← Cross-cutsCross-cut · Hyperscalers

The trillion-dollar bet

Microsoft, Amazon, Alphabet, Meta and Oracle are, together, spending close to half a trillion dollars a year building AI infrastructure. Read side by side, these five Teardown studies turn the “AI capex” headline into a sharper question: each is making the same bet, but with a very different ability to pay for it — and the gap between funding it from today’s profit and borrowing against tomorrow’s is the whole story.

The same bet, at five different scales of risk

Every one of these companies has roughly tripled its capital spending in two years to chase AI compute. The numbers are staggering and converging: Microsoft guides to $120–190B, Amazon to ~$200B, Alphabet to $175–185B, Meta to $125–145B. Microsoft is now spending more on AI infrastructure in a single year than it earns in net income — the most expensive capital program in corporate history. The cross-cut’s first insight is that this is not five separate strategies; it is one industry-wide arms race, and the differentiator isn’t whether to bet but whether your existing business can absorb a bet this size.

Follow the money that funds the bet

The decisive variable is the profit engine underwriting the spend, and it sorts the five cleanly. Alphabet and Meta fund AI from advertising — ~72% and ~98% of revenue respectively — the highest-margin, most reliable cash machines in business. Microsoft and Amazon fund it from enterprise cloud plus (for Amazon) a ~$68.6B ad business. Oracle is the outlier and the warning: its capex leapt against a balance sheet already stretched to a −$24.7B free cash flow and $100B+ of debt, betting on a reported ~$300B compute backlog anchored to a single cash-burning customer. Same bet; one of these is being financed by an unmatched ad monopoly and another by leverage.

The shared question: does the spend earn its cost of capital?

Read the five weighings together and they converge on one contested verdict: nobody can yet prove the capex earns its return. The bull evidence is real and recognized — Azure reaccelerating to +40%, AWS to +28%, Alphabet’s $460B+ and Oracle’s $553B backlogs. The bear evidence is just as concrete — Amazon’s FCF fell 71%, Meta’s could swing negative, and the industry faces a depreciation bill running toward ~$400B a year against AI hardware that obsoletes fast. The studies independently land this as a genuine deadlock because the two sides measure different horizons: backlog is the contracted future, depreciation is the arriving present. Whoever’s revenue keeps pace with the depreciation wins; whoever’s doesn’t is destroying capital.

Alphabet is the one being attacked from both ends

The cluster has an asymmetry worth isolating. For Microsoft, Amazon and Oracle, AI is purely an opportunity to capture. For Alphabet, it is also the knife at the throat of the engine paying for it: generative answer engines threaten to intercept the ad-bearing search queries that fund ~72% of revenue (one study cites clicks halving under AI summaries). Alphabet must spend AI capex to defend against the very technology eroding its cash machine — the only player in the cluster financing the bet from a business the bet itself endangers. That is why its study frames the question as substitution-vs-expansion, where the others ask merely whether the spend pays off.

Where they agree — and where they split

All five agree the AI build-out is real, that they have no choice but to spend, and that the binding constraint is shifting from chips to power and data-center shells. They split on durability of the funding source. The ad-funded players (Alphabet, Meta) have the most reliable engine but, for Alphabet, the most exposed one. The cloud-funded players (Microsoft, Amazon) have recognized, contracted demand but thinning free cash flow. Oracle has the steepest backlog growth and the weakest balance sheet — the highest-beta way to play the same theme. The demand for AI compute is not the question any of these studies doubts. Whether the spending compounds the moat or strains it — and which balance sheet cracks first if the revenue lags — is.

The cluster at a glance

CompanyCloud / AI scale2026 capex betProfit engineCan it fund the bet?
MicrosoftNASDAQ:MSFTAzure +40% · AI ~$37B run-rate~$120–190B/yr guidedCloud + 450M M365 seats$101.8B NI — capex ≈ net income
Amazon.comNASDAQ: AMZNAWS +28% · ~57% of op income~$200B 2026 (from $131.8B)AWS + ads (~$68.6B)FCF fell 71% to $11.2B
AlphabetNASDAQ: GOOGL / GOOGCloud +48–63% · $460B+ backlog$175–185B 2026 (from ~$75B)Ads ~72% of revenueSearch funds it (under threat)
MetaNASDAQ: METANo cloud · AI in the ad stack$125–145B 2026 (from $72.2B)~98% ads · $102.5B FoA op incomeFCF could swing negative
OracleORCLOCI +84% · $553B RPO (+325%)$50B FY26 (from $21.2B)~77% recurring license/supportFCF −$24.7B · debt >$100B

Figures as of each study’s stated date (2026-06); fiscal years differ. Capex figures are company guidance and definitions vary (some include finance leases). See each teardown for sourcing and the full weighing.

The five studies — and the question each turns on

Microsoft CorporationNASDAQ:MSFTWill the most expensive capex program in corporate history compound Microsoft's advantages, or will AI hardware depreciation outrun AI revenue and strain them?Read the full weighing →Amazon.com, Inc.NASDAQ: AMZNIs Amazon's enormous AI/cloud capex bet a generational moat or a margin trap — and does the AWS-plus-ads profit engine stay durable enough to fund it?Read the full weighing →Alphabet Inc. (Google)NASDAQ: GOOGL / GOOGIs generative AI a substitute that erodes Search and the ad engine that funds everything, or an expansion that grows queries while Cloud and TPUs turn a one-engine company into two?Read the full weighing →Meta Platforms, Inc.NASDAQ: METAWill the AI capex pay off before it strains the model — i.e. can the proven ad-stack half of the bet underwrite the unproven 'superintelligence' half at $125–145B a year?Read the full weighing →Oracle CorporationORCLIs Oracle's $553B AI-compute backlog a generational reinvention, or an over-leveraged bet on a single cash-burning customer at the top of the AI capex cycle?Read the full weighing →

This is the kind of reading the Desk does for you

A cross-cut takes five teardowns and one arms race and asks what they say together. The Desk does the same for the companies you actually own — your thesis, the rivals that move it, and the tripwires that would change your mind.

See the Desk →