Five large technology companies entered earnings season with a similar message: demand for AI infrastructure still exceeds available capacity. Their cash-flow outcomes looked far less alike.
Microsoft generated USD 19.6 billion of quarterly free cash flow while continuing to expand AI capacity. Alphabet reported negative quarterly free cash flow, Amazon's trailing-12-month free cash flow turned negative, Meta's quarterly free cash flow narrowed to less than USD 1 billion, and Oracle outlined another year of capital-markets funding.
The share-price reactions also diverged. Based on daily closes around their respective results, Microsoft rose sharply in the following session, while Alphabet, Meta and Oracle fell. Those moves cannot be attributed to capital expenditure alone, but they show that strong AI demand no longer produces one uniform market verdict.
Earnings season has turned the AI investment cycle into a cash-flow audit. Revenue growth still matters, but investors are also asking how long capital remains tied up, when contracted demand becomes cash, and who finances the interval.
Demand no longer settles the debate
The demand evidence across the five calls was unusually strong. Azure and other cloud services grew 43%, AWS grew 36.7%, Google Cloud grew 82%, and Oracle's cloud infrastructure revenue grew 93%. Meta's total revenue rose 28% as it continued to expand AI infrastructure for its advertising and product platforms.
Contracted demand was also substantial. Microsoft reported USD 678 billion of commercial remaining performance obligations, Amazon said its backlog reached USD 496 billion, Alphabet reported USD 514 billion of Cloud backlog, and Oracle ended its fiscal year with USD 638 billion of remaining performance obligations.
Backlog scale versus conversion visibility
USD billions; company-defined backlog and RPO
Company-defined backlog and RPO balances are not perfectly comparable. Conversion schedules are shown only where management disclosed them in the reviewed calls. Meta is excluded because it does not currently operate a comparable third-party cloud business or report a comparable cloud backlog/RPO measure.
Microsoft
678 USD bn- Reported balance
- Commercial RPO
- Conversion visibility
- Timing not disclosed
Oracle
638 USD bn- Reported balance
- RPO
- Conversion visibility
- 12% · next 12 months; 34% · months 13–36
Alphabet
514 USD bn- Reported balance
- Google Cloud backlog
- Conversion visibility
- >50% · within 24 months
Amazon
496 USD bn- Reported balance
- AWS backlog
- Conversion visibility
- Timing not disclosed
The spending plans moved in the same direction. Alphabet raised its 2026 CapEx guidance to USD 195–205 billion. Amazon raised expected 2026 cash CapEx to approximately USD 220 billion, citing higher memory costs, while Meta narrowed its guidance upward to USD 130–145 billion. Oracle expects roughly USD 70 billion of net cash outlay for CapEx in fiscal 2027, and Microsoft said its calendar-2026 CapEx expectation was approximately USD 175 billion after a lease-accounting reclassification.
Company-stated CapEx guidance from the 2026 earnings season
USD billions; company-stated periods and definitions
Company-stated guidance; periods and definitions differ. Oracle is shown as a separate FY27 net-cash-outlay measure so that all five companies remain visible without implying direct comparability.
- Guidance range
- Approximate value
- FY27 net cash outlay (not directly comparable)
Meta
130–145FY2026 · CapEx guidance including finance-lease principal
Microsoft
~175Calendar 2026 · CapEx expectation after lease useful-life reclassification
Alphabet
195–205FY2026 · CapEx guidance
Amazon
~220Calendar 2026 · Cash CapEx guidance
Oracle
~70FY2027 · Net cash outlay for CapEx
Reported CapEx will be USD 20–25bn higher because the net measure excludes customer prepayments and timing effects.
These figures weaken the simple demand-slowdown argument. They do not settle the return question, because backlog is not current revenue and revenue is not free cash flow. The missing variable is time.
The cash residual is already diverging
Microsoft's fiscal fourth quarter provides the strongest current cash-conversion result in the group. It generated USD 55.4 billion of operating cash flow and USD 19.6 billion of free cash flow after USD 41.0 billion of CapEx including finance leases. Free cash flow still fell 23% from a year earlier, but remained clearly positive.
