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

In the first quarter of 2026, Microsoft, Alphabet, Amazon and Meta put a combined $131.6 billion into infrastructure. The same quarter's earnings carried only about $34.0 billion of depreciation, the charge that spreads an asset's cost over its working life.

Cash is going out at roughly 3.9 times the rate the cost shows up in profits. The other $97.6 billion didn't vanish; most of the AI bill just hasn't reached the income statement yet. Below, we show how the lag works, how it stacks up year after year, and the one balance-sheet line that tells you how much more is on the way.

Watch it play outFirst-quarter investment ran $97.6 billion ahead of depreciation

In Q1 2026, Microsoft, Alphabet, Amazon and Meta put $131.6 billion into infrastructure.

Only a slice reached that quarter's profits as depreciation: about $34.0 billion.

Lift it out. The other $97.6 billion didn't vanish; it just hasn't reached the income statement yet.

Cash went out roughly 3.9 times faster than the cost showed up in profits. Most of the AI bill is still on its way.

Replay

The short version

• Cash first, earnings later: depreciation starts only when a data center or server goes into service, so spending hits cash flow well before it hits profit.
• The gap is big everywhere: each company's Q1 spending ran 3.4 to 5.5 times its depreciation. Alphabet alone had $108.597 billion of assets not yet in service.

Watch it play outAlphabet alone had $108.597 billion of assets not yet switched on

Alphabet's depreciation charge for the first quarter: $6.5 billion.

Pull back. Three months earlier, Alphabet had $78.592 billion of assets not yet in service.

By the end of the quarter the pile had jumped to $108.597 billion, dwarfing that depreciation charge.

None of it depreciates until it switches on. Then it starts feeding the earnings charge.

Replay

• The real risk is idle hardware: chips age whether or not paying customers show up.

Figures are as of our July 23, 2026 research cutoff. For each company's full-year plan, see the hyperscaler capex tracker.

Each company spent 3.4 to 5.5 times its depreciation

Q1 2026 investment proxyQ1 depreciation

10203040$50B0

$31.9B$9.0B

MicrosoftQ1 2026 investment proxy$31.9BQ1 depreciation$9.0B

$35.7B$6.5B

AlphabetQ1 2026 investment proxy$35.7BQ1 depreciation$6.5B

$44.2B$12.8B

AmazonQ1 2026 investment proxy$44.2BQ1 depreciation$12.8B

$19.8B$5.7B

MetaQ1 2026 investment proxy$19.8BQ1 depreciation$5.7B

MicrosoftAlphabetAmazonMeta

Why cash leaves before depreciation starts

Capex, short for capital expenditure, is the cash and financing used to buy long-lived assets. For an AI data center, that means land, buildings, power and cooling systems, servers, accelerators, networking gear and assets obtained through finance leases (leases that work like buying on credit).

Depreciation normally starts when the asset is ready to use, not when the company orders it or breaks ground.

A data-center project moves through six stages:

1. The company signs equipment, construction, lease and power commitments.
2. Cash goes out as suppliers are paid.
3. Unfinished assets sit in "construction in progress" or a similar not-yet-in-service bucket.
4. The data center becomes available for use.
5. Depreciation begins.
6. Revenue and savings decide whether the asset earns an acceptable return.

Stages one to three can eat huge amounts of capital with no matching depreciation. Stage four starts the earnings clock; stage six decides whether the spending created value.

The AI bill runs on four clocks

1. Commit (Orders, leases, power and land) to Build (Construction in progress does not depreciate)
2. Build (Construction in progress does not depreciate) to Switch on (Placed in service; depreciation begins)
3. Switch on (Placed in service; depreciation begins) to Monetize (Revenue and savings decide the return)

This lag is how long-lived assets are normally treated. The mistake is judging today's margins against today's spending as if every new dollar were already running.

Check the claimThree claims to test as you read

ClaimThe 3.9x gap means profits are overstatedTap to checkMythNo. Depreciation starts when an asset goes into service, so today's earnings correctly exclude assets still being built. The risk is misreading what comes next.Flip backClaimDepreciation is a cash costTap to checkMythNot in the year it is booked. It reflects cash spent earlier, which is why free cash flow falls first and earnings follow later.Flip backClaimA longer useful life makes a company better offTap to checkPartly trueOnly if the hardware really lasts longer. The change lowers yearly depreciation but brings back no cash.Flip back

A simple example

Say a company spends $12 billion on a data-center campus: $4 billion on buildings expected to last 20 years, and $8 billion on servers, accelerators, storage and networking expected to last five. Assume straight-line depreciation and no leftover value.

Once the whole campus runs, yearly depreciation is about:

Cash flow saw almost the full $12 billion. Earnings see $1.8 billion in the first full year, and if the campus opens halfway through the year, closer to $900 million.

