$500b in ai spending: disaster or disguised cash machine?

in early 2026, amazon, alphabet and microsoft guided to a combined ~$500 billion in capital expenditure. half a trillion dollars. in one year. the market panicked. microsoft dropped 10% in a day, amazon fell 6%.

i’d been trying to make sense of these numbers when i came across a blog post by @deadneurons that framed the investment through a cash flow lens rather than a gaap earnings lens. the author clearly had a model behind the numbers, but the assumptions weren’t fully explained and the derivations were hard to follow. i wanted to verify the claims but couldn’t, because the steps between “here’s the capex” and “here’s the return” were mostly hidden.

claude opus 4.6 had just come out, and this felt like the right task to throw at it. i rebuilt the analysis from scratch. the goal: a step-by-step model where every assumption is labeled, every sourced fact is cited, and every number can be traced back to its inputs. the approach was to aim for conservative estimates throughout; lowest reasonable revenue, highest reasonable costs. if the model still showed decent returns, it would mean something.

the core insight is surprisingly simple. when you spend $500b on data centers, about 30% goes into buildings that last 20 years. but the servers inside them only last 5. so the natural analysis window is 5 years, and within that window, depreciation on the buildings drags down reported earnings even though the buildings will keep generating revenue for another 15 years after the servers are gone. the income statement says you earned 6% annualized. the cash flow tells a different story.

@deadneurons made a good analogy in his post: these companies are starting to look like reits. and it’s true. nobody evaluates a rental property on net income after depreciation. the entire reit industry reports “funds from operations,” which is just net income plus depreciation, because decades ago real estate investors figured out that gaap earnings are misleading for long-lived physical assets. data centers are buildings full of equipment. the accounting should be evaluated the same way.

the model runs on eight labeled assumptions: capex split, depreciation schedules, revenue estimates, operating costs, tax rates, maintenance. every one is sourced where possible and flagged as a judgement call where not. the revenue assumption alone (how much annual revenue does $1 of data centre capex generate?) spans a $185b range depending on how aggressive you want to be. i tried to stay on the cautious side, but that’s a judgement call too.

the result: under these assumptions, the cash flow return is roughly 4x the gaap return. a mechanical gap caused by how accounting handles long-lived assets. whether that gap matters depends on whether the revenue actually shows up. the cash flow case assumes these data centers fill up with paying customers; if cloud demand disappoints, neither lens looks good. right now, cloud revenue is growing at 34% blended across the three companies.

the analysis also stress-tests the model against the scenarios that break it: hardware obsolescence forcing early replacement, weak demand leaving data centers half empty, financing costs, customer concentration risk. it tries to show what has to go wrong, and by how much, for the investment to fail.

you can read the full analysis here.