Last week Microsoft reported good earnings and told investors it would spend about $150 billion on capital expenditure this year. The stock dropped 10%, which erased $357 billion in market capitalisation in a single day, which is the largest single-day dollar loss in the history of publicly traded companies. On Thursday Amazon said it would spend $200 billion, and fell 8-10%. Alphabet, which had the good sense to pair its capex guidance with a huge earnings beat, got away roughly flat despite guiding to $175-185 billion, about $60 billion above what analysts expected.
The combined 2026 capex budget for these three companies is approximately $530 billion, which is roughly the GDP of Sweden, or about 60% of the entire US defence budget. The market looked at this number and sold everything.
I think the market is wrong, and I think the reason it is wrong is kind of funny, because it involves forgetting something you learn in the first week of an accounting course.
Here is how the bear case works. You spend $530 billion on GPUs, servers, and data centres. About two-thirds of that, roughly $333 billion, is servers and GPUs that depreciate over 5-6 years. The remaining third, about $197 billion, is buildings, power infrastructure, and land, most of which depreciates over 15-25 years. Together, those schedules produce about $70 billion per year in total depreciation expense ($61 billion from the servers, $9 billion from the buildings), which lands in the cost of revenue line, where it sits on top of the actual cash costs of running the things (electricity, cooling, maintenance), and together they squash your gross margins. Meanwhile the data centres take a year or two to fill up with paying customers, so in the early period you are eating full depreciation charges on half-empty buildings, which looks terrible.
If you model the full profit and loss for a single year’s capex cohort over the server useful life of about 5.5 years, assuming all the capacity is new and needs to ramp from zero, you get cumulative net income of about $137 billion on $530 billion invested. That is a 26% total return, or about 4-5% annualised, which is genuinely terrible.
Now, that raw model is too pessimistic in a couple of ways. About 35% of each year’s capex replaces fully depreciated equipment that already has paying customers, so $100 billion or so in annual revenue does not need to ramp at all, it is there from day one. The incremental operating expenses (sales, R&D, overhead) on new capacity are also much lower than for a business built from scratch, because the platform, the sales teams, and the engineering organisation already exist. Adjusting for both of these, which is reasonable, gets you to about $392 billion in cumulative net income, or roughly 11% annualised. That is a more realistic GAAP picture, but it is still the sort of return that makes a technology investor reaching for the sell button.
This is the analysis that drives the sell-side downgrades and the panicky headlines. It is also, I think, the wrong way to evaluate this kind of investment.
Forget the income statement for a moment. Think about cash.
You hand $530 billion to Nvidia and various construction companies in 2026. That cash is gone. In return you get buildings full of GPUs. Those buildings full of GPUs generate revenue. The question is simply: how much cash comes back, and when?
In steady state, once the data centres are reasonably full (call it years 3-5), the 2026 cohort generates about $100 billion per year in operating profit. Now, that $100 billion is calculated after subtracting $70 billion of depreciation. Depreciation is a real economic concept, representing the fact that your servers are slowly wearing out, so I am not going to tell you it is fake. What I am going to tell you is that it is not a cash cost. You do not write Nvidia a cheque for $70 billion every year as your GPUs get older. The cash left in 2026. The depreciation is just the accounting system’s way of remembering that it happened.
So your actual annual cash generation is the $100 billion in operating profit plus the $70 billion in depreciation that did not actually require you to spend any money, minus maybe $25 billion in genuine maintenance capex (replacing failed drives, patching things up, keeping the lights on). That $25 billion is important, because depreciation is not purely fictional: it signals that these assets are wearing out and will eventually need replacing with real cash. The maintenance line is where that reality shows up. After accounting for it, you get free cash flow of about $145 billion per year, on $530 billion invested, which is a 27% annual cash yield. The cohort pays for itself in about 3.5 years.
To recap: the raw GAAP analysis says 4-5% annualised, which is awful; a more realistic GAAP analysis says 11%, which is mediocre. The cash flow analysis says 27% annual yield, pays back in 3.5 years, which is excellent. Same $530 billion. Same assets. Same customers. Completely different conclusions about whether it was a good idea, depending on which number you choose to care about.
This is not a magic trick. The gap has a specific, boring explanation, and it lives on the balance sheet.