Now repeat that every year. The 2025 fleet is still depreciating when the 2026 fleet switches on, and a 2027 purchase adds another layer, so the expense grows even if annual capex stops growing.

Make a guess

Spending stays flat at $12 billion a year, one campus at a time. By year five, what is the yearly depreciation charge?

Tap your guess

$0–3BYour guess$3–6BYour guess$6–10BYour guess$10B or moreYour guess

$0123456789.0123456789BYear five

Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was $9.0B.

Five campuses are depreciating at $1.8B each, so the charge reaches $9.0B, three-quarters of annual capex, even though spending never grew.

Year five$9.0B

Year one$1.8B

Annual capex$12.0B

Here's the same hypothetical campus built once a year, with spending flat at $12 billion and each campus running for its full first year:

Flat spending, rising depreciation

Capex that yearYearly depreciation

246810$12B0

$12.0B$1.8B

Year 1Capex that year$12.0BYearly depreciation$1.8B

$12.0B$3.6B

Year 2Capex that year$12.0BYearly depreciation$3.6B

$12.0B$5.4B

Year 3Capex that year$12.0BYearly depreciation$5.4B

$12.0B$7.2B

Year 4Capex that year$12.0BYearly depreciation$7.2B

$12.0B$9.0B

Year 5Capex that year$12.0BYearly depreciation$9.0B

Year 1Year 2Year 3Year 4Year 5

Spending never grew, yet by year five depreciation is $9.0 billion, three-quarters of annual capex ($9.0bn ÷ $12bn). After that, the first campus's servers are fully depreciated and the charge grows far more slowly. That's why a capex boom squeezes free cash flow right away and reported margins later.

What the first-quarter numbers show

Closest disclosed measures, quarter ended March 31, 2026.

• Microsoft reported $31.9 billion of capex on its fiscal third-quarter earnings call. Roughly two-thirds went to short-lived assets, mainly GPUs and CPUs; the rest went to assets meant to support monetization over 15 years or more.
• Alphabet's first-quarter Form 10-Q showed $35.674 billion of property purchases and $6.482 billion of depreciation. Assets not yet in service jumped to $108.597 billion from $78.592 billion three months earlier.
• Amazon's first-quarter Form 10-Q showed $44.203 billion of cash property purchases and $12.834 billion of property depreciation and amortization. Net property additions were $54.757 billion, as non-cash items and unpaid equipment bills open another gap.
• Meta reported $19.84 billion of capex, including finance-lease principal, in its first-quarter results, and its Form 10-Q showed $5.68 billion of property depreciation.

The line that shows what's coming

Alphabet's filing includes the most useful single number here: assets not yet in service. That's spending already made that hasn't started depreciating.

It's a net figure, since assets also leave the pile when they go into service, but the message is plain. The balance of spending waiting to hit earnings grew by an amount equal to about 84% of the quarter's purchases. Every dollar in that $108.597 billion pile will start depreciating once it's switched on.

The not-in-service pile grew by 84% of the quarter's purchases

Alphabet property purchases, Q1 2026$35.7B

• 84%$30.0BIncrease in assets not yet in service

Don't multiply these ratios forward to forecast depreciation; asset mix, timing, leases and useful lives differ. They answer one question: are these companies building assets faster than today's income statement uses them up? For all four, yes. Our accounting comparison explains how the definitions differ.

Member research: The 3.9x gap is a four-company average. Issue 03 runs the audit company by company, including the assets waiting to enter service, commitment stacks and useful-life assumptions, and shows where the margin test lands next.

Why the gap doesn't prove a bubble

Depreciation reflects cash spent earlier, not a new payment. If a new data center is full, pricing is sensible and revenue grows faster than running costs, a big depreciation charge can sit alongside excellent economics.

Building ahead can be rational, too. Microsoft said it expected to stay capacity constrained through at least 2026, and a company with signed demand can lose value by building too slowly.

Each platform earns differently. Microsoft and Amazon sell capacity to cloud customers. Alphabet earns through Google Cloud and uses the same hardware for Search, ads and its own models. Meta earns mostly indirectly, through recommendations, engagement and ad conversion.

Each platform earns back the spending differently

Microsoft and Amazon

• Sell capacity to cloud customers

Alphabet

• Earns through Google Cloud
• Uses the same hardware for Search, ads and its own models

Meta

• Earns mostly indirectly
• Through recommendations, engagement and ad conversion

The real risk: idle depreciation

Underused depreciation is what destroys value. An accelerator ages whether it's serving a paying customer or waiting for one, and an empty data-center shell still costs money for maintenance, power, staff and financing. Three gaps can cause trouble:

• Capacity arrives before demand. Campuses take years to plan, permit, power and build. If big customers pull back, new capacity can hurt margins before the next phase can be cancelled.
• Hardware ages before it pays back. Accelerators improve fast. Older chips can move to inference (running trained models) or internal work, but their value can fall faster than the accounting life assumes.
• Price cuts absorb efficiency gains. Better chips and software lower the cost per result, but competition can hand those savings to customers. Revenue can grow while the return on each new dollar falls.

