Issue 03: The AI bill is arriving late to earnings

2. The real risk is not depreciation. It is underutilized depreciation. A fully occupied data center with strong pricing can produce excellent economics while carrying a large non-cash charge. A half-used campus with rapidly aging accelerators cannot.

3. The companies do not face the same test. Microsoft and Amazon can sell external cloud capacity. Alphabet can monetize infrastructure through Cloud, Search and its own models. Meta funds infrastructure primarily through an advertising engine. The path from silicon to revenue is different in each case.

4. Useful-life assumptions matter more than most earnings models admit. Extending a server's accounting life lowers annual depreciation without changing the cash already spent. That may be economically justified, but it can also make near-term margins look better while technological obsolescence accelerates.

5. The next signal is marginal margin, not total revenue growth. Revenue can rise while the return on the newest dollar of infrastructure falls. Investors need to compare cloud growth, gross margin, depreciation, capital expenditure and free cash flow together.

Our base case is that the leading hyperscalers can absorb the next stage of depreciation because demand remains strong and the largest platforms have multiple ways to monetize the same infrastructure. Our conviction is not uniform. Microsoft has the clearest contracted-demand bridge. Meta has the strongest current advertising subsidy but the largest visible commitment stack. Amazon already shows the cash-flow cost most clearly. Alphabet has the largest near-term placement-in-service queue relative to current depreciation.

The point is not to declare a capex bust. It is to identify the numbers that would reveal one before the market narrative catches up.

The accounting bridge most AI models skip

Capital expenditure is a cash-flow event. Depreciation is an earnings event. They meet only after an asset is placed in service.

Suppose a company commits $10 billion to a new data-center campus. The sequence can look like this:

1. The company signs equipment, construction, lease and power contracts.
2. Cash leaves as land is acquired, buildings are constructed and equipment is delivered.
3. The spending sits in property and equipment or construction in progress.
4. The asset becomes available for its intended use.
5. Depreciation begins over the estimated useful life.
6. Revenue depends on how quickly customers or internal products use the capacity.

The first two steps pressure cash flow. The fourth and fifth pressure reported margin. The sixth determines whether either burden earns an acceptable return.

Buildings and machines also run on different clocks. A data-center shell may be depreciated over decades. Servers and network equipment generally run through earnings over roughly two to six years, depending on the company and asset class. Microsoft says computer equipment is generally depreciated over two to six years. Alphabet uses six years for servers and network equipment. Meta uses five to 5.5 years. Amazon uses five to six years.

Servers run through earnings in two to six years

Useful life	Microsoft	Alphabet	Amazon	Meta
Servers and network equipment	2–6 years	6 years	5–6 years	5–5.5 years

This mix creates a deceptively smooth income statement. A company can report strong operating profit while its cash capital intensity rises because only a fraction of the new asset base is depreciating. Later, several annual investment cohorts overlap. Depreciation from the 2024 fleet remains in the income statement while the 2025 and 2026 fleets join it.

That is the depreciation clock.

It is not hidden accounting. The filings explain the policy. The analytical failure is treating today's depreciation as though it represents today's capital intensity.

The first-quarter gap

This table does not forecast future depreciation by multiplying the latest quarter. That would be wrong. Capital expenditure includes land and long-life buildings, supplier payments can move between quarters, finance leases are treated differently and some spending replaces assets already being depreciated.

The table answers a narrower question: is the physical asset base being funded faster than the current income statement is consuming it?

For all four companies, the answer is yes.

All four are funding assets faster than they depreciate them

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

The ratio is highest at Alphabet, but the more important clues sit one level below the cash-flow statement. Alphabet and Meta disclose unusually large queues of assets and contracts that have not yet entered normal operation.

The next cost wave is already contracted

At 31 March 2026	Alphabet	Meta
Unfinished assets	$108.6bn not yet in service	$61.0bn construction in progress
Leases not yet commenced	$75.6bn	$182.9bn
Non-cancelable commitments	—	$237.7bn

Do not add the boxes in the graphic. The categories can overlap and represent different claims. Use them as a map of how much infrastructure is still waiting to move from commitment into operation.

