The AI Boom Runs on Unproven ROI

The obvious story is that artificial intelligence is running into shortages of chips, memory, power, and data-center capacity. Those constraints are real. They are also the less interesting part of the business story.

The harder constraint is capital. More precisely, it is the gap between the cost of financing AI infrastructure and the still-unproven cash returns generated by the customers expected to pay for it.

That gap creates a circular business model. Infrastructure providers invest heavily because demand appears greater than supply. AI companies and cloud operators absorb the capacity because future usage is expected to grow. Enterprises experiment with the tools because the technology may eventually improve productivity. Investors finance the entire chain because they assume the future cash flow will justify today’s spending.

Every link depends on the next one producing a return that has not yet been demonstrated at scale.

This is why the comments from Broadcom CEO Hock Tan and Intel CEO Lip-Bu Tan matter. Their central warning was not that AI has no value. It was that revenue growth, infrastructure demand, and technological progress are being treated as proof of a business model before the business model has actually been proven.

The Overlooked Angle

The narrow issue is the financing dependency between AI infrastructure spending and enterprise ROI.

AI infrastructure can continue expanding only if businesses eventually generate enough measurable economic value from using it. That value must be large enough to cover several layers of cost: model access, cloud consumption, software integration, employee training, data preparation, security, compliance, and the financing cost embedded in the infrastructure itself.

Most commentary treats these as separate questions. It discusses whether companies will buy more GPUs, whether data centers can secure electricity, or whether model capabilities will improve. The more consequential question is simpler:

Can customer-level productivity gains arrive quickly enough, and at sufficient scale, to support the debt and capital commitments made upstream?

If the answer is yes, financing accelerates expansion. If the answer is uncertain, every new data center becomes a larger fixed-cost bet placed against an unproven customer revenue stream.

That is the mechanism worth examining. The AI boom is not merely a technology investment cycle. It is a leveraged conversion of expected future productivity into present infrastructure spending.

Why This Small Detail Matters

An infrastructure shortage usually strengthens suppliers. When demand exceeds available capacity, suppliers can raise prices, improve utilization, and secure long-term contracts. That is the attractive part of the current AI market.

But scarcity does not automatically produce healthy economics. It only produces pricing power while customers remain willing and able to pay. If customers are buying capacity to generate real returns, the shortage can support a durable market. If they are buying capacity to avoid being left behind, demand is more fragile than it looks.

This distinction separates productive investment from speculative absorption.

An enterprise may purchase AI capacity for several reasons:

• It has identified a process where automation will reduce labor or operating costs.
• It expects to create a new revenue stream.
• It wants to learn before competitors do.
• Its leadership fears being seen as technologically behind.
• Its existing cloud or software contracts make experimentation easy.

Only the first two reasons directly establish an economic return. The others create demand without necessarily creating profit.

That is the problem hidden by high usage. Tokens consumed, queries processed, and model calls made are activity measures. They are not proof of value. A business can spend heavily on AI while producing no meaningful improvement in throughput, margin, or customer retention.

The arcade analogy used by Hock Tan is commercially precise. If users consume tokens because they are experimenting, the model provider earns revenue. But the enterprise may simply be converting its budget into usage without learning anything that changes the economics of the business.

That revenue can support the next infrastructure investment for a while. It cannot support an entire capital structure indefinitely.

The Economic Mechanism

The basic economics of the AI buildout are often presented as a straightforward supply-and-demand story. Demand for compute exceeds supply, so more capacity must be built. Yet supply expansion introduces a stack of costs that must ultimately be recovered from customers.

At the infrastructure level, the cost base includes:

• Accelerators and networking equipment
• Data-center construction and leasing
• Electricity generation, transmission, and cooling
• Engineering and maintenance staff
• Software and systems integration
• Depreciation on rapidly aging hardware
• Interest on debt used to fund expansion
• The cost of reserving capacity before utilization is certain

The final item is easy to underestimate. A data center is not built only after demand has been fully proven. Capacity must often be committed in advance. Land, power contracts, equipment, and construction schedules require large expenditures before the revenue stream is visible.

That creates a timing mismatch. Costs arrive early and are relatively rigid. Revenue arrives later and may be variable.

A simplified infrastructure business therefore looks like this:

The model works only when utilization, pricing, and customer value move in the right direction together.

Suppose an AI provider signs customers who use a large amount of compute but have no clear economic reason to continue. The provider may report strong revenue growth while absorbing high inference costs. If pricing is kept low to encourage adoption, gross margins remain under pressure. If pricing rises, customers may reduce usage or build alternative systems.

Now add financing costs. When interest rates and credit spreads rise, the required return on every infrastructure project increases. A project that looked acceptable when capital was cheap may no longer produce enough cash to cover interest, depreciation, and operating expenses.

The infrastructure supplier may still sell equipment. The data-center operator may still report demand. The model company may still report revenue. But the combined system can become less profitable as it expands.

This is the central illusion of a capital-intensive technology boom: growth at one layer can conceal deterioration at another.

The supplier sees orders. The customer sees experimentation. The lender sees collateral and contractual commitments. The investor sees future market share. Nobody necessarily sees a mature cash return.

That can persist while financing remains abundant. It becomes unstable when the cost of capital rises or lenders begin demanding evidence rather than promises.

The Strategic Consequence

The immediate beneficiaries are the companies positioned upstream of uncertain customer outcomes. Chip designers, networking suppliers, equipment makers, power providers, and specialized data-center operators can monetize the buildout before enterprise ROI is fully established.