The same depreciation charge, opposite economics

Full data center

• Pricing is sensible
• Revenue grows faster than running costs
• A big depreciation charge sits alongside excellent economics

vs

Idle capacity

• The accelerator ages whether or not a customer pays
• The shell still costs maintenance, power, staff and financing
• Underused depreciation destroys value

Useful lives are a choice, not a law of physics

Management picks each asset's useful life. Stretching a server's life lowers yearly depreciation, and that can be reasonable if software keeps older hardware productive. But it brings back no cash; it only moves the cost in time.

Hypothetical: take the $8 billion of servers from the example above and stretch their life from five years to six.

Reported profit rises by about $270 million a year, and not a dollar of cash changes.

Stacked campuses and server lives set the charge together

Pick how many years of flat $12 billion spending, and the server life

Years of building

12345

Server life

5 years6 years

Yearly depreciation$1.80B$1.53B$3.60B$3.07B$5.40B$4.60B$7.20B$6.13B$9.00B$7.67B

−85.0%−87.3%−70.0%−74.4%−55.0%−61.7%−40.0%−48.9%−25.0%−36.1%below Annual capexbelow Annual capexbelow Annual capexbelow Annual capexbelow Annual capexbelow Annual capexbelow Annual capexbelow Annual capexbelow Annual capexbelow Annual capex

Annual capex $12.00B

$1.53B$9.00B

When a big buyer stretches server or network lives, read the accounting-policy note. Check whether it's still replacing its newest fleet quickly, whether older hardware is well used, whether impairments or assets held for sale are rising, and whether gross margin improves beyond the lower charge. A longer life backed by longer use is a real gain; one paired with fast replacement just flatters near-term margins.

Michael Burry has made these extensions the center of his AI bear case. We checked his $176 billion claim against the filings in Is Michael Burry right about AI depreciation?

Try it yourself

A 10-minute check using one company's filing:

1. Open Alphabet's first-quarter Form 10-Q and find property purchases, property depreciation and assets not yet in service. Recompute the 5.5x ratio ($35.674bn ÷ $6.482bn).
2. Repeat step 1 for one more company from the table, using its linked filing.
3. Check the company's full-year plan on its capex page, such as Alphabet's, to see how big this year's spending cohort will be.
4. Open the free AI capex, depreciation and return model. Split spending between equipment and buildings and watch the depreciation build year by year.
5. Add one year to the equipment's useful life and note what happens to earnings and to free cash flow.

You should now be able to answer: how much depreciation is already locked in for the next few years, and how much revenue growth would it take to absorb it?

What we'd watch

Our read: the $132 billion first-quarter figure is neither a victory nor a verdict. It's a funded claim on future gross profit. The good outcome needs revenue, utilization and efficiency to catch the cost as the fleet switches on. The bad one doesn't need AI demand to disappear, only the return on new capacity to fall below what today's spending implies.

Five numbers belong in the same model: capex, assets not yet in service, depreciation, gross margin (revenue left after the direct costs of serving it) and free cash flow (operating cash minus capex). On earnings calls, we'd listen for these signals:

What matters most is the return on the newest capacity, not on hardware installed years ago. The free AI capex ROI scorecard turns these checks into five tests.

Questions readers ask

Does the 3.9x gap mean profits are overstated?

No. Depreciation normally starts when an asset goes into service, so today's earnings correctly exclude assets still being built. The risk is misreading the numbers: judging a price-to-earnings multiple on current margins without allowing for the depreciation still to come.

Is depreciation a cash cost?

Not in the year it's booked. It reflects cash spent earlier, when the asset was bought. That's why free cash flow falls first and earnings follow later.

When will depreciation catch up with spending?

It can't be forecast by multiplying the latest quarter, because asset mix, timing, leases and useful lives differ. Watch assets not yet in service: when that balance moves into service, depreciation rises. The stacking example shows the charge can keep growing for years even if capex goes flat.

Why would a company extend useful lives?

Because it expects hardware to stay productive longer, for example when software keeps older chips useful for inference. The change lowers yearly depreciation but doesn't change cash. It's a real gain only if the hardware really does last longer.

Go deeper with member research

Depreciation is where the bill shows up. These member reports follow it into valuation:

• Meta jumped 11%. What must happen next to justify the price?: Muse, subscription economics, the cash-flow burden, legal costs and the earnings assumptions baked into the new share price.
• 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

• The four companies define spending and depreciation differently, so the 3.9x ratio is a warning sign, not an accounting identity.