Alphabet has the cleanest visible depreciation queue

Alphabet's first-quarter filing provides the clearest picture of the lag.

The company spent $35.674 billion on capital expenditure in the quarter, more than double the $17.197 billion spent one year earlier. Property depreciation rose from $4.5 billion to $6.5 billion. The charge increased 44 percent, which is hardly benign, but spending still grew much faster.

Make a guess

At 31 March, what share of Alphabet's $281 billion of net property and equipment was not yet in service, and so not yet depreciating?

Tap your guess

0–10%Your guess10–20%Your guess20–30%Your guess30–40%Your guess40% or moreYour guess

01234567890123456789.0123456789%Not yet in service

Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was 38.6%.

$108.6B of $281.0B was still waiting to be placed in service, up from $78.6B three months earlier. Depreciation begins only when those assets are ready for use.

The balance sheet shows why the pressure can persist. Alphabet had $108.597 billion of assets not yet in service at 31 March, up from $78.592 billion three months earlier. That category equaled 38.6 percent of the company's $281.020 billion of net property and equipment.

The Form 10-Q states that these assets consist primarily of technical infrastructure and may take months or years to become ready for use. Depreciation begins when they are ready.

Alphabet also disclosed $75.6 billion of future payments under data-center leases that had not yet commenced. Those leases are expected to begin between 2026 and 2031. Again, this is not a number to add mechanically to construction in progress. It is evidence that the capacity pipeline extends beyond assets already visible in quarterly depreciation.

The bull case is unusually strong. First-quarter revenue rose 22 percent to $109.9 billion. Google Cloud revenue rose 63 percent to $20.0 billion. Consolidated operating margin expanded from 34 to 36 percent even as depreciation increased. Alphabet said it had signed agreements to supply multiple gigawatts of TPU hardware to customers, with most of the associated revenue expected in 2027. It also generated $45.8 billion of operating cash flow in the quarter.

In other words, the company is not spending into a revenue vacuum. It is growing into a capacity shortage.

The issue is timing. Alphabet spent $35.7 billion of cash on property in a quarter that produced roughly $10.1 billion of free cash flow before acquisitions and financing, calculated as operating cash flow less capex. It also issued $31.1 billion of senior notes and paused share repurchases during the quarter. The debt issuance had multiple corporate uses, so it should not be labeled AI financing alone. It still marks a change in the capital structure of a company that historically funded growth while returning large amounts of cash.

Capex consumed most of Alphabet's $45.8 billion of operating cash

Operating cash flow$45.8B

Capital expenditure−$35.7B= $10.1B

Free cash flow$10.1B

The most useful Alphabet signal is not whether total revenue beats consensus. It is whether Cloud growth and consolidated operating margin remain strong as assets not yet in service begin moving into depreciation. A slowing Cloud backlog combined with accelerating depreciation would weaken the thesis. Stable or rising Cloud margin while the depreciation charge expands would validate management's utilization and pricing assumptions.

Meta has the largest commitment stack

Meta's current margins make the infrastructure burden look manageable. First-quarter revenue rose 33 percent to $56.3 billion. Family of Apps operating margin was 48 percent. The consolidated operating margin held at 41 percent even though costs and expenses increased 35 percent.

That advertising engine is why Meta can spend at a level that would be impossible for most companies. It does not need to sell every accelerator-hour to an external customer. Better recommendation, ranking and advertising systems can monetize the capacity indirectly by improving engagement, ad conversion and price.

The cash commitment is still extraordinary.

Meta recorded $19.84 billion of first-quarter capital expenditure, including finance-lease principal, and raised its 2026 guidance to $125 billion to $145 billion. At the midpoint, annual spending would be more than seven times the $18.0 billion of property depreciation recorded in 2025.

