Amazon’s AI CAPEX Debate Is Using the Wrong Unit of Analysis

| CORPORATE ANATOMY SERIES |

Yoshinori Maeda (Astrocyte LLC / Business Strategist & MBA)

Expertise: Corporate Strategy, AI Implementation, Business Model Mutation, and Financial Analytics for C-Suite Executives

The Board Should Underwrite Reversibility, Regulatory Exposure and Asset-Level Economics—not Aggregate Spending

Amazon’s second quarter produced an unusual combination: extraordinary operating performance and negative free cash flow.

Revenue reached $200.6 billion. Operating income rose to $27.5 billion. AWS grew 37% to $42.2 billion and produced $16.6 billion of operating income, a 39.4% operating margin. At the same time, trailing-twelve-month operating cash flow increased 33% to $161.4 billion while free cash flow fell to negative $7.6 billion, primarily because of higher property and equipment investment associated with AI.

The obvious debate is whether Amazon is investing too aggressively.

That is probably the wrong unit of analysis.

The public numbers can tell us how much financial capacity Amazon is committing. They cannot tell us whether a particular power agreement, data-center site, cooling design, network build or accelerator deployment will earn an acceptable return.

That distinction matters.

The real capital-allocation problem is not how much Amazon is spending on AI.

It is which commitments deserve to become irreversible before the future architecture of AI is known.

And that requires a different framework.

Public Data Can Establish the Risk—but Not Underwrite the Assets

There is enough public evidence to establish three things.

First, AWS currently has exceptional economics. It accounts for roughly one-fifth of Amazon’s revenue but about 60% of consolidated operating income in Q2.

Second, Amazon is converting a growing amount of operating cash flow into AI-related infrastructure rather than allowing it to appear as near-term free cash flow.

Third, the technology and regulatory environment against which those assets will earn their returns is not fixed.

But outsiders cannot observe many of the variables that would actually determine whether an individual investment is attractive.

We do not know, project by project, Amazon’s power acquisition costs, regional capacity constraints, cancellation rights, utilization assumptions, customer concentration, expected accelerator replacement cycles, conversion costs, residual asset values or hurdle rates.

That means an external analyst cannot credibly conclude that Amazon is either overinvesting or underinvesting in aggregate.

But this does not leave us with nothing to say.

It tells us what the Board should demand before approving the next tranche of capital.

Instead of one AI CAPEX budget, Amazon should effectively be underwriting a portfolio of assets against four different tests.

Test 1: Does the Asset Still Earn Its Return in a More Open Market?

The previous version of the AI infrastructure argument was too simple: if Amazon builds a lower-cost, higher-performance and more available infrastructure platform, regulation should be less threatening.

That is directionally useful, but incomplete.

The European Commission has reached a preliminary view that AWS and Azure should be designated as gatekeepers under the Digital Markets Act. No final AWS designation has yet been made.

The significance is broader than interoperability alone.

The DMA framework addresses issues including self-preferencing, access to data, interoperability and the ability of business users and consumers to exercise meaningful choice. EU enforcement experience also shows that technical integration, privileged treatment and data advantages can matter alongside simple portability.

So the Board’s regulatory stress test should not merely ask:

Would this investment still work if customers could switch clouds more easily?

It should ask:

Would the economics still work if Amazon could not rely on preferential placement, privileged integration, discriminatory capacity allocation, exclusive access to relevant data or commercial arrangements that advantage affiliated services?

That is a harder test.

It separates two sources of return.

One is productive advantage: lower cost, greater availability, better performance, faster deployment and more reliable infrastructure.

The other is ecosystem advantage: returns that may partly depend on how tightly infrastructure, models, data, distribution and commercial relationships are connected.

Both can be economically powerful.

But they do not carry the same regulatory durability.

The FTC’s existing monopolization case against Amazon is focused on its e-commerce conduct, not evidence of unlawful AWS conduct. It should therefore not be treated as proof of a cloud antitrust problem. It does, however, reinforce why Amazon’s Board should be careful about strategies whose economics depend heavily on exclusionary conduct rather than superior customer economics.

For capital approval, the relevant question is therefore not whether the asset creates a moat.

It is:

How much of the projected return survives if the moat has to remain open?

Test 2: Is the Asset Reconfigurable After Conversion Costs?

A second assumption also needs tightening.

Land, power and network access are often described as having high option value because they can support multiple generations of computing.

Sometimes they will.

But scarcity alone does not create option value.

A data-center site can be scarce and still become economically unattractive if the next computing architecture requires a materially different power density, cooling system, network topology, latency profile or physical configuration.

A long-term power commitment can be valuable while the infrastructure attached to it becomes poorly located for future workloads.

Cooling infrastructure can remain physically functional while becoming economically inferior to the configuration required by the next generation of accelerators.

The right concept is therefore not simply reusability.

It is conversion-adjusted residual value.

For every long-duration asset, the Board should ask:

> What would this asset be worth in its next-best use, after subtracting the capital required to convert it, the downtime involved, and the value lost while that conversion occurs?

That changes the hierarchy of apparently “safe” AI infrastructure.

Power, land and fiber should not automatically receive long-duration capital because they appear technology-agnostic.

They deserve long-duration capital only when their economics survive several materially different workload configurations.

