Five Big Tech Companies Spent $556.9 Billion on Capex Against $542.9 Billion of Operating Cash Flow. The API You Build On Is Downstream of That. Here's the Margin Rule I'd Use.
In the first nine months of 2026, Microsoft, Alphabet, Amazon, Meta, and Oracle spent $556.9 billion on capital expenditure while generating $542.9 billion in operating cash flow, according to a Reuters analysis of LSEG data published October 30. That is the first time capex has exceeded operating cash flow in data going back to 2011. Combined free cash flow turned into a $14 billion deficit, against a $143.5 billion surplus a year earlier.
If your product calls one of these companies' models, or a lab that rents their data centers, you are downstream of that number. The useful question is what you should do about it, and the honest answer is smaller than the headlines suggest.
What the numbers say
The same Reuters analysis reports that the five companies returned 9.3% of their cash to shareholders in the period, down from 27.4% a year earlier. Buybacks fell 83% to $13.6 billion from $81.8 billion. Dividends rose to $37 billion from $33.9 billion. Total payouts of $50.6 billion were effectively funded from cash reserves, borrowing, or share sales.
The borrowing is visible. Amazon raised $48.1 billion of long-term debt in the third quarter. Alphabet raised about $29 billion from debt and about $12.4 billion from share sales. Operating cash flow itself hit a record $542.9 billion, up 29%, so this is not a story about weak businesses. It is a story about spending that has roughly doubled.
What the numbers do not say
They do not say prices will rise. Nothing in the Reuters analysis connects capex to API pricing, and I'd be inventing a causal chain if I claimed otherwise. Model prices have moved in both directions this year. Earlier this week I wrote about eight LLM price moves in nine weeks, some of them up to 80%, and about Gemini 4 Argon's announced price doubling on December 1.
They also do not say the spending is wasteful. A capex-heavy year can precede cheaper inference, because data centers built now lower the cost per token later. That has been the historical pattern with cloud compute.
What the numbers do say is narrower and still relevant: the companies setting your input costs are running with no free-cash-flow cushion. They are funding growth with debt and equity. When margins get squeezed at that scale, pricing is one of the first levers anyone looks at. I can't tell you the direction. I can tell you the lever is within reach.
The margin rule
Here is the rule I'd use for any feature that calls a model.
Price the feature so it still earns a profit if your model cost doubles. Not "can survive," but is profitable. Run your real numbers: tokens per task, times your current price per million tokens, times two. If the margin is still above zero after that, you have headroom. If it isn't, you are one pricing announcement from losing money on every customer.
For a subscription product, express it per user per month. If a typical user costs you $3 in model spend and you charge $15, a doubling takes your cost to $6 and you are fine. If a typical user costs you $9 and you charge $15, a doubling erases your margin. Fix that before the price moves, not after.
Two supporting habits make the rule cheap to keep. Build a swap layer, meaning one place in your code where the model name and provider live, so changing vendors is a config edit and not a rewrite. And don't prepay for tokens. A prepaid balance is a bet that prices stay flat, and a credit balance can't be refunded when a price drops.
Where this could be wrong
The counter-argument is that I'm drawing a conclusion from one quarter-by-quarter quirk. Capex can exceed cash flow in an investment cycle and normalize afterward, and the companies involved have access to credit markets that you and I don't. A deficit of $14 billion on more than half a trillion dollars of cash flow is small. Dividends still went up, which suggests the companies themselves are not treating this as an emergency.
If that is right, the doubling test is conservative and costs you nothing but a spreadsheet.
What I'd actually do
Open your cost model this week. Take your three most-used model calls, double the price, and write down which products go negative. For any that do, either raise the price, add a cheaper fallback model for easy requests, or cap usage per user. Then check that your provider and model names live in one file. That is a Saturday afternoon of work, and it protects you from this risk and from the next eight price moves.
Author
Lukas
@lukcombinator