Meta's 2026 capex guide is more than seven times last year's depreciation

2026 capex guidance$125–$145B

2025 property depreciation$18.0B

$0$50$100$150B

The March 2026 10-Q shows $61.017 billion of construction in progress, mostly data centers, network infrastructure and servers. It also shows $182.88 billion of lease obligations that had not begun and $237.67 billion of non-cancelable contractual commitments, mostly for third-party cloud capacity, servers, network infrastructure, data centers and Reality Labs hardware.

Those categories may overlap. They should not be presented as a $481 billion total. Their importance is contractual rigidity. If AI demand disappoints, Meta cannot reverse all of the spending with one annual budget decision.

Meta also offers the clearest example of useful-life sensitivity. In January 2025, it extended the estimated useful life of most servers and network assets to 5.5 years. The company estimated that the change would reduce 2025 depreciation by approximately $2.9 billion on the assets already in service at the end of 2024.

There is nothing inherently aggressive about that change. Servers can remain economically useful for longer when software scheduling improves, older chips handle less demanding inference and internal workloads are matched to the right hardware. The accounting estimate may be entirely reasonable.

Investors should still understand the mechanism. Cash did not return when the useful life changed. The annual earnings charge simply spread across more periods. If AI hardware becomes obsolete faster than expected, the risk reappears through lower utilization, impairment or weaker economics on the old fleet.

Meta's decisive metric is incremental monetization. Ad impressions rose 19 percent and average price per ad rose 12 percent in the first quarter. If those measures remain healthy while depreciation rises, the advertising flywheel is funding the build. If infrastructure costs accelerate while ad price, impressions or engagement soften, the subsidy begins to crack.

The 29 July earnings call should answer three questions:

• How much of the higher capex guide reflects component inflation rather than added capacity?
• How quickly is depreciation expected to rise in the second half?
• Are AI-driven advertising gains still expanding faster than infrastructure cost?

Microsoft has the best demand bridge and a real margin trade-off

Microsoft presents the strongest version of the bull case.

Fiscal third-quarter revenue rose 18 percent to $82.9 billion. Microsoft Cloud revenue exceeded $54 billion and grew 29 percent. Azure grew 40 percent. Management said its AI business had surpassed a $37 billion annual revenue run rate, up 123 percent. It also described commercial revenue still to be recognized above $600 billion and demand that continued to exceed available capacity.

That is the demand bridge investors want to see. Microsoft is not building solely against an internal forecast. It has cloud commitments, customer workloads and first-party applications competing for the same incoming capacity.

The cost is no longer invisible. Company gross margin fell to 68 percent, and Microsoft attributed the pressure to AI infrastructure investment and growing AI product usage. Its March 2026 10-Q shows quarterly depreciation of $9.0 billion, up from $5.8 billion one year earlier. Nine-month depreciation reached $24.0 billion, up from $15.7 billion.

Microsoft's depreciation is up more than 50% in a year

Fiscal 2025Fiscal 2026

5101520$25B0

$5.8B$9.0B

Fiscal third quarterFiscal 2025$5.8BFiscal 2026$9.0B

$15.7B$24.0B

First nine monthsFiscal 2025$15.7BFiscal 2026$24.0B

Fiscal third quarterFirst nine months

At the same time, net property and equipment increased from $205.0 billion at 30 June 2025 to $283.2 billion at 31 March 2026. Servers, network equipment and software at cost increased by $58.0 billion in nine months. Buildings and improvements increased by $34.3 billion.

Spot the red flag

Microsoft spent $31.9 billion in the quarter. Which detail means that spending reaches earnings fastest?

Microsoft fiscal Q3 2026: what the filing and call disclosed

Net property and equipment rose to $283.2B from $205.0BThat shows the asset base growing, not how quickly it depreciates.Buildings and improvements rose $34.3B in nine monthsBuildings depreciate over long lives, so they reach earnings slowly.Roughly two-thirds of capex went to short-lived assets, mainly GPUs and CPUsShort-lived assets run through depreciation in a few years, so they hit margins soonest.Quarterly depreciation was $9.0B, up from $5.8BThat is the charge already showing. The question is what drives the next one.

