling), and guarantees returns to other investors if the lease isn’t renewed. Meta’s auditor doesn’t entirely agree with Meta’s claim to have no control over the entity, which is necessary for Meta keep the project off the balance sheet.
What I’m Pondering: How does Meta earn a return on “industry-leading levels of compute”? The other hyperCAPEXers have 11- or 12-digit cloud businesses that foot a lot of their infrastructure bill. Is ad optimization or synthetic engagement really enough to justify hundreds of billions of CAPEX? Or does Meta have to get into the cloud platform business? Its history as a platform provider is a succession of rug pulls, so a Meta cloud would be a battle against not only AWS, Azure and GCP, but also its own genetics.
Microsoft
Microsoft was first to experience the ChatGPT moment via their OpenAI partnership and accelerated accordingly. CAPEX was up 45% in 2023 (the first time they materially outspent AWS), up 83% in 2024, and 56% in 2025. Microsoft doesn’t give guidance for the full calendar year, but their 2026 spending likely will exceed $140 billion. At that spend, Microsoft would drop from our top infrastructure spender of the last two years to third in 2026.
Most Common CAPEX Error: Microsoft’s finance leases are often neglected. In 2025, about a quarter of their spend, $29 billion, was via finance lease. So Microsoft’s spending on those charts about crazy levels of AI spending is often actually understated.
What I’m Pondering: What the hell happened at Microsoft in late 2024? What did Amy see? OpenAI’s discontinuity was dubbed “the blip”, while Microsoft had the “The Big Pause” (an episode begging for a more dramatic name). The company slowed data-center builds, abandoned gigawatts of leases, and gave investors guidance that CAPEX growth would slow which was reversed a quarter later. This occurred while OpenAI and Microsoft were renegotiating their relationship to non-exclusivity. Microsoft has its foot back on the gas, but it isn’t fully floored and the direction of travel is different. What are they optimizing for? (besides disembarking from OpenAI’s “To Infinity and Beyond” CAPEX train).