AI Capex ROI Scorecard: The Evidence That Spending Is Paying Off

Direct answer: AI capital spending is paying off only when new capacity produces durable revenue or operating gains faster than depreciation, power, financing, and replacement costs grow. Revenue growth alone is not enough. The strongest evidence is a combination of demand, utilization, unit economics, margin resilience, and free-cash-flow recovery.

Microsoft, Amazon, Alphabet, and Meta plan a combined $695 billion to $720 billion of 2026 capex. The AI capex tracker establishes the spending envelope. This scorecard tests the outcome.

The five-part AI capex ROI scorecard

Test	Evidence that supports the thesis	Evidence that weakens it
Demand	Backlog converts, usage grows, capacity remains constrained	Commitments slip, customers optimize away spend
Utilization	New clusters reach productive load quickly	Energized assets sit below planned utilization
Unit economics	Cost per useful result falls while price holds	Efficiency gains are passed through faster than cost falls
Earnings	Gross profit absorbs depreciation and operating costs	Margin falls as new capacity enters service
Cash	Operating cash flow catches capex after the build	Investment stays structurally above cash generation

No single quarterly result can pass all five tests. The point is to make the evidence explicit before the market narrative changes.

Test one: is demand contracted or merely expected?

Capacity constraints are constructive when they reflect paying demand. They are less useful when management is building against a broad forecast without a clear route to workload deployment.

Microsoft reported that Azure demand continued to exceed available capacity and disclosed a large commercial remaining-performance-obligation balance in its fiscal third-quarter earnings call. That supports the demand side of the scorecard. The next test is whether capacity placed online converts those commitments into revenue without sacrificing economics.

Amazon’s second-quarter 2026 results reported 37% AWS sales growth and annualized run rates above $25 billion for both its AI and chip businesses. Those are stronger operating signals than a press release announcing a future campus. They show customers using the platform now.

Still, backlog and run rates are not guaranteed returns. Contract duration, pricing, credits, power availability, and the capital required to serve the workload determine the result.

Test two: how quickly does capacity become productive?

An ordered accelerator is not revenue. A completed building without a grid connection is not usable capacity. Even an energized cluster can destroy value if software, networking, or customer deployment keeps utilization low.

The useful sequence is:

Capital is committed.
Equipment and facilities are delivered.
Power and networking become available.
The asset is placed in service.
A workload reaches productive utilization.
Revenue or an internal operating benefit arrives.

Track the time between each step. Microsoft said it reduced dock-to-live times for new GPUs and brought a major data center online ahead of schedule. Those are operational improvements because they shorten the period between cash investment and revenue-producing capacity.

Utilization is rarely disclosed as one clean percentage. Investors have to use capacity constraints, deployment timing, cloud growth, backlog conversion, and gross-margin commentary as a combined proxy.

Test three: do efficiency gains reach shareholders?

AI systems can become dramatically cheaper per task while the return on infrastructure still falls. Competition may pass the savings to customers through lower prices. Model demand may expand so quickly that total spending rises despite falling unit cost. Older hardware may be reassigned to lower-value workloads.

The relevant equation is not simply “tokens per dollar improved.” It 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 relative to the latest silicon in its fleet. Amazon reports strong adoption of custom Trainium and Graviton chips. Those claims support a cost-control path, but they should be tested against external usage, gross margin, and the engineering cost of maintaining a proprietary stack.

The custom AI chips versus GPUs guide explains why a heterogeneous fleet can be economically rational without one architecture “winning” every workload.

Test four: can gross profit outrun the depreciation clock?

Capital expenditure reaches cash flow before the full cost reaches earnings. Depreciation begins when assets are placed in service and can accelerate as several investment cohorts overlap.

Microsoft reported a lower company gross-margin percentage partly because of AI infrastructure investment 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 in its official results.

None of those facts proves the investment is failing. A productive capacity ramp can pressure margin before scale and pricing absorb the cost. The warning appears when depreciation accelerates without a matching improvement in revenue, engagement, advertising yield, cloud growth, or cost per useful result.

Use the AI depreciation analysis to connect assets not yet in service with the expense still ahead.

Test five: does free cash flow recover?

Free cash flow exposes the immediate cost of the build. Amazon’s trailing free cash flow moved to a $7.6 billion outflow in the second quarter, driven primarily by a $66.1 billion year-over-year increase in property-and-equipment purchases net of incentives. At the same time, AWS growth accelerated.

That is the central AI investment debate in one company: strong operating demand and a large current cash cost can both be true.

Meta reported only $784 million of second-quarter free cash flow while spending $31.08 billion on capex under its stated definition. Microsoft reported $15.8 billion of quarterly free cash flow after $31.9 billion of capex. These snapshots are not directly comparable, but they show why earnings per share cannot carry the entire analysis.

A successful build does not require free cash flow to rise every quarter. It requires a credible bridge from current investment to future cash generation. If that bridge keeps moving further into the future while commitments and financing expand, the required return should rise.

Company-specific routes to a return

Company	Primary monetization route	Confirmation signal	Breakpoint
Microsoft	Azure consumption and paid AI products	Capacity converts into cloud growth with stable unit economics	Capex and depreciation rise while Azure demand or margin weakens
Amazon	AWS services, custom chips, and wider company automation	AWS profit and cash generation outgrow the infrastructure bill	Companywide investment remains above cash generation without a clearer split
Alphabet	Cloud revenue plus Search, ads, and model efficiency	Cloud backlog and ad economics absorb depreciation	Infrastructure cost rises faster than Cloud and advertising contribution
Meta	Engagement, ad conversion, and future enterprise products	Revenue and ad yield improve with manageable infrastructure cost	Indirect benefits cannot offset depreciation, leases, and replacement spending

The routes differ, so the same metric should not be forced onto every company. Meta does not need to sell external cloud capacity for AI infrastructure to create value. Amazon should not receive credit for the full companywide capex plan based only on AWS growth.

The next-quarter checklist

Before the next earnings cycle, write down the expected answer to each question:

Is demand still above available supply?
How much new capacity entered service?
Did cloud or product usage accelerate when it arrived?
Did gross margin improve after known one-time items?
Did depreciation grow faster or slower than gross profit?
Did free cash flow improve because operations strengthened or because investment slipped?
Were finance leases and purchase commitments still expanding off the cash-flow statement?

This turns a narrative into a falsifiable thesis.

The decision takeaway

The AI capex bull case is not “spending is large.” It is that large spending creates scarce, highly utilized capacity whose revenue and efficiency gains compound faster than its cost.

The bear case does not require AI demand to disappear. It only requires the newest dollar of infrastructure to earn less than investors assumed.

Use the scorecard to decide which side the evidence supports. Yield Theory members get the next layer: the portfolio implication, the companies with the strongest and weakest return bridge, the catalysts that can move the call, and the written breakpoint that forces a change.

Research cutoff: August 12, 2026. This article is educational and is not personalized investment advice. Forward-looking guidance and company operating metrics can change.