Tap the line you think is the tell.

Spotted it.Not that line. Microsoft depreciates computer equipment over two to six years. With about two-thirds of the quarter's capex in GPUs and CPUs, the cost moves into earnings quickly, which is why gross margin is already under pressure.

Explain every lineHide the other explanations

The fiscal Q3 earnings call adds the most important detail: roughly two-thirds of the quarter's $31.9 billion of capex was for short-lived assets, primarily GPUs and CPUs. Those assets are more directly connected to near-term revenue than buildings, but they also move through depreciation faster.

Microsoft expects calendar 2026 capex of about $190 billion, including roughly $25 billion of higher component pricing. Fiscal fourth-quarter capex was expected to exceed $40 billion. The company says it can accelerate Azure growth as the new capacity comes online and that efficiency improvements will help offset the cost.

This creates a clean test. Microsoft does not need depreciation to stop rising. It needs Azure growth, Copilot monetization and fleet efficiency to grow gross profit dollars faster than depreciation and energy costs.

The risk is visible in product usage. An AI service can grow revenue while consuming much more infrastructure per user than a traditional software subscription. Revenue growth therefore does not guarantee software-like incremental margin. Microsoft has already changed GitHub Copilot pricing to align more closely with usage and value. That is a rational response to a product whose compute cost scales with activity.

For Microsoft, the most important 29 July numbers are:

• Actual fiscal Q4 capex versus the greater-than-$40-billion guide
• The short-lived asset mix
• Azure growth and the capacity constraint
• Microsoft Cloud gross margin
• Fiscal 2027 depreciation and capex direction
• Evidence that Copilot revenue per user is rising faster than inference cost per user

If Azure accelerates and cloud gross margin stabilizes, the depreciation wave is being absorbed. If Azure accelerates but gross margin keeps falling, growth is being purchased at a lower incremental return.

Same Azure growth, two different verdicts

Depreciation absorbed

• Azure accelerates
• Cloud gross margin stabilizes

The depreciation wave is being absorbed

vs

Growth bought at a lower return

• Azure accelerates
• Gross margin keeps falling

Growth is purchased at a lower incremental return

Amazon is already showing the cash-flow version

Amazon's income statement looks strong. AWS first-quarter revenue rose 28 percent to $37.6 billion. AWS operating income reached $14.2 billion, producing a 37.7 percent operating margin. Infrastructure investment did not prevent the cloud segment from expanding profit.

Cash flow tells a harsher story.

Amazon spent $44.203 billion in cash on property and equipment in the first quarter. Total net additions to property and equipment were even higher at $54.757 billion because that measure includes finance leases and property acquired but not yet paid. AWS alone accounted for $41.516 billion of net additions.

The March 2026 10-Q allocated $7.277 billion of quarterly property depreciation and amortization to AWS. AWS net additions were therefore 5.7 times current allocated depreciation.

Make a guess

Amazon generated $148.5 billion of operating cash flow over the past twelve months. How much free cash flow was left after property purchases?

Tap your guess

$0–10BYour guess$10–30BYour guess$30–60BYour guess$60B or moreYour guess

$0123456789.0123456789BLatest twelve months

Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was $1.2B.

Net property purchases consumed $147.3B, leaving company-defined free cash flow of $1.232B, down from $25.925B a year earlier.

Latest twelve months$1.2B

A year earlier$25.9B

Amazon's trailing-twelve-month operating cash flow reached $148.5 billion. Cash purchases of property, net of sales and incentives, consumed $147.3 billion. Company-defined free cash flow fell to $1.232 billion from $25.925 billion one year earlier.

That does not mean Amazon is running out of money. It means almost all operating cash flow is currently being recycled into physical capacity. The company raised external financing, and cash flow can recover if the capex rate slows or operating income rises. The current figure still removes the comfort of treating depreciation as the whole infrastructure cost.

