Analysis
The defining financial characteristic of the hyperscalers for the last fifteen years was that they generated cash faster than they could spend it. That is what funded the buybacks, the dividends, and ultimately the multiple. CNBC's argument this week is that AI capital expenditure is testing exactly that characteristic, and Axios covered the same shift.
The mechanism is straightforward. Free cash flow is operating cash flow minus capital expenditure. Operating cash flow at Microsoft, Alphabet, Amazon and Meta continues to grow. Capex has grown faster. Data centers, power infrastructure and GPUs are not software costs -- they are physical assets with lead times measured in years and price tags that have to be committed before the revenue exists to justify them.
The accounting question sitting underneath is depreciation. Hyperscalers generally depreciate AI servers over five to six years, which spreads the cost and flatters near-term margins. Nvidia has been shipping a new architecture roughly annually -- Hopper, Blackwell, GB300 -- and the resale market for two-generation-old accelerators is thin. If the useful economic life is shorter than the book life, reported earnings are overstated and the correction arrives as a writedown rather than a miss.
“Nvidia has been shipping a new architecture roughly annually -- Hopper, Blackwell, GB300 -- and the resale market for two-generation-old accelerators is thin.”
The credit market has become the release valve. More than $400 billion of AI-related debt has been raised globally in 2026 by banks and technology companies, and neoclouds like Lambda and CoreWeave finance GPU purchases against contracted revenue. Pulse covered Lambda's $1 billion short-dated facility arranged by JP Morgan this week. Debt-funded capex keeps free cash flow optics intact while moving the risk onto the balance sheet.
The counterweight: none of this is evidence of a bubble by itself. Amazon spent through free cash flow building AWS for years and the trade was correct. Microsoft's Azure AI revenue is real and growing, and Salesforce's raised AI guidance this week argued the demand side is not imaginary. Capex that funds capacity a customer has already contracted for is investment, not speculation.
The distinction to hold onto is contracted versus speculative capacity. In the next earnings cycle, the disclosure that matters is not the capex number -- it is the share of new capacity with signed offtake behind it, and any change in server useful-life assumptions buried in the footnotes.
