AI Capex ROI Scorecard: Is the Spending Paying Off?

Meta spent $31.08 billion on capex in the second quarter and was left with just $784 million of free cash flow. Amazon's trailing free cash flow turned negative, even as AWS sales grew 37%.

Watch it play outMeta spent almost all of its cash on the build

In the second quarter, Meta's operations brought in roughly $31.86 billion of cash.

It poured $31.08 billion of that into capex: data centers, chips and the rest of the build.

Zoom in on what stayed behind: $784 million of free cash flow.

Pull back out. Almost every dollar went back into the build; whether it pays off is still ahead.

Replay

So is AI spending paying off? Not proven yet. It pays off only when new capacity earns revenue or cuts costs faster than depreciation, power, financing and replacement bills grow. Demand is strong right now, but the cash cost is large and the payback is still ahead.

Watch it play outAI spending pays off only if earnings outrun the bills

Every dollar of AI capex buys new capacity: data centers and chips.

Once it runs, the bills start: depreciation, power, financing and, later, replacement.

It pays off only if it earns new revenue or cuts costs...

...faster than those bills grow. Demand is strong, but the cash cost is large and the payback is still ahead.

Replay

Below are five tests for telling which way it's going, plus a two-line calculation that shows how much of each dollar of cash these companies are putting back into the build.

Check the claimThree claims to test as you read

ClaimIf revenue is growing, the AI spending is paying off.Tap to checkMythRevenue can rise while the return on each new dollar of infrastructure falls, if price cuts or rising depreciation and power costs eat the gains.Flip backClaimA falling gross margin means AI spending is failing.Tap to checkMythNot necessarily. A productive ramp can squeeze margins first. The warning is depreciation speeding up without matching gains.Flip backClaimMeta's $784 million of free cash flow means it's in trouble.Tap to checkMythIt implies about $31.86 billion of operating cash; the low figure reflects how much Meta chose to reinvest.Flip back

The short version

• The scale: Microsoft, Amazon, Alphabet and Meta plan a combined $720 billion to $740 billion of 2026 capex, the money spent on long-lived assets like data centers and chips (tracker; Meta's $135 billion to $145 billion is from its third-quarter results).

Watch it play outFour companies, up to $740 billion of capex in one year

Microsoft, Amazon, Alphabet and Meta are all building at once.

Pour their 2026 plans together: $720 billion to $740 billion of capex, the money spent on long-lived assets like data centers and chips.

For scale, Meta's whole second-quarter spend, $31.08 billion, is this small circle.

Replay

• Demand is real: Microsoft said Azure demand exceeded capacity, and Amazon's AI and chip businesses each run above $25 billion a year.
• Cash is the pressure point: Amazon's trailing free cash flow swung to a $7.6 billion outflow, driven mainly by a $66.1 billion year-over-year jump in property and equipment purchases.
• Five tests decide it: demand, utilization, unit economics, earnings and cash. No single quarter can pass all five.

Figures are as of our August 12, 2026 research cutoff. The AI capex tracker has each company's spending plan; this scorecard tests the outcome.

The five-part scorecard

The point is to write down the evidence before the market's story changes.

1. Is demand signed, or just expected?

Capacity shortages are a good sign when paying customers are waiting. They mean less when a company is building against a broad forecast with no clear workloads to fill it.

Microsoft said Azure demand kept exceeding available capacity, and it disclosed a large commercial remaining-performance-obligation balance (contracted revenue not yet delivered) on its fiscal third-quarter earnings call. The next test is whether new capacity turns those commitments into revenue without hurting margins.

Amazon's second-quarter 2026 results showed 37% AWS sales growth and annualized run rates above $25 billion for both its AI and chip businesses. That beats a press release about a future campus: customers are using the platform now.

Amazon's demand shows up as revenue now

+37%AWS sales growth, Q2 2026

$25B+AI business, annualized run rate

$25B+Chip business, annualized run rate

Backlog and run rates still aren't guaranteed returns. Contract length, pricing, credits, power and the capital needed to serve each workload decide the result.

2. How fast does capacity start earning?

An ordered chip isn't revenue. A finished building without a grid connection isn't usable capacity. Even a powered cluster can lose money if software, networking or slow customer rollouts keep it underused.

