Pricing the AI buildout: an interactive model
The five hyperscalers are spending on AI infrastructure at a pace with no precedent. This piece explains the four things that actually drive what that spending is worth — growth, margin, terminal capex, and the discount rate — and hands you an interactive model to set those assumptions yourself. Educational, not a prediction, not investment advice.
Microsoft, Alphabet, Amazon, Meta, and Oracle are collectively committing capital to AI data centers at a pace with no precedent — capital expenditure now runs two to four times depreciation at each. What that spending is worth depends entirely on four things: how fast revenue grows, what margin it earns, whether capex ever normalizes, and the rate you discount it at. The model this piece introduces makes all four explicit and hands you the dials.
The point is not to tell you what these companies are worth. It is to make it easy to try your own assumptions and see, immediately, what they imply — because the answer is far more sensitive to those four inputs than most headlines about the "AI trade" let on.
What the model holds fixed
The structure below is the same for every company and every scenario. It is a cash-flow model — its value drivers are cash in and cash out, not accounting earnings — and it holds a deliberate set of choices constant so that the only things moving are the assumptions you control.
The structure:
- A 15-year explicit forecast, then a Gordon terminal value at 4% perpetual growth (roughly nominal GDP).
- Three-phase growth: the starting rate holds for a short plateau, fades to half by year 10, then to 4% by year 15.
- Capex glides from today's elevated level down to a terminal intensity over 10 years, then holds.
The accounting choices:
- Depreciation is non-cash. It is modeled from capex vintages, but only its tax shield affects value. The "depreciation wave" from the buildout is not itself a value driver — a point that matters enormously to reported margins and not at all to cash.
- Leases are treated as a phased cash expense — not-yet-commenced commitments spread over their life — and kept out of invested capital.
- Base-year revenue, capex, cash, debt, shares, tax rate, and lease commitments come from each company's latest 10-K.
Your four dials
You control four inputs. Pick a scenario preset to load a coherent set of assumptions, then move any slider to explore your own. The presets are built from each company's own history: growth from the cleaned last-five-year range, margins anchored to through-cycle levels rather than today's peak, and capex reverting toward pre-AI norms (bull) or staying elevated (bear).
The first three dials are the real risks in the story:
- Growth — does AI revenue actually materialize?
- Operating margin — do real costs (not depreciation) compress it?
- Terminal capex intensity — does the buildout ever normalize, or is this the new baseline?
The fourth is the discount rate (WACC), and it is scenario-specific by design: a bull world is a lower-risk world and earns a lower discount rate; a bear world demands a higher return. The presets pair each scenario's cash flows with its own rate, but you can override it like any other dial.
One thing the model deliberately does not do: it outputs a fair value per share only. It shows no market price and makes no buy-or-sell call. Comparing that value to a price, and deciding what to do about it, is your step — not the tool's.
Early return signals
The scenarios are forward assumptions, but the near-term data gives them context. In the first half of calendar 2026, the cloud franchises' operating margins are landing in the mid-to-high 30s and expanding: AWS ~39%, Microsoft's Intelligent Cloud ~40%, and Google Cloud ~36% — the last up from roughly 18% a year earlier — on cloud revenue growing 30–80% year over year. Trailing returns on invested capital run about 18–32% across the group.
One caveat runs through all of it: a depreciation lag. Most of 2026's capex has not yet reached the income statement — it depreciates over roughly six years — so today's margins reflect cheaper past vintages. The depreciation crest arrives in 2027–2029. These are two quarters of data, not a trend. Because this model values cash flows, that lag changes the reported-margin optics more than it changes value — but it is exactly the thing to watch as you judge the margin dial.
How to read the output — and a caveat
Under the assumptions you select, the model returns a fair value per share. It deliberately does not show or compare against any market price, and it does not tell you whether a stock is a good investment. Whether these assumptions are realistic — and whether the resulting value makes something worth owning — is entirely your judgment.
This is an educational tool for exploring how assumptions drive value, not a prediction of what is likely to happen, and not investment advice.