AI Spent $132 Billion. Earnings Show Only Part of the Bill.

Microsoft, Alphabet, Amazon, and Meta invested a combined $131.6 billion in first-quarter 2026 infrastructure proxies. The same companies recorded about $34.0 billion of property depreciation or depreciation and amortization.

The definitions are not identical, so the comparison is not an accounting identity. It is still a useful warning. The cash committed to the current infrastructure build is running at roughly 3.9 times the expense moving through earnings.

That does not mean the other $97.6 billion vanished. It means most of the economic test has not reached the income statement yet.

The distinction matters because investors often use today's operating margin to judge a fleet that is still being ordered, constructed, connected to power, and placed in service. Strong current profit can coexist with a large future cost pipeline. The right conclusion is not that reported earnings are false. It is that cash flow and earnings are showing different moments in the same investment cycle.

Cash leaves before depreciation begins

Capital expenditure and depreciation answer different questions.

Capital expenditure records the cash and financing used to acquire long-lived assets. For an AI data center, that can include land, buildings, electrical systems, cooling, servers, accelerators, networking equipment, and assets obtained through finance leases.

Depreciation allocates the cost of an asset across the years management expects to use it. It normally begins when the asset is ready for its intended use, not when the company first orders it or starts building the site.

A simplified data-center investment can move through six stages:

1. The company signs equipment, construction, lease, and power commitments.
2. Cash leaves as suppliers are paid and construction advances.
3. Unfinished assets remain in construction in progress or another not-yet-in-service category.
4. The data center becomes available for use.
5. Depreciation begins over the estimated useful lives of its buildings and equipment.
6. Revenue and operating savings determine whether the asset earns an acceptable return.

The first three stages can consume enormous amounts of capital without producing a matching depreciation charge. The fourth stage starts the earnings clock. The sixth determines whether the spending created value.

This lag is not an accounting trick. It is how long-lived assets are normally treated. The analytical mistake is comparing current margins with current spending as if every dollar of new infrastructure were already operating and fully expensed.

A simple example

Assume a company spends $12 billion on a data-center campus.

Suppose $4 billion funds buildings and site infrastructure expected to last 20 years. The remaining $8 billion funds servers, accelerators, storage, and networking equipment expected to last five years. Ignore residual values and assume straight-line depreciation for simplicity.

Once the whole campus is operating, annual depreciation would be approximately:

• $200 million for the buildings and site infrastructure
• $1.6 billion for the computing and networking equipment
• $1.8 billion in total

The company spent $12 billion, but the first full year of depreciation is only $1.8 billion. If the campus becomes available halfway through the year, the first calendar-year charge may be closer to $900 million.

Cash flow saw nearly the entire construction bill. Earnings saw only the portion assigned to assets already in service and only for the months they were operating.

Now repeat the exercise every year. The 2025 fleet is still depreciating when the 2026 fleet enters service. A 2027 purchase adds another layer. The expense becomes material when several investment cohorts overlap, even if annual capital expenditure stops accelerating.

That is why a capex boom can pressure free cash flow immediately while reported margin deteriorates later.

What the first-quarter numbers actually say

The four-company comparison uses the closest disclosed infrastructure measures available for the quarter ended March 31, 2026.

Company	Q1 2026 investment proxy	Q1 property depreciation or D&A	Ratio
Microsoft	$31.9bn of capex	$9.0bn	3.5x
Alphabet	$35.7bn of cash capex	$6.5bn	5.5x
Amazon	$44.2bn of cash property purchases	$12.8bn	3.4x
Meta	$19.84bn of capex including finance-lease principal	$5.68bn	3.5x
Combined	$131.6bn	$34.0bn	3.9x

Microsoft said capital expenditure was $31.9 billion in its fiscal third-quarter earnings call. Roughly two-thirds went to short-lived assets, primarily GPUs and CPUs. The rest went to assets expected to support monetization over 15 years or more. That split matters because a dollar spent on a server reaches depreciation much faster than a dollar spent on a building.

Alphabet's first-quarter Form 10-Q reported $35.674 billion of property purchases and $6.482 billion of property depreciation. It also reported $108.597 billion of assets not yet in service, up from $78.592 billion three months earlier. Those assets are a visible bridge between cash already committed and depreciation that can arrive later.

Amazon's first-quarter Form 10-Q reported $44.203 billion of cash property purchases. Its segment note reported $12.834 billion of property depreciation and amortization. The same filing showed $54.757 billion of net property additions because non-cash activity and unpaid equipment create another difference between physical investment and the cash-flow statement.

Meta reported $19.84 billion of capital expenditure, including finance-lease principal, in its first-quarter results. Its Form 10-Q reported $5.68 billion of property depreciation.

The ratios should not be used to forecast future depreciation by multiplying the latest quarter. The asset mix, payment timing, leases, replacement spending, and useful lives differ. The comparison answers a narrower question: are these companies funding the physical asset base faster than the current income statement is consuming it?

For all four, the answer is yes.

The gap does not prove an AI bubble

A high capex-to-depreciation ratio is not automatically bearish.

Depreciation is not a cash payment in the period it is recognized. It represents cash spent earlier. If a new data center is full, pricing is rational, and revenue grows faster than the cost of operating the site, a large depreciation charge can coexist with excellent economics.

Supply constraints can also make advance spending rational. Microsoft said it expected to remain capacity constrained through at least 2026. A company that already has contracted demand may destroy value by building too slowly.

The platforms also have more than one route to monetization.

Microsoft and Amazon can sell capacity to external cloud customers. Alphabet can earn revenue through Google Cloud while using the same infrastructure to improve Search, advertising, and its own models. Meta primarily monetizes compute indirectly through better recommendations, engagement, and ad conversion.

