Cloud giants keep spending on AI. When does cash catch up?

The second quarter of 2026 presented two compatible facts: demand for cloud services was strong, and funding the next stage of capacity was expensive. Google Cloud revenue rose 82%, AWS grew roughly 37% and Azure grew 43%. Alphabet nevertheless reported negative quarterly free cash flow, while Amazon’s trailing twelve-month free cash flow also turned negative.^[1][2][6] Revenue growth does not eliminate the financing interval before new facilities earn cash.

The useful question is how long today’s construction payments take to return through customer collections. Buildings, electricity connections and servers follow different schedules. Lease accounting adds another difference. A single sum of four companies’ budgets conceals those distinctions; investors need to understand the asset being purchased, when it can be used and who pays while it waits.

Start with the spending definitions

Microsoft expects approximately $175 billion of capital expenditure in calendar 2026. Alphabet guides to $195 billion–$205 billion; Meta to $130 billion–$145 billion, including finance-lease principal payments; Amazon management expects roughly $220 billion of cash capital expenditure.^[1][3][5][8] These describe a large investment cycle, but they are not a fully comparable measure of AI spending.

Their business coverage also differs. Amazon purchases assets for its fulfillment network as well as technology infrastructure. Other groups support offices, conventional cloud services and non-AI products. Without a disclosed allocation, group capital expenditure cannot all be assigned to AI, much less translated directly into one semiconductor or power-equipment supplier’s orders.

Microsoft illustrates the accounting problem. Extending the estimated useful life of certain data-center buildings from fifteen to twenty-five years changed the classification of some leases and reduced its calendar-year capex expectation from roughly $190 billion to $175 billion. The change concerns buildings and lease classification. It does not establish a longer GPU life or show that $15 billion of physical infrastructure was canceled.^[1]

A comparison should therefore retain cash purchases of property and equipment, new finance leases and lease principal repayments separately. They describe current cash spending, new asset commitments and repayment of existing obligations. Similar economic investment can produce different reported capex and free-cash-flow paths when those structures differ.

Profitable cloud services can coexist with a construction cash deficit

Alphabet generated $39.069 billion of quarterly operating cash flow and purchased $44.924 billion of property and equipment, producing negative free cash flow of $5.855 billion. Google Cloud generated $24.768 billion of revenue and $8.814 billion of operating income.^[2] A profitable segment and a group-level cash deficit can coexist. Segment operating profit is not the cash return on a newly constructed data center.

Meta spent $31.08 billion on capex, against operating cash flow of approximately $31.86 billion and free cash flow of $784 million. Amazon generated $161.4 billion of operating cash flow over the trailing twelve months, but free cash flow was negative $7.6 billion. Its second-quarter cash capex was $53.1 billion.^[4][6][7] Amazon’s twelve-month figures should not be ranked directly against quarterly figures at its peers.

Microsoft generated $55.4 billion of quarterly operating cash flow, spent approximately $35.8 billion on cash property purchases and reported $19.6 billion of free cash flow, alongside $5.6 billion of finance leases.^[1] That demonstrates strong cash generation, but lease structure, working capital and business mix also shape the comparison. The difference cannot all be attributed to AI efficiency.

The more informative test follows each company over time. If revenue and operating collections catch up with new capacity, the construction deficit may be absorbed. If spending keeps rising while collections slow, receivables increase or margins deteriorate, shareholders face a larger financing requirement. The direction of conversion matters more than one quarter’s headline.

Long contracts still have a revenue schedule

Alphabet reported $514 billion of cloud backlog and said slightly more than half should become revenue over the next twenty-four months.^[3] The figure provides contractual coverage across different durations and products. It is neither next year’s revenue nor a pool that can all be valued at the present cloud-services margin.

Microsoft reported $678 billion of commercial remaining performance obligations, with approximately 30% expected to become revenue in the next twelve months.^[1] Applying that percentage gives about $203.4 billion, leaving roughly $474.6 billion over a longer horizon. The companies’ measures cover different businesses and contractual definitions; their sizes alone do not rank the certainty of AI demand.

Contracts support returns through pricing, cancellation rights, delivery conditions and payment arrangements. Customer prepayments can reduce construction funding needs while creating future service obligations. Minimum purchases can improve visibility, but a supplier that misses its delivery date may still face remedies or penalties. Contract value does not resolve these terms.

Valuation must also reflect time and product type. A five-year service agreement has different costs and working-capital needs from a system delivered once. Turning backlog into a model requires a schedule for construction, revenue recognition and collection. Assuming all three happen together can make an expensive project look deceptively cash generative.