Amazon also reversed part of an earlier useful-life extension. It moved server lives from five to six years in 2024, then shortened a subset of servers and networking equipment from six to five years in 2025. The 2025 change increased annual depreciation and amortization by $1.4 billion and reduced net income by $1.0 billion.

That adjustment is useful because it demonstrates that useful lives can move in both directions. New software can extend hardware value. Faster product cycles, energy inefficiency or technical obsolescence can shorten it.

The Amazon test is therefore two-sided:

• AWS operating margin must remain strong enough to justify the asset base.
• Consolidated free cash flow must recover before financing becomes a permanent substitute for internally funded growth.

Amazon could pass both. AWS sales are growing quickly, operating margin is high and much of the infrastructure supports internal retail, advertising and logistics workloads as well as external cloud customers. The burden of proof is simply higher when a business converts $148.5 billion of operating cash flow into $1.2 billion of free cash flow.

The bull case is stronger than the accounting warning sounds

A depreciation-lag thesis can become a lazy bear argument if it ignores why the companies are spending.

The strongest counter-hypothesis is straightforward: the buildout is demand-backed, the platforms remain capacity constrained and the new fleet will produce revenue faster than depreciation rises.

Microsoft says demand exceeds supply and expects Azure growth to accelerate as capacity arrives. Alphabet's Cloud revenue rose 63 percent and its TPU supply agreements include revenue expected in 2027. AWS revenue rose 28 percent while its operating margin remained near 38 percent. Meta's advertising system produced simultaneous growth in impressions and price.

All three clouds grew 28% or faster

Google Cloud revenue+63%

Azure+40%

AWS revenue+28%

The companies also have several efficiency levers.

Custom silicon can lower unit cost. A first-party accelerator does not need to beat every general-purpose GPU. It needs to lower the cost of a high-volume internal workload. Microsoft, Alphabet, Amazon and Meta all have some ability to match models to their own chips, networking and software.

Software can increase utilization. Better schedulers, lower-precision computation, model distillation, caching and workload placement allow the same installed fleet to produce more tokens or serve more requests.

Old hardware can move down the workload stack. The newest accelerator can train or serve the most demanding model. An older server can handle ranking, recommendations, search retrieval, smaller models or batch jobs. That supports a longer useful life.

Revenue can arrive through more than cloud rental. Meta can improve ad conversion. Alphabet can improve Search, YouTube and Cloud. Microsoft can sell Azure capacity and raise the value of Microsoft 365, GitHub and security products. Amazon can monetize AWS while using the same infrastructure across advertising, retail and logistics.

Long-lived assets have option value. Land, grid access, buildings and network rights can remain scarce even if a specific accelerator generation becomes obsolete. Replacing machines inside an energized campus is much easier than creating a new powered site.

These points prevent a simplistic conclusion. High capex does not automatically imply low return. The hyperscalers have the best balance sheets, distribution and engineering talent in the industry. If anyone can make the economics work, they can.

The correct question is not whether the spending is large. It is whether marginal revenue and gross profit are arriving before the overlapping depreciation cohorts become a drag.

Three ways the economics can break

1. Utilization arrives too slowly

Capacity can be technically available but economically underused. Enterprise deployments may take longer than expected, model providers may optimize away some compute demand, or customers may resist prices needed to produce an adequate return.

The earliest signal would be a gap between new capacity and cloud growth. Management would stop describing supply as the constraint. Backlog growth would slow. Discounting or long-term contract incentives would increase.

2. The fleet ages faster than the accounting life

AI accelerators can remain functional while becoming uneconomic. A newer chip may provide much better performance per watt or token. If electricity, cooling and floor space are scarce, an older accelerator can lose value before it stops working.

That risk does not necessarily appear first as an impairment. It can show up as lower utilization, higher cost per unit or the need to buy a new fleet while the old fleet is still depreciating.

