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

In one quarter, Microsoft reported $31.9 billion of capex, $30.9 billion of cash paid for property and equipment, and $4.7 billion of finance leases. Three numbers, one spending program. Below, we reconcile them step by step, then show which one to use for which question.

Watch it play outMicrosoft's $31.9 billion of capex was not $31.9 billion of cash

In one quarter Microsoft reported $31.9 billion of capex. Circles are sized by amount.

$4.7 billion of it came through finance leases: assets now, no cash upfront. That leaves $27.2 billion bought outright.

But cash paid ran $3.7 billion higher than those outright purchases, as payment timing caught up.

Add it in and $30.9 billion of cash went out for property and equipment.

Three numbers, one program: capex for the build, cash for the hit to cash flow, leases for obligations a cash view misses.

Replay

That's the problem with comparing Big Tech's AI budgets. Microsoft, Amazon, Alphabet and Meta guide to $720 billion to $740 billion of combined 2026 capex (capital expenditure, the money spent on long-lived assets like data centers and servers). But none of them reports a line called "AI capex," and each one counts differently.

Watch it play outOne $720 billion total, four different ways of counting

Each of the four guides to a 2026 capex number. Circles are sized by guidance.

Pour them together and you get $720 billion to $740 billion of combined capex.

But none of them reports a line called "AI capex".

And each counts differently: different scope, lease treatment and calendar. Adding them shows scale, not comparable AI spending.

Replay

The short version

• Scope differs: Amazon's roughly $220 billion plan is companywide, covering fulfillment, robotics, transport and satellites as well as AWS and AI.
• Leases differ: Meta's capex explicitly includes finance-lease principal; Amazon reports cash spending net of proceeds and incentives.

Watch it play outAmazon's $220 billion plan is companywide, not an AI number

Amazon's 2026 capex plan is roughly $220 billion.

But the plan is companywide.

AWS and AI are major reasons it grew. Fulfillment centers, robotics, transportation and satellites eat capital too.

So calling the whole amount "Amazon AI spending" overstates what the company disclosed.

Replay

• Calendars differ: Microsoft's fiscal year ends in June, but its roughly $170 billion guidance is for calendar 2026.
• The fix: decide what question you're asking (physical build, cash drain, earnings hit or return) before you compare anything.

Figures are as of our August 12, 2026 research cutoff.

Check the claimThree claims to test as you read

ClaimBig Tech tells you how much it spends on AI.Tap to checkMythNone of the four files a line called AI capex. Total capex is the best public stand-in, and it covers far more than GPUs.Flip backClaimCapex is the cash a company spent that quarter.Tap to checkMythMicrosoft reported $31.9 billion of capex and paid $30.9 billion of cash in the same quarter. Leases and payment timing open the gap.Flip backClaimYou can add up the four companies' guidance.Tap to checkPartly trueAs a measure of scale, yes: $720 billion to $740 billion. As a measure of AI spending per company, no. The definitions differ.Flip back

The comparison in one table

The company guidance and source links are in the free 2026 AI capex dataset. The AI capex tracker shows each company's number and where the spending flows.

Why "AI capex" is always a stand-in

No company files a standard line item that isolates every dollar spent only on AI. One data center can run ordinary cloud workloads, advertising systems, model training, inference (running a trained model to answer requests), storage and security at the same time. The building, network and power gear around the servers can outlast several generations of chips.

One data center serves many jobs, so no line isolates AI

1. One data center (Shared building, power, network) to Cloud
2. One data center (Shared building, power, network) to Ads
3. One data center (Shared building, power, network) to AI models
4. One data center (Shared building, power, network) to Storage

So the best public proxy is total capex on the wider infrastructure build. It shows the scale without pretending every dollar bought a GPU.

This matters most for Amazon. AWS and AI are major reasons its plan grew to about $220 billion, but fulfillment centers, robotics, transportation, devices and satellites also eat capital. Calling the whole amount "Amazon AI spending" overstates what the company disclosed.

Total capex vs cash capex

An asset can show up in a company's capex before the same amount leaves as cash. Finance leases, unpaid equipment bills, construction timing, supplier incentives and asset sales all open gaps between the two.

A finance lease works like buying an asset on credit: the company gets the equipment or building now and records an obligation to pay for it over time.