Earlier capacity can justify a higher purchase price

Alphabet plans to use more third-party capacity in the third quarter while building internal facilities, accepting some near-term margin pressure.^[3] That is a concrete commercial choice: wait for lower-cost owned capacity, or buy access sooner and capture revenue and customer relationships that might otherwise be delayed.

Consider an illustrative service capable of $1 billion in annual revenue. If owned capacity arrives six months later than rented capacity, and the rented service earns a 25% cash contribution margin, immediate access could generate $125 million over that interval. This excludes migration, deposits and integration costs. It illustrates timing value, not an Alphabet project return.

The reverse risk arises when a temporary shortage is met with a multiyear fixed purchase obligation. Investors then need to know when external capacity can be released, how long minimum payments last and whether rent overlaps with the opening of owned facilities. Early revenue must be compared with those later obligations.

This also explains why a cloud platform and its landlord need not earn the same return from a shortage. The landlord may receive rent and advance payments while the platform bears end-customer usage and pricing risk. Cash certainty improves further only if the contracts allocate those risks effectively. The actual agreement identifies the beneficiary.

Buildings and servers carry different risks

Amazon management separates data-center buildings from servers and networking equipment: building expenditure precedes service, while equipment purchasing is closer to observed demand. It argues that a long-lived facility can support multiple server generations. Those are management’s economic judgments and still require validation through utilization and replacement costs.^[8]

Meta likewise separates near-term capacity from future flexibility. It is maximizing available capacity for 2026 and 2027, while preparing land, power and networks for 2028 and beyond before making later server decisions. APIs, enterprise agents and direct compute sales are potential monetization routes, not established revenue streams.^[5]

Phased investment has a sensible purpose. Preparing electricity and buildings takes time, so early work can reduce future delays. Servers evolve more quickly, and buying too early can consume useful life before cash earnings begin. Poor coordination leaves equipment waiting for power, or completed facilities waiting for paying customers. Both reduce capital efficiency.

Project analysis therefore needs intermediate milestones: a workable grid connection, delivery of major equipment, cooling and network acceptance, and the start of customer billing. Building completion is only one step. Suppliers similarly face different dates for orders, production, shipment and revenue acceptance within the same project.

A larger budget can buy the same physical output

Amazon raised its expected annual cash capex from roughly $200 billion to $220 billion, citing higher memory costs, while management still expected capacity constraints in 2026 and 2027.^[8] A budget increase can contain both price and quantity effects. Treating the entire increase as additional servers or useful compute overstates physical expansion.

In a simple stress test, suppose a $10 billion project spends 30% on an equipment category whose price rises 20%, with all other costs unchanged. Total project cost rises $600 million. Maintaining the original cash return then requires higher pricing, better utilization or savings elsewhere. Both percentages are illustrative assumptions, not estimates of Amazon’s purchasing mix.

A supplier may gain revenue from higher prices without retaining the full increase as profit; a cloud operator may face a larger capital burden from the same change. Industry revenue growth does not distribute shareholder benefits evenly. The relevant questions concern cost pass-through, repricing rights and whether existing contracts cover new costs.

Efficiency changes the comparison too. Higher throughput, lower energy per completed task and better scheduling can generate more billable output from the same assets. If competition passes those savings to customers, however, the supplier retains only part of the benefit. Unit contribution and saleable output belong in the same analysis.

What the next results must demonstrate

The strongest positive evidence would be new capacity producing incremental cash: fewer delivery delays, expanding paid usage and rising operating collections without dependence on ever-larger prepayments or looser payment terms. Temporarily low free cash flow can coexist with value creation if that conversion becomes visible.

The adverse case is spending growth driven mainly by inflation, with project delays, equipment depreciation and customer demands for lower prices occurring together. Revenue might still grow while cash earned per dollar of capital falls. A high growth rate cannot settle the return question; equally, one negative cash-flow quarter does not prove the investment cycle has failed.

Supplier selection can be more specific. Check whether orders belong to projects with financing, permits and electricity arranged; whether customers bear cancellation costs; and how long collection takes after delivery. At cloud platforms, track conversion into billable capacity, contracted revenue timing and replacement spending needed to sustain existing services.

The results show substantial cloud revenue associated with growing AI demand. Future shareholder returns still depend on construction cost, timing and service pricing. Budgets measure investment intensity. Value depends on the cash that those investments can retain and the wait before it arrives.