Useful-life changes deserve scrutiny for this reason. A longer estimate is bullish only if the company can keep the asset productive.

3. Fixed commitments outlive demand

Construction projects, power contracts and leases cannot be cancelled as quickly as a quarterly equipment order. Meta's uncommenced leases and non-cancelable commitments, Alphabet's future data-center lease payments and Microsoft's finance leases extend the investment clock beyond current capex.

If demand remains strong, those commitments secure scarce capacity and protect future growth. If demand slows, they turn flexibility into fixed cost.

This is why lease-adjusted analysis matters. A company can slow reported cash capex while previously signed leases begin, keeping the economic commitment alive.

Three ways the economics can break

Utilization arrives too slowly

• Management stops citing supply as the constraint
• Backlog growth slows
• Discounts and contract incentives rise

The fleet ages faster than its accounting life

• Utilization falls
• Cost per unit rises
• A new fleet is bought while the old one still depreciates

Fixed commitments outlive demand

• Leases and power contracts cannot be cancelled quickly
• Signed leases begin even as reported capex slows
• Flexibility turns into fixed cost

A practical dashboard for members

The next two years should be read through six linked measures.

No single row decides the thesis.

A falling capex-to-depreciation ratio can be healthy because spending normalizes. It can also be unhealthy because management cuts investment after demand weakens. Rising depreciation can be healthy when a revenue-producing campus goes live. It becomes dangerous only when revenue, utilization and margin fail to follow.

The most revealing comparison is therefore a three-part bridge:

Capacity placed in service → incremental revenue → incremental gross profit.

Companies disclose the first and third pieces imperfectly. Investors must reconstruct them from capex, assets not in service, cloud growth, margin commentary and free cash flow.

Company scorecard

The scorecard deliberately avoids a ranking by share-price upside. This is an operating-economics test, not a recommendation to buy or sell the four stocks.

Microsoft currently has the most visible external demand bridge. Alphabet has the strongest combination of cloud acceleration and diversified internal monetization, but also the largest disclosed assets-not-in-service queue. Meta has extraordinary advertising economics and equally extraordinary fixed commitments. Amazon has the clearest current warning in free cash flow, offset by a highly profitable AWS segment.

Three scenarios through 2028

Base case: revenue catches the cost

Capacity remains constrained through 2026, cloud and AI revenue grow faster than the installed asset base, and depreciation rises without causing a large margin reset. Capex growth slows in 2027 as the first build phase matures. Free cash flow recovers.

In this scenario, the market should reward companies that disclose a clean link between capacity additions and revenue. Suppliers remain supported, but the return begins migrating from component scarcity toward platform utilization and software efficiency.

Bull case: the new fleet is more productive than the old accounting suggests

Custom silicon, model optimization and higher utilization lower cost per token faster than prices decline. AI improves existing advertising, productivity and cloud products while creating new revenue. Older hardware remains useful across a tiered workload stack.

Depreciation rises, but gross profit rises faster. Useful lives prove conservative rather than aggressive. The hyperscalers fund the build internally and resume larger capital returns without sacrificing growth.

Bear case: overlapping fleets expose weak marginal returns

Enterprise adoption grows more slowly, model efficiency reduces required compute per task and competitive pricing compresses cloud revenue per unit. The 2024 to 2026 fleets depreciate together while companies continue buying newer accelerators to remain competitive.

Cloud growth slows before capex and lease commitments can be reduced. Free cash flow remains depressed. Companies extend useful lives, move assets to held-for-sale or take impairments. The market shifts from rewarding total AI revenue to penalizing poor asset turnover.

The bear case does not require AI to fail. It requires the infrastructure return to fall below what current valuations assume.