The bull case is straightforward: demand keeps the new fleet highly utilized, custom chips and software reduce the cost of each unit of work, and gross profit grows faster than depreciation.

If that happens, the capex surge will look like the front-loaded cost of a productive new platform.

The real risk is idle depreciation

Depreciation alone does not destroy value. Underutilized depreciation does.

An accelerator begins aging whether it is serving a paid workload or waiting for one. A data-center shell still carries maintenance, power, staffing, and financing costs before utilization reaches an efficient level. The economic clock does not pause because customer demand arrived late.

Three gaps can create trouble.

Capacity can arrive before demand

Cloud providers build ahead because campuses take years to plan, permit, power, and complete. That lead time is unavoidable. It also means demand forecasts must be made before the revenue is visible.

If enterprise adoption slows or large customers reduce commitments, newly opened capacity can depress margin before management can cancel the next construction phase.

Hardware can age before it pays back

AI accelerators improve quickly. A server may remain physically useful for years while becoming less competitive for the most valuable workloads. Older equipment can often be moved to inference, internal services, or less demanding tasks, but that reuse must be real.

The risk is not that every old chip becomes worthless overnight. It is that price and performance fall faster than the accounting useful life assumes.

Price competition can absorb efficiency gains

Better chips, model compression, and software optimization can reduce the cost of generating a useful result. Those savings do not automatically reach shareholders. Competition may pass them to customers through lower prices.

Revenue can grow while the return on each new dollar of infrastructure falls. That is why unit economics and gross margin matter more than headline AI usage.

Useful lives are an earnings assumption, not a physical law

Management chooses estimated useful lives based on expected economic use. Extending a server's life reduces annual depreciation because the same historical cost is spread over more years.

The change can be reasonable. Software optimization can keep older hardware productive. A diverse workload stack can assign cutting-edge accelerators to frontier models and move mature workloads to older machines.

But useful-life changes do not recover cash. They change when the cost appears in earnings.

Investors should read the accounting-policy note whenever a large infrastructure buyer changes the expected life of servers or networking equipment. Then compare that assumption with:

• how quickly the company is replacing its newest fleet
• whether older hardware is still used at attractive utilization
• whether assets held for sale or impairments are rising
• whether gross margin improves after adjusting for the lower depreciation charge

An extension supported by longer productive use is an economic gain. An extension paired with aggressive replacement can make near-term margin look stronger without improving the underlying return.

Five numbers belong in the same model

No single accounting line can answer whether AI infrastructure is earning its cost. Investors need to connect five.

1. Capital expenditure

Capex shows the speed and scale of the build. Separate cash purchases, finance leases, and companywide investment where the disclosure permits it.

2. Assets not yet in service

Construction in progress and similar categories show how much funded infrastructure has not started depreciating. A growing balance can be constructive if deployment and demand are advancing together. It is concerning when projects remain unfinished or unpowered for longer than planned.

3. Depreciation

Track both the absolute charge and its growth rate. Read the useful-life policy. Distinguish property depreciation from broader depreciation and amortization measures that may include acquired intangibles, content, or lease assets.

4. Gross margin

Gross margin reveals whether pricing, utilization, and efficiency are absorbing the cost of the fleet. Consolidated margin can hide business-level differences, so cloud or segment margin is more useful when available.

5. Free cash flow

Free cash flow shows the immediate capital burden. A company can report rising operating income while free cash flow falls because capex arrives before depreciation.

The strongest result is not merely faster AI revenue. It is revenue growth accompanied by stable or improving gross margin, a manageable depreciation ramp, and free-cash-flow recovery after the build matures.

What to listen for on the next earnings calls

Management will rarely provide a perfect return-on-invested-capital calculation for AI. The clues are still visible.

Listen for whether capacity remains constrained, how quickly new sites are energized, and whether customer commitments are firm or merely expected demand. Compare cloud growth with capex growth. Watch whether management discusses gross-margin pressure from depreciation, energy, and component prices.

Then look one level below the headline:

• Is construction in progress rising faster than assets placed in service?
• Is depreciation accelerating after useful-life changes?
• Are finance leases and purchase commitments growing after reported capex slows?
• Is free cash flow recovering because operating cash flow improved, or only because investment was delayed?
• Is management describing lower cost per unit while segment margin still deteriorates?

The most important signal is the return on the newest capacity, not the average return on infrastructure installed years ago.

What this means for the AI investment debate

The first stage of the AI cycle rewarded access to scarce compute. The next stage will distinguish access from productive use.

Today's income statements already include substantial AI infrastructure costs, but they do not include the full economic burden of the assets still being constructed and commissioned. That creates a delayed test. Depreciation from several investment cohorts can overlap just as competitive pricing and hardware replacement become more demanding.

The optimistic outcome does not require capex to fall immediately. It requires revenue, utilization, and efficiency to catch the cost as the fleet enters service.

The pessimistic outcome does not require AI demand to disappear. It only requires the marginal return on new capacity to fall below the cost implied by today's spending.

The $132 billion first-quarter figure is therefore neither a victory nor a verdict. It is a funded claim on future gross profit.

Yield Theory's member issue performs the full company-by-company audit, including the assets waiting to enter service, commitment stacks, useful-life assumptions, and the specific signals that would confirm or break the thesis.

Read Issue 03: The AI bill is arriving late to earnings →

---

Research cutoff: July 23, 2026. This article is for educational purposes only and is not personalized investment advice. Company disclosures use different definitions, and forward-looking guidance can change.