Alphabet's second quarter shows the opposite side of the same expansion. Operating cash flow reached USD 39.1 billion, while CapEx of USD 44.9 billion pushed quarterly free cash flow to negative USD 5.9 billion. At the same time, Google Cloud revenue reached USD 24.8 billion and its operating margin rose to 35.6%. The cloud business was improving even as infrastructure spending absorbed more cash at the group level.
Meta reported USD 31.1 billion of CapEx including finance-lease principal and only USD 784 million of free cash flow for the quarter. Amazon disclosed USD 53.1 billion of Q2 cash CapEx, primarily for AWS and generative AI, while its trailing-12-month free cash flow was negative USD 7.6 billion. Oracle generated USD 32 billion of operating cash flow in fiscal 2026 against USD 48 billion of net cash outlay for CapEx under its own non-GAAP definition.
Five-company cash-funding scorecard
USD billions; reported periods and definitions differ
Reported periods and CapEx definitions differ. This is a cash-funding scorecard, not a like-for-like profitability ranking.
- Latest investment measure
- Reported cash outcome (FCF / OCF)
Microsoft
- Latest investment measure
- 41.0 USD bnCapEx including finance leasesFY26 Q4
- Reported cash outcome (FCF / OCF)
- 19.6 USD bnFree cash flowFY26 Q4
- Demand signal
- +43%Azure and other cloud services growth
- Contracted demand
- 678 USD bnCommercial RPO
- Essential caveat
- FCF fell 23% year over year; about two-thirds of CapEx went to short-lived assets.
Alphabet
- Latest investment measure
- 44.9 USD bnCapExQ2 2026
- Reported cash outcome (FCF / OCF)
- −5.9 USD bnFree cash flowQ2 2026
- Demand signal
- +82%Google Cloud revenue growth
- Contracted demand
- 514 USD bnGoogle Cloud backlog
- Essential caveat
- Trailing-12-month FCF remained positive at USD 53.3bn.
Amazon
- Latest investment measure
- 53.1 USD bnCash CapExQ2 2026
- Reported cash outcome (FCF / OCF)
- −7.6 USD bnFree cash flowTrailing 12 months to Q2 2026
- Demand signal
- +36.7%AWS revenue growth
- Contracted demand
- 496 USD bnAWS backlog
- Essential caveat
- Periods differ; do not calculate a ratio from these figures.
Meta
- Latest investment measure
- 31.1 USD bnCapEx including finance-lease principalQ2 2026
- Reported cash outcome (FCF / OCF)
- 0.784 USD bnFree cash flowQ2 2026
- Demand signal
- +28%Total company revenue growth
- Contracted demand
- Not disclosed on a comparable basisComparable backlog
- Essential caveat
- Demand growth is company-wide, not a cloud-segment metric.
Oracle
- Latest investment measure
- 48.0 USD bnNet cash outlay for CapExFY2026
- Reported cash outcome (FCF / OCF)
- 32.0 USD bnOperating cash flow (OCF)FY2026
- Demand signal
- +93%Cloud infrastructure revenue growth
- Contracted demand
- 638 USD bnRPO
- Essential caveat
- Net cash outlay is Oracle's non-GAAP funding measure; OCF is not FCF.
These numbers are not a clean ranking. The reporting periods differ, and so do the CapEx definitions. Oracle's net cash outlay deducts certain customer prepayments and short-term financing effects, while Microsoft and Meta include specified finance-lease amounts in their headline CapEx figures.
The comparison is still useful because it reveals the cash residual each management team must defend. A company with positive free cash flow can fund more of the cycle internally. A company with a funding gap needs stronger evidence on contract duration, utilisation, margins and access to external capital.
Investment duration is the core variable
Amazon gave the clearest explanation of the timing mismatch. Management separated data-centre investment from the servers and networking equipment installed inside it. Data-centre capital can be spent about two years before monetisation begins, but the resulting facility may operate for more than 30 years and host several generations of equipment.
