Six of the largest companies in the world spent $579 billion on property and equipment in the last twelve months, most of it on data centers. That spending lands directly on the cash flow statement: capital expenditure comes out of operating cash flow, what remains is free cash flow, and free cash flow yield is one of the four inputs to the value pillar in this site's quality-growth score.
The arithmetic is unforgiving. Microsoft, Alphabet, Amazon and Meta all post quality and growth scores in the eighties and nineties, and value scores in the forties and fifties. The screen marks them down, partly for the spending and more for the price.
There is a version of this site where I quietly explain away a score I don't like. This isn't that — I think the model is right, and the interesting question is what an investor should do with a number that is correct and inconvenient at the same time.
One disclosure before the numbers, because it is the obvious objection: I own Alphabet, Amazon, Microsoft, Apple and NVIDIA, and I added Meta and Microsoft on the pullback over the last few months. I do not own Oracle or Tesla, which are the two names this argument will exclude.
Where these numbers come from
Before any of the argument, the source, because this is the sort of article where a reader should be able to check me.
Every capital-spending figure below is the filed number from each company's own cash-flow statement — us-gaap:PaymentsToAcquirePropertyPlantAndEquipment, assembled trailing twelve months from the SEC's XBRL company facts. Not an estimate, not a vendor's summary, and not a number I inferred by subtracting one thing from another. You can pull the same file I did at data.sec.gov/api/xbrl/companyfacts/CIK0001652044.json for Alphabet, and the equivalent CIK for each of the others.
The periods differ slightly because fiscal years do. Alphabet, Amazon, Meta, Microsoft and Tesla run trailing twelve months to 30 June 2026. Oracle runs to 31 May, Apple to 27 June, NVIDIA to 26 July. That is close enough to compare how hard each one is spending, and not close enough to rank anybody by a decimal place.
What the numbers actually say
Those five builders spent $566 billion on property and equipment in twelve months. Add Tesla and it is $579 billion. That is not a forecast, and not an analyst's estimate of an AI budget. It is cash that has already left the building.
The model has no opinion about artificial intelligence. It reads what that spending did to the cash.
The gray bar is operating cash flow. The gold is what survived as free cash flow after the spending. Alphabet turned $186 billion of operating cash into $53 billion of free cash. Microsoft kept $67 billion of $183 billion. Amazon and Oracle converted none of it — both report negative free cash flow for the trailing year.
Amazon is the one that stops people. A company generating $161 billion of operating cash flow, one of the great cash machines ever assembled, reports negative free cash flow because it spent $173 billion on infrastructure. Nothing is wrong with the business. The cash simply went somewhere.
Why the score falls where it does
Split into its three legs, the pattern is almost monotonous. Quality holds up: 87 at Microsoft, 86 at Meta, 85 at Alphabet. Growth holds up: 91, 92, 61. Value is where every one of them breaks — 43, 53, 51 — but the spending turns out to be only part of why.
The value pillar reads four things: the PEG ratio, which carries the most weight; the company's P/E against its own history; free cash flow yield; and EV/EBIT. Only one of those is a cash-flow measure. Free cash flow yield is the leg capital spending hits immediately, and it carries 0.4 of the pillar's total weight of 2.4. The PEG ratio and EV/EBIT are built on earnings, which capital spending reaches only later, as the new equipment depreciates. A company that spends 71% of its operating cash flow, as Alphabet did, is going to look expensive on free cash flow yield however good the underlying business is.
So the spending is a real drag on the score, but it isn't the biggest one. For Microsoft, Alphabet and Amazon, about half of the gap between their value score and a perfect score comes from the PEG ratio — what investors are paying relative to expected growth. Free cash flow yield accounts for between a fifth and just under a third of it, and Meta's gap is spread more evenly across all four. Take that leg out entirely and their overall scores would rise by only one to three points. Alphabet and Amazon even trade near the cheapest P/E of their own recent history. The price against expected growth marks them down most; the spending adds a second, smaller weight on top.
Apple is the exception that keeps this honest. It converts 93% of operating cash flow into free cash — $10 billion of spending against $147 billion of operating cash — and it still scores 19 on value, the lowest number on the page. Capital spending explains nothing about Apple. It is simply expensive on the other measures. If you take one methodological point from this article, take that one: the value pillar is not a capex detector, and any story that treats it as one is overfitting.
Placed against each other the shape is clear. The builders cluster low and to the left. NVIDIA sits alone in the top right, converting 94% of its operating cash into free cash, because in this cycle it is the seller rather than the buyer. And Apple sits off the pattern entirely — high conversion, lowest value score — which is the visual version of the paragraph above.
The model is not wrong
I want to deal with the obvious objection, because I think it is mistaken.
The argument goes: the value pillar is naive, it cannot tell the difference between a company wasting money and a company investing in its future, and penalizing Alphabet for building capacity is a flaw to be corrected.
I don't believe the model is unfairly punishing capex-heavy compounders. The reality is that they are spending a lot and it is impacting free cash flow. That is what the value pillar is designed to show you. A screen that quietly forgave capital spending because the spender had a good story would be worth less, not more — and it would have to forgive Oracle and Tesla on the same page as Alphabet, using the same exemption.