Microsoft's fiscal third-quarter 2026 earnings call spells this out. The company reported:

• $31.9 billion of capex
• $4.7 billion of finance leases
• $30.9 billion of cash paid for property and equipment

Roughly two-thirds of that capex went to short-lived assets, mainly GPUs and CPUs. The rest went to longer-lived infrastructure.

Each number answers a different question. Total capex describes the physical build. Cash paid shows the immediate hit to cash flow. Finance leases reveal assets and obligations that a cash-only view misses.

Microsoft's quarter: three numbers, three questions

$31.9BTotal capexHow much it built

$30.9BCash paid for property and equipmentHow much cash left

$4.7BFinance leasesObligations a cash view misses

Meta uses yet another convention. Its second-quarter 2026 results define capex to include principal payments on finance leases. On that basis, Meta spent $31.08 billion in the quarter and narrowed full-year guidance to $130 billion to $145 billion. Its third-quarter results narrowed it again, to $135 billion to $145 billion, on the same lease-inclusive basis.

Worked example: reconciling Microsoft's three numbers

Here is the fiscal third quarter, using only the figures Microsoft gave on its call:

Find the gap

One line shows that cash paid in a quarter doesn't match the assets added in it. Tap it.

Microsoft, fiscal Q3 2026$ billions

Total capex31.9The physical build, including leased assets.Finance leases, no cash paid upfront−4.7These explain why capex can exceed cash, but here cash ended up higher.Cash paid beyond this quarter's outright purchases3.7Microsoft paid $3.7 billion more than its outright purchases, for example bills for equipment booked earlier.Cash paid for property and equipment30.9The cash hit, which doesn't line up neatly with this quarter's assets.

Tap the line you think is the tell.

Spotted it.Not that line. Step 5 is the lesson: $30.9 billion of cash against $27.2 billion of outright purchases. The call doesn't break the gap down.

Explain every lineHide the other explanations

Step 5 is the lesson. Cash paid in a quarter doesn't line up neatly with assets added in that quarter. Paying bills for equipment booked in an earlier period is one common reason; the call doesn't break the gap down.

So which number is Microsoft's "real" spending? It depends on the question:

• How much did it build? $31.9 billion, including leased assets.
• How much cash left the bank? $30.9 billion.
• What obligations were added that a cash view misses? $4.7 billion of finance leases, which have to be paid over time.

A $1 billion gap between capex and cash sounds small. Hypothetically, if a gap that size recurred every quarter at all four companies, it would add up to $16 billion a year ($1bn × 4 quarters × 4 companies).

Fiscal calendars create false comparisons

Microsoft's fiscal year ends in June, while its roughly $170 billion guidance covers calendar 2026. You can't multiply a Microsoft fiscal-quarter figure by four and set it next to a calendar-year range from Meta or Alphabet.

Microsoft's calendar 2026 guidance spans two fiscal years

1. Jan to Mar 2026Fiscal Q3 2026$31.9B of capex
2. Apr to Jun 2026Fiscal Q4 2026Fiscal year ends June 30
3. Jul to Sep 2026Fiscal Q1 2027
4. Oct to Dec 2026Fiscal Q2 2027

Label every period, keep each company's own definition, and only normalize once you know what you're asking:

One number can't do all four jobs.

Member research: The cash-pressure row is where financing enters the picture. Issue 05 follows AI infrastructure into private credit, insurance balance sheets and securitized debt, and shows how NVIDIA turns its roadmap knowledge into cheaper capital for customers.

Chips and buildings hit earnings on different clocks

A dollar spent on an accelerator doesn't last as long as a dollar spent on land, power systems or a data-center shell. Servers go into service quickly and become obsolete faster. Buildings take years to finish and then house several hardware generations.

That difference shows up in depreciation, the yearly charge that spreads an asset's cost over its useful life. Alphabet has described a mix of roughly 60% servers and 40% data centers and networking. Microsoft splits its spending into short-lived compute and assets meant to support monetization for 15 years or more. Those splits connect today's capex to tomorrow's depreciation.

To see why the split matters, take Microsoft's $31.9 billion quarter and apply its "roughly two-thirds" short-lived share. The useful lives below are a hypothetical we chose for round numbers, not Microsoft's accounting policy.

Make a guess

Short-lived compute is about two-thirds of Microsoft's $31.9 billion quarter. Depreciate compute over 5 years and sites over 20. What share of the yearly charge comes from compute?