Likewise, accelerators should not automatically be treated as poor long-term assets. If supply is exceptionally constrained, utilization is high and payback occurs before technological depreciation becomes material, a short-lived asset can still be an excellent investment.

Asset life and strategic value are not the same thing.

What matters is whether the expected economic return arrives before the asset’s relevant advantage disappears.

Test 3: Is Demand Independent Enough to Justify the Commitment?

Strong demand does not by itself establish investment quality.

Amazon has disclosed rapid AWS growth, significant expansion in its AI and chips businesses, and major commitments involving leading AI labs.

But a Board underwriting infrastructure should distinguish between different qualities of demand.

Demand backed by diversified enterprise workloads is different from demand concentrated in a handful of frontier-model providers.

Committed revenue with strong take-or-pay economics is different from a commercial relationship with material cancellation, repricing or technology-performance contingencies.

Demand for training is different from demand for inference.

Demand tied to one accelerator is different from demand portable across multiple architectures.

Public investors cannot make these distinctions from Amazon’s aggregate disclosures.

The Board can.

For each large infrastructure commitment, it should therefore see at least three demand cases:

Base demand: the current expected mix of customers and workloads.

Partner-loss demand: economics if a major model partner materially reduces usage or shifts architecture.

Architecture-shift demand: economics if the mix changes significantly between training, centralized inference, efficient models and edge processing.

This is where the central AI-efficiency debate becomes more nuanced.

More efficient models do not necessarily mean lower total infrastructure demand.

Lower costs may stimulate far more usage, applications and automated workloads. Total demand could continue rising even while compute required per unit of capability falls.

The capital risk is therefore not simply AI demand collapses.

It is that **the composition of demand changes faster than the asset can change with it**.

Test 4: Can Capital Move Before Sunk Cost Becomes Strategy?

The final test is organizational rather than technological.

Large infrastructure programs create their own momentum.

Once management has secured land, contracted power, begun construction, ordered equipment and committed to customers, the argument for finishing the program becomes progressively easier to make.

That creates a classic danger: a strategy that began as forward-looking capital allocation gradually becomes justified by money already spent.

The Board therefore needs reallocation rules before the downside case occurs.

Not simply:

When do we stop investing?

But:

What evidence changes where the next dollar goes?

A decline in accelerator economics might shift capital toward networking.

A change in power density might alter site design.

A move from training toward inference might change hardware requirements.

A regulatory constraint on preferential integration might change the economics of a vertically connected model-and-cloud strategy.

A major customer concentration threshold might move capacity from dedicated builds toward more fungible infrastructure.

These should be treated as capital-reallocation decisions, not admissions that the AI thesis was wrong.

That distinction is important because technological transitions rarely invalidate an entire market at once.

They change which layer captures the economics.

Commit, Stage or Rent

These four tests lead to a more useful Board framework than a single debate over annual CAPEX.

Every material AI infrastructure commitment should fall into one of three capital postures.

・Commit

Make the long-duration commitment when the asset remains scarce across multiple plausible futures, produces acceptable economics without relying on regulatory closure, serves sufficiently diversified demand and retains meaningful conversion-adjusted residual value.

This is where irreversible capital can become an advantage.

・ Stage

Build in phases when the underlying capacity is strategically valuable but the future configuration remains uncertain.

Secure what is genuinely scarce while delaying architecture-specific decisions.

Use modularity, milestone-based expansion and contractual flexibility to purchase information before committing the full amount.

This is where option value should be created deliberately rather than assumed.

・Rent

Use shorter-duration capacity, leases, external supply or rapidly replaceable assets where technological depreciation is fast and ownership does not create sufficient strategic benefit.

Paying a higher unit cost can be rational when it prevents the company from owning the wrong generation of capacity.

This is where flexibility can be worth more than nominal cost efficiency.

The Board Question Is Smaller—and Harder—Than the CAPEX Number

Amazon’s Q2 results establish that the company has extraordinary operating capacity and is deploying a substantial portion of it into AI infrastructure.

They do not establish whether Amazon is spending too much.

Nor can public information identify which individual investments should be approved.

That conclusion is not a weakness in the analysis. It defines the proper boundary between external analysis and internal capital allocation.

What public evidence does allow us to conclude is that aggregate AI CAPEX is too coarse a variable for the decision Amazon now faces.

A dollar committed to durable power access is not economically equivalent to a dollar committed to a particular accelerator generation.

A scarce site with expensive conversion requirements is not necessarily more strategic than a less scarce but highly adaptable one.

A customer relationship that fills capacity today is not necessarily evidence of durable demand.

And an infrastructure advantage whose return depends heavily on privileged ecosystem control should not be valued as though regulation cannot change that control.

The Board therefore does not need one answer to the question, Are we spending too much?

It needs hundreds of smaller answers about what each major commitment is buying.

For every material asset, the decision should be explicit:

Commit. Stage. Or rent.

And the standard should be demanding:

Would we still approve this investment if the winning model were different, the workload mix changed, customers had more freedom to move, and the asset had to earn its return without privileged ecosystem treatment?

If the answer is still yes, Amazon may be buying strategic capacity.

If the answer depends on today’s architecture, today’s regulatory assumptions and today’s partner structure all remaining intact, it may simply be buying a very expensive version of the present.