Each step takes time:

1. Capital is committed.
2. Equipment and buildings are delivered.
3. Power and networking come online.
4. The asset is placed in service.
5. A workload reaches productive use.
6. Revenue or an internal cost saving arrives.

Six steps stand between spending cash and earning from it

1. Step 1CommitCapital is committed
2. Step 2DeliverEquipment and buildings arrive
3. Step 3Power upPower and networking come online
4. Step 4In serviceDepreciation starts
5. Step 5WorkloadReaches productive use
6. Step 6EarnRevenue or a cost saving

Microsoft said it cut dock-to-live times for new GPUs and brought a major data center online ahead of schedule. Both shorten the gap between spending cash and earning from it.

Companies rarely disclose utilization (how busy the hardware is) as one clean number. Piece it together from capacity comments, deployment timing, cloud growth, backlog conversion and gross margin.

3. Do efficiency gains reach shareholders?

AI can get dramatically cheaper per task while the return on infrastructure still falls. Competitors may pass savings to customers as lower prices, demand may grow so fast that total spending rises anyway, and older hardware may drift to lower-value work.

The real test is:

Price received minus the full cost of compute, memory, networking, power, cooling, software, and capital.

Microsoft said its Maia 200 accelerator improved tokens per dollar compared with the latest silicon in its fleet. Amazon reports strong adoption of its custom Trainium and Graviton chips. Both support a cost-control path, but check them against outside usage, gross margin and the engineering cost of running a proprietary stack.

Cheaper AI doesn't automatically mean better returns

Gains reach shareholders

• Cost per useful result falls while price holds
• Custom chips like Maia 200 improve tokens per dollar
• Trainium and Graviton adoption supports cost control

vs

Gains leak away

• Competitors pass savings on as lower prices
• Demand grows so fast total spending rises anyway
• Older hardware drifts to lower-value work

Our custom AI chips versus GPUs guide explains why a mixed fleet can make economic sense without one chip "winning" every job.

4. Can gross profit outrun depreciation?

Capex hits cash flow before its full cost reaches earnings. Depreciation, the yearly charge that spreads an asset's cost over its useful life, starts when the asset goes into service. It speeds up as several years of spending pile on top of each other.

The early pressure is visible:

• Microsoft's company gross-margin percentage fell, partly because of AI infrastructure and product usage.
• Alphabet has warned that depreciation and data-center operating costs will rise.
• Meta's second-quarter operating margin fell while quarterly capex, including finance-lease principal, reached $31.08 billion, according to its official results.

None of that proves failure. A productive ramp can squeeze margins before scale and pricing catch up. The warning comes when depreciation speeds up without matching gains in revenue, engagement, ad pricing, cloud growth or cost per result. Our AI depreciation analysis connects assets not yet in service to the expense still coming.

5. Does free cash flow recover?

Free cash flow, operating cash minus capex, shows the immediate cost of the build.

These snapshots aren't directly comparable, but they show why earnings per share can't tell the whole story. Amazon captures the debate in one company: strong demand and a heavy cash cost at the same time.

A successful build doesn't need free cash flow to rise every quarter. It needs a believable path from today's spending to future cash. If that path keeps sliding further out while commitments and borrowing (leverage) grow, investors should demand a higher return.

Member research: When the path to free cash flow stretches out, someone has to fund the gap. Issue 05 shows who is stepping in, as AI infrastructure moves into private credit, insurance balance sheets and securitized debt.

Worked example: how much cash goes back into the build?

Free cash flow plus capex gives you, roughly, the operating cash a company generated. Dividing capex by that figure shows how much of each operating dollar was reinvested. We use the simple definition from this article (free cash flow = operating cash minus capex); company definitions vary, so treat the results as approximate.

Make a guess

Meta spent $31.08 billion on capex in Q2 and kept $784 million of free cash flow. What share of its operating cash went back into the build?

Tap your guess

Under 50%Your guess50% to 75%Your guess75% to 90%Your guess90% or moreYour guess

01234567890123456789%Meta, Q2 2026

Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was 98%.

About 98%: $31.08 billion of capex out of about $31.86 billion of implied operating cash. Microsoft reinvested about 67% in its fiscal third quarter.

Meta, Q2 202698%

Microsoft, fiscal Q367%

The quarters differ (Meta's is April to June; Microsoft's fiscal third quarter is January to March), but the contrast is clear. Meta is putting almost every operating dollar back into infrastructure. Microsoft keeps about a third.