About $197 billion of the $530 billion goes into long-lived assets: buildings, power connections, fibre, and land. The buildings and infrastructure depreciate over 15-25 years; the land does not depreciate at all. When the GAAP P&L says “this cohort only earned $392 billion against $530 billion invested,” the missing $138 billion is not lost. It is sitting there in the form of data centre buildings that will keep generating revenue for another two decades. The GAAP framework just has no way of telling you that within a 5.5-year analysis, because the buildings have not finished depreciating yet.
There is a simpler version of this, which is: you buy a house for $500,000. You rent it out for $30,000 a year after expenses. Your accountant allocates $100,000 to the land (not depreciable) and $400,000 to the building, which gets depreciated over 27.5 years at about $14,500 per year. Your annual “net income” is $30,000 minus $14,500 in depreciation, or $15,500. Over ten years you have earned $155,000 on a $500,000 investment, which your accountant will solemnly inform you is a 31% total return.
Your cash return, meanwhile, is $300,000 over ten years, plus you still own a house worth at least $500,000. Nobody in the history of property investing has evaluated a rental property on net income after depreciation. Everyone uses cash-on-cash return. This is so well understood that an entire asset class, REITs, reports a metric called “funds from operations” which is literally just “net income but we add back the depreciation because come on.” Yet when Amazon builds a data centre, which is the same physical object as a building that a REIT would own, analysts evaluate it on earnings per share, which includes the depreciation, and panic.
I am being somewhat unfair to the bears, so let me steelman their position properly.
The 27% FCF yield assumes maintenance capex of about $25 billion a year. If hardware gets obsolete faster than expected, and there is a reasonable case for this given that Nvidia releases a new GPU architecture every 12-18 months, then “maintenance” is really “complete replacement” and costs more like $60 billion a year. That drops the FCF yield to about 21%, which is still good but less spectacular.
The scenario that genuinely breaks things is accelerated obsolescence combined with disappointing demand. If you are replacing hardware every three years and your data centres are half empty because enterprises decided they did not actually need that much AI compute, your FCF yield falls to maybe 14%, which does not justify a 30x earnings multiple. That is the real risk: not “they are spending too much” but “they are spending roughly the right amount on assets that will be useful to fewer customers for less time than they expect.” Those are different problems with different implications, and the market is mostly trading on the first one, which is the less concerning of the two.
Here is one more number that I think is clarifying. By 2027, three overlapping capex cohorts (2024, 2025, and 2026) will be generating about $250-260 billion per year in total depreciation. To maintain 60% cloud gross margins, combined cloud revenue from these three companies needs to reach about $460-480 billion, which implies blended growth of roughly 25-27%.
That sounds like a lot until you look at the current growth rates: AWS is at 24%, Azure at 39%, Google Cloud at 48%. None of them need to accelerate. They can all slow down meaningfully and still clear the bar. The model breaks at about 18-20% blended growth, which would require a sharp deceleration from where they are today, not a gentle one.
I find this more useful than the headline capex number because it gives you something concrete to watch. You do not need to have a philosophical view on whether artificial intelligence will transform the economy. You just need to track quarterly cloud revenue growth and check whether it is above 20%. If yes, the capex is probably fine. If it starts drifting towards 15%, start worrying.
My theory is that the market sees “$200 billion capex” in an Amazon press release and does some quick mental arithmetic: that is more than Amazon’s entire $140 billion in annual operating cash flow, which is a lot. Analysts model the depreciation, see the margin compression that will show up in 18 months, lower their EPS estimates, lower their price targets, and the stock sells off. Nobody in this chain is making an error, exactly. The depreciation really will compress margins. The EPS estimates really should come down. The price targets really do follow from the models.
The problem is that the models use the wrong metric. They evaluate infrastructure investments on accounting earnings rather than cash returns. This is a category error, the sort of thing that is completely understandable given how technology companies have always been evaluated, and also completely wrong given that the marginal dollar of capex is now going into concrete and copper and cooling systems rather than into software engineers.
So you have three of the most dominant businesses in history, with combined operating cash flow north of $400 billion a year, investing in assets with 27% cash yields, and the stocks sell off 10% because the income statement will look bad for a couple of years. That is probably an opportunity. It will not feel like one for a while, because the depreciation will hit the P&L exactly as the bears predict, and quarterly EPS will disappoint, and there will be many confident people on television explaining why the AI bubble has burst. You will have to wait for the cash to prove them wrong, which it will, because cash always wins in the end.