Three scenarios through 2028

Base caseRevenue catches the cost

• Capacity stays constrained through 2026
• Capex growth slows in 2027
• Free cash flow recovers

Bull caseThe new fleet is more productive

• Cost per token falls faster than prices
• Older hardware stays useful in a tiered stack
• Gross profit rises faster than depreciation

Bear caseOverlapping fleets expose weak returns

• Cloud growth slows before commitments can be cut
• Free cash flow stays depressed
• Longer lives, held-for-sale moves or impairments

What would change our view

We would become more constructive if three developments occur together:

• Cloud and AI revenue accelerate as new capacity is placed in service.
• Gross margin stabilizes despite faster depreciation.
• Free cash flow improves because operating cash flow grows faster than capex.

We would become more cautious if:

• Management stops describing capacity as constrained while capex remains elevated.
• Construction in progress or assets held for sale rise without a clear customer bridge.
• Useful lives are extended again while fleet replacement remains aggressive.
• Lease-adjusted commitments keep growing after reported capex slows.
• Cloud price competition increases as depreciation and energy expense accelerate.

The sharpest red flag would be a company reporting strong total revenue growth while cloud or advertising gross profit fails to absorb the newest infrastructure charge. That would show that AI is increasing activity without increasing economic value at the same rate.

The final judgment

The AI infrastructure boom has passed its first test. Revenue is real. Microsoft, Alphabet, Amazon and Meta are not building empty demonstration sites. Their cloud, advertising and software businesses are growing, and several remain supply constrained.

The second test is harder.

The first wave asked whether the companies could buy enough accelerators and secure enough power. The next wave asks whether each dollar of installed capacity can produce enough gross profit before the hardware ages and the next generation must be purchased.

From access to compute to return on compute

The first wave asked

• Can the companies buy enough accelerators?
• Can they secure enough power?

Passed: the revenue is real

vs

The next wave asks

• Does each dollar of installed capacity produce enough gross profit?
• Does it do so before the hardware ages and the next generation must be bought?

Decided by the next margin bridge

Current earnings do not fully answer that question because the spending and the charge arrive on different clocks. Cash capex is already enormous. Construction in progress, assets not yet in service and uncommenced leases show more capacity waiting behind it. Depreciation is rising, but it is still catching up.

That does not make today's profit fictional. It makes the next margin bridge decisive.

The companies that win will not merely report the fastest AI revenue growth. They will convert capital into available capacity quickly, keep that capacity busy, lower cost per unit through hardware and software, and preserve free cash flow after the build matures.

The market has spent two years valuing access to compute. It is about to start valuing the return on compute.

---

Primary source ledger

1. Microsoft fiscal Q3 2026 earnings call, 29 April 2026. Capex, short-lived asset mix, AI revenue run rate, Azure outlook and calendar 2026 spending guidance.
2. Microsoft Form 10-Q for the quarter ended 31 March 2026. Property and equipment, depreciation, margins and cash-flow data.
3. Microsoft fiscal Q4 2026 earnings-date announcement, 8 July 2026.
4. Alphabet Form 10-Q for the quarter ended 31 March 2026. Capex, depreciation, assets not yet in service, leases, debt issuance, Cloud revenue and operating margin.
5. Alphabet investor FAQ on capex and depreciation. Asset categories, placement-in-service policy and useful lives.
6. Alphabet fiscal 2025 Q4 earnings call, 4 February 2026. 2026 investment mix and depreciation outlook.
7. Amazon Form 10-Q for the quarter ended 31 March 2026. Cash property purchases, net additions, AWS depreciation, AWS revenue and margin, and free cash flow.
8. Amazon Form 10-K for the year ended 31 December 2025. Server useful lives and the earnings effect of the 2025 estimate change.
9. Meta Form 10-Q for the quarter ended 31 March 2026. Capex, depreciation, construction in progress, leases, commitments, advertising metrics and segment margin.
10. Meta first-quarter 2026 results, 29 April 2026. Cash flow, full-year capex guidance and financial highlights.
11. Meta Form 10-K for the year ended 31 December 2025. Server and network useful lives and 2025 depreciation.
12. Meta Q2 2026 earnings-date announcement, 14 July 2026.