AI Capex Accounting: Compare Microsoft, Amazon, Google & Meta

Direct answer: Microsoft, Amazon, Alphabet, and Meta do not report one comparable measure called “AI capex.” Their 2026 guidance adds to $695 billion to $720 billion, but the total combines different reporting periods, lease treatments, cash-flow definitions, and non-AI investments. A useful comparison must preserve those differences instead of treating the midpoint as a single audited number.

That distinction changes the investment conclusion. Two companies can announce similar capital budgets while creating very different near-term cash costs, depreciation schedules, and paths to revenue.

The accounting comparison in one table

Company	2026 guidance	Reporting scope	Lease treatment	Main comparability problem
Microsoft	About $190bn	Calendar-year company capex	Total capex can include finance-lease additions	Calendar guidance does not align with its June fiscal year
Amazon	About $200bn	Companywide capex	Cash-flow reporting is net of proceeds and incentives	Includes AWS, fulfillment, robotics, transport, and satellites
Alphabet	$175bn–$185bn	Company property and equipment	Quarterly cash capex can move with payment timing	Does not disclose a standalone audited AI total
Meta	$130bn–$145bn	Company capex	Explicitly includes finance-lease principal	Much of the return arrives indirectly through advertising

The underlying company guidance and source links are available in the 2026 AI capex dataset. The broader AI capex tracker explains where the spending flows.

Why “AI capex” is usually a proxy

None of the four companies files a standardized line item that isolates every dollar spent only for artificial intelligence. A data center can serve conventional cloud workloads, internal products, advertising systems, model training, inference, storage, and security at the same time. The network, building, and power equipment around the servers can support several generations of compute.

The most defensible public proxy is therefore capital expenditure associated with the wider infrastructure build. It captures the scale of investment without pretending that every dollar bought a GPU.

This matters most for Amazon. Its approximately $200 billion plan is companywide. AWS and AI are major reasons for the increase, but fulfillment centers, robotics, transportation, devices, and satellite infrastructure also consume capital. Calling the full amount “Amazon AI spending” would overstate what the company disclosed.

Total capex and cash capex answer different questions

An asset can enter a company’s capital-expenditure measure before the same amount appears as cash paid for property and equipment. Finance leases, unpaid equipment, construction timing, supplier incentives, and asset sales all create differences.

Microsoft’s fiscal third-quarter 2026 earnings call is unusually explicit. The company reported $31.9 billion of capex, $4.7 billion of finance leases, and $30.9 billion of cash paid for property and equipment. Roughly two-thirds of capex went to short-lived assets, primarily GPUs and CPUs, with the rest directed to longer-lived infrastructure.

Those numbers are related, but they are not interchangeable. Total capex helps describe the physical investment cohort. Cash paid helps explain the immediate cash-flow burden. Finance leases show assets and obligations that an analysis based only on cash purchases can miss.

Meta uses another convention. Its second-quarter 2026 results define capex to include principal payments on finance leases. The company reported $31.08 billion of quarterly capex on that basis and narrowed full-year guidance to $130 billion to $145 billion.

Fiscal calendars can create false comparisons

Microsoft reports on a fiscal year ending in June, while its roughly $190 billion guidance is framed for calendar 2026. A quarterly figure from Microsoft’s fiscal third quarter cannot simply be multiplied by four and compared with a calendar-year range from Meta or Alphabet.

The safe approach is to label every period, preserve the company’s own definition, and normalize only after deciding the question being asked.

For physical investment, compare additions including financed assets where available.
For cash pressure, compare property-and-equipment cash payments using each filing’s reconciliation.
For earnings pressure, compare the assets placed in service with depreciation and operating costs.
For return on investment, connect new capacity with revenue, utilization, gross margin, and free cash flow.

One number cannot perform all four jobs.

Short-lived compute and long-lived sites hit earnings differently

A dollar spent on an accelerator does not have the same economic life as a dollar spent on land, power distribution, or a data-center shell. Servers and accelerators can enter service quickly and become obsolete faster. Buildings can take years to complete and then support several hardware generations.

Alphabet has described a mix of roughly 60% servers and 40% data centers and networking equipment. Microsoft separately describes short-lived compute and assets intended to support monetization for 15 years or more. These disclosures provide a bridge between today’s capex and tomorrow’s depreciation.

The AI depreciation guide shows why cash can leave long before the full expense reaches the income statement. An investor who compares current margins with current capex without tracking assets not yet in service is comparing different moments in the cycle.

A normalization workflow that does not invent precision

Use this sequence when a new earnings report arrives:

Record the company’s exact capex label and reporting period.
Separate cash purchases, finance leases, and other non-cash additions.
Identify whether the guidance is AI-specific, infrastructure-specific, or companywide.
Split short-lived compute from long-lived facilities when management provides the mix.
Track construction in progress or assets not yet in service.
Compare new depreciation with capacity, revenue, and gross-margin signals.
Reconcile the investment to free cash flow without assuming every dollar is discretionary.

This method produces a range and a set of caveats, not a false point estimate.

What the accounting can and cannot prove

Capex proves that capital is being committed. It does not prove that demand will arrive, that a facility will be energized on time, or that the newest capacity will earn an attractive return.

Depreciation proves that assets are being consumed through the income statement. It does not reveal utilization by itself. Free cash flow shows the immediate burden after capital investment, but it can be distorted by payment timing and working capital.

The investment question is resolved only when these accounting measures are connected to operating evidence. Microsoft points to capacity constraints and Azure demand. Amazon’s second-quarter update shows both rapid AWS growth and a trailing free-cash-flow outflow driven by property investment. Meta monetizes much of the infrastructure through advertising performance rather than a separately reported cloud unit.

The decision takeaway

The combined capex total is useful as a map of the infrastructure cycle. It is dangerous as a valuation shortcut.

Treat every headline number as the start of the work. The durable comparison asks three separate questions: how much physical capacity was added, how much cash left, and what operating result arrived. When those three lines improve together, the investment is beginning to validate itself. When spending and depreciation accelerate while utilization, margins, or free cash flow weaken, the thesis needs a harder review.

Yield Theory members get that review as an explicit investment call: what the market is pricing, which evidence confirms the thesis, which catalysts matter next, and what would break the position.

Research cutoff: August 12, 2026. This article is educational and is not personalized investment advice. Company definitions and forward-looking guidance can change; verify current figures against the linked primary sources.