Their business model is comparatively clean. They sell tools, capacity, or construction inputs to buyers who are making the larger strategic gamble.

The pressure falls downstream. AI labs must turn expensive compute into recurring revenue. Cloud providers must keep infrastructure highly utilized. Enterprises must convert pilot projects into measurable operating improvements. Debt-financed operators must maintain cash flow while their assets depreciate and their financing costs fluctuate.

This shifts bargaining power in a specific way. During scarcity, upstream suppliers can capture attractive economics because customers compete for limited capacity. But that power is conditional. It depends on the downstream buyer continuing to believe that additional capacity will produce additional value.

If enterprise ROI remains weak, customers do not necessarily disappear overnight. They become more selective. They shorten commitments, demand usage-based pricing, consolidate vendors, and push providers to absorb more implementation risk.

That changes the economics of the infrastructure market.

Long-term capacity reservations may become harder to secure. Customers may resist paying for idle capacity. Cloud buyers may demand flexibility rather than fixed commitments. Model providers may need to subsidize usage to maintain growth. Data-center projects may be delayed because the expected utilization curve no longer justifies the financing burden.

The result is not necessarily a collapse in AI demand. It is a repricing of demand quality.

There is a major difference between:

• Contracted demand backed by a profitable customer workflow
• Demand backed by a company’s strategic experimentation budget
• Demand backed by a financing structure that assumes future refinancing
• Demand generated by investors competing to fund the next visible growth story

Only the first category provides strong protection against a capital-market reversal.

This is also where scale can become a disadvantage. Large infrastructure commitments create operating leverage, but they also create fixed obligations. If demand accelerates, scale lowers unit costs. If demand disappoints, scale magnifies the damage because expensive assets remain in place while utilization falls.

The same capacity that creates an advantage during a shortage becomes a liability when customers stop paying for speculative excess.

What Most Commentary Gets Wrong

The lazy interpretation is that skepticism about AI ROI means skepticism about the technology itself. That is a convenient argument because it turns an economic question into a tribal one.

The real issue is not whether AI can be useful. It clearly can. The issue is whether usefulness arrives with enough measurable value, quickly enough, to justify the cost structure built around it.

Another weak interpretation is that strong revenue growth settles the question. It does not. Revenue is only the top line. In a capital-intensive business, the relevant test is cash generation after compute costs, infrastructure spending, employee costs, depreciation, interest, and customer acquisition.

A company can produce enormous revenue while destroying value if every additional dollar of sales requires more than a dollar of total economic cost.

The same mistake appears in infrastructure forecasts. Analysts often treat current capacity shortages as evidence that demand will remain strong for years. But a shortage proves only that current supply is insufficient for current orders. It does not prove that future customers will earn enough from the capacity to keep buying it.

Shortages can be resolved through supply expansion. Demand can also weaken when prices rise, alternatives improve, or customers discover that their use cases are less valuable than expected.

The third mistake is treating bottleneck removal as unambiguously bullish. Memory shortages, wafer constraints, and power limitations restrict supply. Solving them may enable more deployment, but it also removes scarcity pricing and exposes the market to the quality of underlying demand.

When supply is constrained, almost any customer appears valuable. When supply expands, providers must distinguish between customers who generate returns and customers who merely consume budgets.

Finally, commentary often assumes that technological improvement automatically creates economic improvement. A more capable model may produce better outputs, but the value to an enterprise depends on adoption, workflow redesign, error management, integration costs, and accountability. Technical progress does not bypass operational friction.

A model that performs better in a benchmark may still fail to reduce costs in a real department. An AI assistant that saves time may not reduce headcount, increase output, or improve customer retention. The distance between technical capability and financial return is where many investment cases quietly weaken.

The Hard Business Lesson

The AI infrastructure cycle will not be validated by the number of chips shipped, data centers announced, or tokens consumed. It will be validated when customers can show that AI produces returns greater than the full cost of deploying it.

That full cost includes more than a software subscription. It includes infrastructure, financing, integration, oversight, security, process redesign, and the opportunity cost of management attention.

The critical question for every participant is therefore not whether demand exists. It is whether demand survives financial discipline.

For infrastructure suppliers, the lesson is to distinguish backlog from durable customer economics. A customer can sign a contract and still be a weak credit or an unprofitable source of demand.

For AI labs, the lesson is harsher. Usage growth is not enough. The product must move from experimentation to a workflow that customers cannot economically abandon.

For enterprises, the correct discipline is to measure AI projects against avoided cost, incremental revenue, error reduction, cycle time, and cash impact. A pilot that produces enthusiasm but no operating change is not an investment. It is a technology demonstration paid for by the finance department.

For investors and lenders, the question is whether projected cash flow can withstand higher capital costs, lower pricing, slower adoption, and faster hardware obsolescence. If the answer depends on all assumptions going right, the asset is not merely growth-sensitive. It is structurally fragile.

The quiet part is not that AI has no future. The quiet part is that the industry has built present-day obligations against a future that still needs to prove its economics.

That is why financing may become the decisive bottleneck. Chips can be manufactured, power can be added, and data centers can be constructed. Capital, however, eventually demands evidence.

When it does, the winners will not be the companies with the loudest forecasts. They will be the companies whose customers can prove that every additional unit of AI capacity creates more cash than it consumes.