There is a deeper reason not to patch it. The moment you add a rule that says "but this kind of spending is good," you have replaced a measurement with a forecast. The measurement is durable: it will say the same thing in five years about a company in the same position. The forecast is a bet on this particular cycle, dressed up as methodology. I would rather have a score that is occasionally inconvenient than one that has my current opinions baked into it.
The model does its job. It turns the numbers into a score. The question is what you, as the investor, do with it.
That distinction matters more than anything else on this page. A score is a summary of what has been reported. It is not a prediction, and it was never meant to substitute for a judgment about the future.
Where I come down
My position is that this spend is not forever.
Amazon, Alphabet and Meta are strong cash-flowing businesses. They may never see a return on these particular data center buildouts — I want to be honest that I am not claiming they will. Plenty of infrastructure booms have ended with more capacity than demand, and the people who financed them did not get their money back. That is a real possibility here and I am not going to argue it away.
What I am claiming is narrower. When the spending slows, these companies replenish their balance sheets quickly, because the underlying machine throws off cash at a rate the capital spending is temporarily hiding. Alphabet's $53 billion of free cash flow is what is left after $132 billion of building. The moment that building normalizes toward maintenance levels, the free cash flow line does not improve gradually. It steps.
In the best case the buildout becomes a catalyst that improves free cash flow rather than suppressing it, because the capacity earns something. My thesis does not require that. It requires the businesses underneath the spending to still be what I think they are.
That is a weaker claim than the one usually made for AI capital spending, and I prefer it that way. It does not need the data centers to pay off. It needs Google to still be Google.
The number that would change my mind
Revenue.
If revenue started to slow, it would tell me they did not have the safety net to actually back all this spending — that the capital was going out the door without a business growing fast enough behind it to absorb the mistake. A company can survive overbuilding. It cannot easily survive overbuilding while its core decelerates, because then the spending is not early, it is desperate.
It has not slowed. Meta is growing at 27.7%, Alphabet at 20.1%, Microsoft at 17.8%, Amazon at 15.8%, trailing twelve months. The part people lose track of is the base those rates sit on. Amazon is compounding at 15.8% on three quarters of a trillion dollars of revenue.
So the test I set has a reading, and today it says hold. I would rather publish the tripwire in advance and be held to it than explain afterward why whatever happened was consistent with what I always thought.
Two names this argument does not cover
This is where a thesis earns its keep, by saying what it excludes.
Oracle is spending 174% of its operating cash flow, against a debt-to-equity ratio of 3.89. That is not a fortress balance sheet absorbing a spending cycle. Alphabet's D/E is 0.19, NVIDIA's 0.17, Microsoft's 0.29, Meta's 0.43, Amazon's 0.46 — they are funding this from surplus. Oracle is funding it with leverage, and the recovery I have described depends on having something to recover with. A company that borrows to build has a second way to be wrong: the capacity can be fine and the financing can still break it.
Tesla is the other one, and it is my own test firing. Trailing revenue growth of 11.8%, the slowest on the board, while still spending $12.9 billion. And on a fiscal-year basis, which is the harsher and arguably more honest read, Tesla's revenue fell 2.9%. Its value pillar is 6 — not 43 or 51, but 6. That is the model saying something categorically different from what it says about Alphabet, and it is saying it correctly.
I do not own Oracle or Tesla. I would rather point that out myself than have you notice it and wonder whether the thesis was reverse-engineered from my holdings. The test I set excludes two companies, and they happen to be the two I do not hold — which is either evidence the test is doing real work, or evidence I built it to flatter my book. You are entitled to decide which, and the numbers to check me are above.
What I would actually tell you to do with it
Read the value pillar as a description, not a verdict.
When Microsoft shows 87 on quality, 91 on growth and 43 on value, the model has told you something true and useful: this is an excellent business whose cash generation is currently suppressed, and you are not getting it cheaply on that measure. What it cannot tell you is whether the suppression is temporary. That requires a view about what happens when the buildout ends, and a view is not a number.
The practical version is that a low value pillar on a high-quality name is a question, not an answer. Sometimes the question resolves to "expensive," which is Apple. Sometimes it resolves to "spending heavily against a growing business," which is Alphabet. The score cannot tell those apart, and it is not supposed to. You open the cash flow statement and decide.
I have a view. It might be wrong. What I have tried to do here is say exactly which number would prove it wrong, show you that number rather than wait to be asked for it, and cite where every figure came from so you can check the whole thing without taking my word for any of it.
I own Alphabet, Amazon, Meta, Microsoft, Apple and NVIDIA at the time of writing, and added to Meta and Microsoft within the last few months. I do not own Oracle or Tesla. Capital expenditure, operating cash flow, free cash flow and revenue growth are trailing-twelve-month figures assembled from SEC XBRL company facts; pillar scores are from this site's published edition and can be checked on each company's page. Nothing here is a recommendation — invest at your own risk.