Tap your guess

About two-thirdsYour guess72% to 85%Your guess85% or moreYour guess

01234567890123456789%Compute's share of yearly depreciation

Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was 90%.

Close to 90%: about $4.3 billion of a $4.8 billion yearly charge, because compute wears out four times faster than the buildings.

Compute's share of yearly depreciation90%

Compute's share of spending67%

The compute is two-thirds of the spending but close to 90% of the yearly charge (about $4.3bn ÷ $4.8bn). A company whose mix tilts toward servers will see depreciation climb faster, even with the same capex headline.

Our AI depreciation guide shows why cash leaves long before the full expense reaches the income statement. If you compare current margins with current capex while ignoring assets not yet in service, you're comparing two different moments in the cycle.

A seven-step check for each earnings report

1. Record the company's exact capex label and reporting period.
2. Separate cash purchases, finance leases and other non-cash additions.
3. Note whether the guidance is AI-specific, infrastructure-specific or companywide.
4. Split short-lived compute from long-lived facilities when management gives the mix.
5. Track construction in progress, the assets bought but not yet in use.
6. Compare new depreciation with capacity, revenue and gross margin.
7. Reconcile the spending to free cash flow (operating cash minus capex), without assuming every dollar is optional.

The output is a range with its caveats attached, not a false point estimate.

Try it yourself

Ten minutes, one company, three definitions:

1. Open Microsoft's fiscal third-quarter call and find the three numbers from the worked example. Confirm they match.
2. Open the company you follow in the free 2026 AI capex dataset and note which capex measure its guidance uses.
3. Check the matching company page, such as Meta's, to see which measure its quarterly spend rows use.
4. From the company's latest cash-flow statement, take revenue, operating cash flow and cash paid for property and equipment. Enter them into the free capex-to-free-cash-flow coverage calculator.
5. Run it a second time with total capex (including finance leases) in place of cash paid, and note how much free cash flow changes.

You should now be able to answer: which of this company's "capex" numbers would you use to judge cash pressure, and which to judge the size of the physical build?

Our read

Capex proves money is being committed. It doesn't prove demand will show up, that a site will get power on time, or that new capacity will earn a good return. Depreciation shows assets being used up, but not how busy they are. Free cash flow shows the immediate burden, though payment timing can distort it.

You only get an answer when the accounting meets 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 spending. Meta earns much of its return through ad performance rather than a separately reported cloud business.

We'd track three lines for each company: how much physical capacity was added, how much cash left, and what operating result arrived.

• The case strengthens when all three improve together.
• The case needs a harder look when spending and depreciation speed up while utilization, margins or free cash flow weaken.

The combined total is a good map of the infrastructure cycle. It's a dangerous shortcut for valuing any one company, the same way a price-to-earnings ratio is meaningless if two companies define earnings differently.

Questions readers ask

Is a finance lease basically debt?

Economically, it's close. The company gets the asset now and records an obligation to pay for it over time, much like borrowing to buy it. That's why finance leases matter for leverage as well as for capex, and why a cash-only view can understate what a company has committed.

How is an operating lease different?

An operating lease works more like rent. The company pays for use of the asset over the lease term, and those payments don't count as capex. Two companies with identical data centers can report different capex if one uses finance leases and the other operating leases.

Why can't I just add up the four companies' guidance?

You can, as a measure of scale, and the sum is $720 billion to $740 billion. But it mixes a June fiscal-year company's calendar guidance, cash-basis and lease-inclusive definitions, and Amazon's companywide plan. It tells you how big the build is, not how much each company spends on AI.

Which capex number should I use?

Match the number to the question. Use total capex including leases for the physical build, cash paid for property and equipment for cash pressure, and depreciation plus assets placed in service for the earnings hit.

Go deeper with member research

Once you can read the accounting, the next question is what it does to margins and cash. That's where member research picks up:

• Issue 03: The AI bill is arriving late to earnings: Hyperscalers put almost four dollars into first-quarter infrastructure for every dollar of depreciation. The report shows where the margin test lands next.
• Issue 02: The AI Capex Shock: Where the next AI infrastructure dollar goes, which bottlenecks keep pricing power, how the boom is financed and what could break first.

Limits of this data

• Company definitions and guidance can change after our August 12 cutoff.
• No company discloses a standalone audited AI capex figure, so every comparison here uses proxies.