Actionable insights — Amazon: The CapEx Equation

1. Split a capital budget into short-payback and long-life buckets before judging the burn

1. Break reported CapEx into (a) revenue-generating equipment bought close to deployment, and (b) long-lived structures built well ahead of monetization — they have completely different risk profiles and should never be judged with one number.
2. For the equipment bucket, compare the payback period to the contract length behind it. Payback shorter than the contract means the risk is operational, not speculative.
3. For the structures bucket, compare the build-to-monetization lag to the asset's economic life and how many equipment generations it will host — that ratio is what makes an early build defensible.
4. Only then interpret negative free cash flow: a timing mismatch on contracted demand is a different animal from spending into hope.

• Any lengthening of equipment payback or shortening of contract length; depreciation-life assumptions on the long-lived bucket; and whether the equipment bucket starts being bought further ahead of demand.

2. Test whether the buildout is contracted, not speculative — look for reserved forward capacity

1. Find the forward-capacity disclosure: what percentage of next year's (and the year after's) capacity is already reserved or committed.
2. Cross-check it against the backlog figure and the phrase management uses about supply — "demand exceeds capacity" is a claim you can validate against margin and pricing.
3. Watch margins during the buildout. If margins expand while capacity is being added at record pace, the new capacity is being absorbed at price, not discounted to fill.
4. Discount any capacity claim that isn't paired with either a backlog number or a reservation percentage.

• The reservation percentage rolling forward each quarter; backlog growth versus revenue growth; and the first quarter where margin expansion stalls while CapEx keeps rising.

3. Map a new market's demand as a barbell and locate the missing middle

1. Segment current demand into the extremes actually paying today (here: frontier labs and breakout applications at one end, narrow point use cases at the other).
2. Identify the middle that structurally should adopt but hasn't yet — usually existing production workloads with the largest absolute base.
3. Size the middle against the extremes: if it is the largest absolute pool, the market's duration is much longer than the current adopters imply.
4. Check whether the vendor's product roadmap actually reaches the middle (tooling, agents, applications) rather than only serving the extremes.

• Evidence the middle is actually converting (per-customer inference spend in ordinary enterprises, not labs); CPU/core-cloud growth reaccelerating alongside AI; agent-product usage disclosures.

4. Normalize a guide for calendar shifts and FX before calling it a slowdown

1. When guided growth decelerates, list the mechanical distortions first: promotional-event timing shifts between quarters, extra/fewer weeks, FX, divestitures.
2. Add those points back to get the underlying rate, then compare it to the prior quarter's underlying rate rather than the reported one.
3. Cross-check with the profit guide — if operating income is still guided to grow multiples of revenue, the margin story is intact and the revenue optics are noise.

• The reciprocal effect in the next quarter's comparison; whether the profit guide keeps outgrowing the revenue guide; and management quantifying the distortions itself (if it doesn't, be more sceptical).

5. Track the cash engine across the whole peer group, not one name

1. Aggregate trailing operating cash flow across the peer set and compare its growth to aggregate CapEx growth — the crossover point is when self-funding ends for the industry, not just for one company.
2. Rank the peers by how far through that transition each is (still positive FCF → near zero → negative), because the laggard sets the sector's financing narrative.
3. Then read each name's chosen funding instrument — retained cash, debt, or equity — as a confidence signal. Equity issuance from a cash-rich balance sheet is the loudest.
4. Recognize the reflexivity: once one peer's FCF turns negative, the sector's cost of capital and the market's tolerance for further CapEx both change.

• New debt issuance and its spread; any equity raise from a cash-rich balance sheet; the first peer to cut CapEx guidance; and whether aggregate operating cash flow growth reaccelerates enough to close the gap.

6. Weigh turning an internal advantage into a merchant product

1. When a company considers selling an internally developed component externally, separate the addressable-market expansion from the erosion of the exclusivity that made the component valuable.
2. Check whether the internal advantage is the component itself or the system around it — if the surrounding stack is the moat, selling the component is close to free upside.
3. Size the customer set that would buy it (here: firms running their own data centres) and whether they are also the platform's competitors.

• Any named external Trainium customer; whether AWS pricing advantages narrow afterwards; and, on grocery, whether purchase frequency and advertising inventory actually move rather than just grocery revenue.