Now a hypothetical to show why free cash flow swings so hard. Suppose Meta's operating cash rose 10% with capex unchanged:

A 10% gain in operating cash would roughly quintuple free cash flow, from $0.784 billion to about $3.97 billion. The reverse is also true. When nearly all cash is being reinvested, a small shortfall in operations or a small rise in capex can push free cash flow below zero, which is what happened to Amazon's trailing figure.

When nearly all cash is reinvested, small changes swing free cash flow

Move Meta's operating cash and capex

Operating cash vs Q2

−10%Unchanged+10%

Capex vs Q2

Unchanged+10%

Hypothetical free cash flow−$2.41B−$5.51B$0.78B−$2.33B$3.97B$0.86B

−$3.19B−$6.29B−$0.00B−$3.11B+$3.19B+$0.08Bbelow Meta's Q2 actualbelow Meta's Q2 actualbelow Meta's Q2 actualbelow Meta's Q2 actualabove Meta's Q2 actualabove Meta's Q2 actual

Meta's Q2 actual $0.78B

−$5.51B$3.97B

How each company gets paid back

Don't force one metric on all four. Meta doesn't need to sell cloud capacity for its AI spending to pay off. Amazon shouldn't get credit for its whole companywide plan based on AWS growth alone.

Try it yourself

Score one company in about 10 minutes:

1. Open its capex page, for example Microsoft or Amazon, and note its latest quarterly capex.
2. From the results linked in this article, find the same quarter's revenue, operating cash flow and the prior quarter's capex.
3. Enter them into the free capex-to-free-cash-flow coverage calculator. Then repeat the worked example: what share of operating cash went back into the build?
4. Open the free AI capex, depreciation and return model and compare its bear and base cases to see how much revenue the build needs to earn a return.
5. Mark each of the five tests in the scorecard table as pass, fail or unclear.

You should now be able to answer: which tests does this company pass today, and which result next quarter would change your mind?

Our read

The bull case is that heavy spending builds scarce, busy capacity whose revenue and efficiency gains grow faster than its cost. The bear case doesn't need AI demand to vanish. It only needs the newest dollar of infrastructure to earn less than investors assumed.

The bear case doesn't need AI demand to vanish

Bull case

• Heavy spending builds scarce, busy capacity
• Revenue and efficiency gains grow faster than cost

vs

Bear case

• Demand can stay strong
• The newest dollar of infrastructure earns less than investors assumed

Before the next earnings cycle, write down what you expect for each question:

• Is demand still above available supply?
• How much new capacity entered service?
• Did cloud or product usage speed up when it arrived?
• Did gross margin improve after one-time items?
• Did depreciation grow faster or slower than gross profit?
• Did free cash flow improve because operations got stronger, or because spending slipped?
• Were finance leases and purchase commitments still growing off the cash-flow statement?

Mostly yes on demand, usage and margins would move us toward the bull case. Rising depreciation and commitments with flat margins would move us the other way.

Questions readers ask

If Meta kept only $784 million of free cash flow, is it in trouble?

Not on this evidence alone. The worked example implies roughly $31.86 billion of operating cash in the quarter; low free cash flow reflects how much Meta chose to reinvest. The real question is whether that $31.08 billion of capex earns a return through engagement and ad pricing.

Why isn't revenue growth enough?

Revenue can rise while the return on each new dollar of infrastructure falls. If price cuts pass efficiency gains to customers, or depreciation and power costs grow faster than revenue, growth doesn't reach shareholders. That's why the scorecard also tests unit economics, earnings and cash.

Does a falling gross margin mean AI spending is failing?

Not necessarily. A productive capacity ramp can squeeze margins before scale and pricing catch up. The warning sign is depreciation speeding up without matching gains in revenue, engagement, cloud growth or cost per result.

How does this connect to a stock's valuation?

A price-to-earnings multiple assumes a path for future profits. If the AI build needs more capex, or pays back more slowly than that path assumes, the earnings behind the multiple are at risk even if revenue keeps growing.

Go deeper with member research

The scorecard gives you the tests. Member reports run them against real companies and put a number on what the share price needs:

• 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.
• 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.

Limits of this data

• Figures and guidance are as of August 12, 2026 and can change.
• Utilization is rarely disclosed directly, so parts of this scorecard rely on proxies.