Servers follow a shorter cycle. Amazon said they are generally purchased a few months before deployment, take a little less than three years to break even on average, and currently have useful lives of at least five to six years. Most of its AI capacity is being contracted for at least five-year terms. If those assumptions hold, the contract can outlast the break-even period and leave two to three years of cash generation before replacement. They remain forward estimates rather than realised returns.
The other calls supplied different parts of the same duration test. Alphabet said approximately 60% of its Q2 technical-infrastructure investment went to servers and 40% to data centres and networking equipment. It expects just over half of its USD 514 billion Cloud backlog to be recognised as revenue over the next 24 months, while temporary use of third-party capacity will put modest pressure on near-term Cloud margins.
Microsoft said roughly two-thirds of its quarterly CapEx went to short-lived assets, mainly CPUs and GPUs. That mix can be deployed against visible demand more quickly, but it also increases the importance of utilisation, useful-life assumptions and replacement economics. A large backlog helps, yet it cannot remove technology and execution risk.
The relevant question is whether contracted demand can convert into cash before depreciation, replacement costs and financing requirements absorb the expected return.
Funding structure changes who bears the wait
Oracle makes the funding question explicit. For fiscal 2027, management expects about USD 70 billion of net cash outlay for CapEx and said reported CapEx will be roughly USD 20–25 billion higher because the net figure excludes customer prepayments and timing effects. It expects to raise around USD 40 billion of debt and equity, including an announced USD 20 billion at-the-market equity programme.
Customer financing is becoming part of the infrastructure product. Oracle said it signed USD 67 billion of AI infrastructure contracts in the quarter, mostly using bring-your-own-hardware or prepaid structures, taking the cumulative value of those arrangements to USD 75 billion. These structures reduce Oracle's own cash requirement by moving part of the funding burden to the customer. Management said they carry similar or better margins than its other contracts.
That does not make the capital free. Oracle expects 12% of its remaining performance obligations to be recognised over the next 12 months and another 34% in months 13 to 36. Its estimate of project-level steady-state return on invested capital in the high 20s applies only after large projects have ramped. The 97.5% reported global GPU utilisation rate is encouraging operating evidence, but the return still depends on delivery, customer quality and sustained usage.
Meta is also broadening the funding mix through debt, external capital and infrastructure partnerships. Amazon said it has issued debt and retains several funding options, without setting out a new financing plan on the call. Microsoft, by contrast, still expects to remain free-cash-flow positive.
The result is a transfer question as much as a spending question. When customers prepay, partners own part of the infrastructure, or shareholders supply new capital, the waiting period does not disappear. It moves to a different balance sheet.
The three tests investors should apply
The first test is conversion, not backlog size. Contracted demand is more valuable when revenue recognition is near, customers are diversified, capacity is ready and pricing protects returns against component inflation. Backlog that requires years of construction and additional financing deserves a different risk adjustment from capacity already in service.
The second test is duration matching. Investors should compare contract length with construction lead time, equipment break-even, useful life and replacement cadence. A 30-year data-centre shell and a five-year accelerator do not carry the same economic risk even when both appear in the same CapEx number.
The third test is who funds the gap. Internal operating cash flow offers the greatest flexibility. Debt, equity, customer prepayments and infrastructure partnerships can all support a viable project, but each changes the distribution of risk and future cash flows. The funding mix should therefore be read alongside revenue growth, not after it.
What the cash-flow test changes
Negative free cash flow does not by itself prove that AI investment is uneconomic. In a build phase, cash leaves before contracted capacity enters service, and the eventual return may be attractive. Positive cloud growth does not prove the opposite either. Returns can still be eroded by slower delivery, lower utilisation, component inflation, refinancing needs or faster equipment replacement.
The next phase of the AI trade will be judged less by the absolute size of each CapEx announcement than by changes in backlog conversion, utilisation, cloud margins, depreciation and funding needs. The five calls show that management teams now understand the burden of proof: each offered a more detailed account of duration, returns or financing than a simple growth narrative would require.
That is the practical shift for investors. AI CapEx is no longer receiving permission for the entire cycle at once. It has to earn that permission again each quarter through cash conversion and credible financing.
