Opening

The obvious story is that generative AI faces political resistance, higher interest rates, and an increasingly skeptical market. Those risks matter. They are also not the most important risk hiding in plain sight.

The more dangerous problem is buried in the cost of serving the product. Investors have treated AI companies as if they can eventually achieve familiar software economics: high gross margins, low incremental costs, and strong operating leverage once the customer base scales. That assumption is fragile because generative AI does not merely sell software. It repeatedly rents expensive computing capacity every time a user asks a question, creates an image, summarizes a document, or triggers an automated workflow.

This changes the margin equation. The product is not just the model, interface, or subscription. The product includes an ongoing computation bill attached to every unit of customer behavior.

If usage grows faster than pricing, the company can become more popular while becoming less profitable. That is not a temporary inconvenience. It is a structural threat to the valuation logic investors have applied to AI.

The Overlooked Angle

The overlooked angle is the gap between AI revenue growth and inference-cost growth.

Training costs attract the headlines because they involve large data centers, advanced chips, and capital spending. But training is a periodic investment. Inference is the recurring cost of operating the service after the customer arrives. It is incurred across every interaction and expands with usage.

That distinction is commercially decisive.

A traditional software company can add another user at very low marginal cost. The user may consume storage, bandwidth, and support, but the core software has already been built. A generative AI company must allocate meaningful computing resources every time the user consumes the core product. The more useful the product becomes, the more frequently customers may use it. Increased engagement therefore creates a cost obligation rather than simply improving retention.

This creates a narrow but serious investment question: can AI companies convert rising model usage into higher contribution margin, or does usage merely pass through the income statement as revenue matched by a similar increase in compute expense?

The answer depends on pricing power, model efficiency, customer behavior, and control over infrastructure. Those are operational questions, not slogans about technological leadership.

Why This Small Detail Matters

The market has been willing to value AI businesses on the assumption that scale will solve the economics. That is a reasonable assumption for many digital products. It is not automatically valid for computationally intensive services.

Scale helps only when the cost per unit falls faster than the price per unit, or when pricing can rise faster than usage. If neither happens, scale magnifies exposure.

Consider the basic relationship:

The dangerous outcome is not that AI demand disappears. It is that demand grows in a form that is economically inferior to what investors expected.

A customer who uses an AI assistant twice a month is not equivalent to a customer who embeds that assistant into daily operations. The second customer may produce more revenue, but also far more inference demand. If the subscription price is fixed, the supplier carries the usage risk. If the price is variable, the customer may reduce usage or switch to a cheaper provider.

This is the central tension. AI companies want adoption to be frictionless, so they often package usage inside simple subscriptions or offer generous access. But simple pricing hides the marginal cost from the customer while leaving it visible to the provider.

That is how a growth metric becomes a margin problem.

The Economic Mechanism

The economics begin with a simple distinction between fixed and variable costs.

Model development, data acquisition, and infrastructure construction contain substantial fixed or semi-fixed costs. These costs are painful, but they can be spread across a large revenue base if demand materializes. Inference is different. It is closer to a consumption cost. Each additional request requires processing capacity, energy, networking, storage, and often additional orchestration around the model.

The relevant unit is not merely a customer. It is a completed task, generated token, processed document, image, query, or automated action.

A subscription model can conceal that unit economics are deteriorating. Suppose a provider charges a flat fee while the customer gradually increases usage. Revenue remains stable, but the cost of serving the account rises. The provider may report strong user growth and retention while contribution margin quietly weakens.

The basic account-level equation is:

Contribution margin = customer revenue minus inference cost minus directly related service cost.

The problem is that the denominator is behavioral. It depends on how customers use the product, how long their prompts and documents are, how much context the system must retain, how often the model is called, and whether the workflow requires one response or a chain of model calls.

That makes forecasting harder than it is for conventional software. A sales team can sell a large enterprise contract, but the contract does not reveal the eventual cost unless the provider understands the customer workflow in detail.

Three forces make the problem worse.

Usage expands faster than pricing

AI vendors often market productivity gains rather than compute consumption. Customers buy an outcome, such as faster analysis or automated support. If the product succeeds, they attempt to apply it to more tasks. The provider then processes more requests without necessarily receiving more revenue.

Usage-based pricing can solve this problem, but it introduces sales friction. Customers dislike unpredictable bills, especially when they are still testing whether the technology delivers reliable value. Flat pricing accelerates adoption but transfers demand volatility to the vendor.

There is no free pricing model. The vendor either absorbs the variability or asks the customer to do so.

Better models can increase cost

Technological progress does not guarantee lower total cost. A more capable model may require more computation per request. It may also encourage customers to use AI for higher-value and more complex workflows, which require longer context, multiple steps, validation, and integration with external systems.

Efficiency improvements can reduce cost per request, but those savings may be consumed by richer usage. This is a familiar operational trap: productivity gains lower the cost of each action, then demand expands until the total cost base remains high.

The relevant measure is not whether inference becomes cheaper in isolation. It is whether total inference expense grows more slowly than monetized usage.

The supplier captures part of the value

AI companies may not control the full stack. They can depend on chip suppliers, cloud platforms, data center operators, energy providers, and model infrastructure vendors. Each supplier occupies a position in the cost chain.

If infrastructure capacity is scarce, the provider cannot assume that volume discounts will arrive quickly enough to protect margins. A company can have strong customer demand and still face weak bargaining power if its service depends on constrained computing resources.

This is where the capital intensity of AI becomes relevant. A software company typically owns intellectual property and distributes it cheaply. An AI company may need continuous access to expensive physical infrastructure simply to keep the service running. That creates a recurring dependency that investors cannot dismiss as a one-time buildout.

Price competition passes efficiency to customers

Even if inference costs fall, competition can prevent vendors from retaining the benefit. Rivals can use lower costs to cut prices, increase free access, or bundle AI into broader products. Customers then receive the efficiency gain while providers fight over distribution.

This is particularly dangerous when model capabilities become broadly available. If several providers can offer acceptable performance, the market may price the service near operating cost. Technical superiority creates a temporary advantage, not necessarily durable pricing power.

A cost reduction is strategically valuable only when the company can keep enough of it.

The Strategic Consequence

The immediate winners are companies that control scarce infrastructure, own a large distribution channel, or can monetize AI without carrying unlimited usage risk.

Infrastructure owners can charge across the entire ecosystem. Distribution owners can place AI inside products customers already pay for, reducing acquisition costs and making the feature part of a larger economic bundle. Companies with proprietary workflows can charge for a business outcome rather than for raw access to a model.

The vulnerable group is the standalone AI application with limited differentiation and flat pricing.

Such a company faces a difficult choice. It can cap usage, which harms adoption and customer satisfaction. It can raise prices, which invites substitution. It can move to usage-based billing, which makes demand less predictable. Or it can absorb the cost and hope that future efficiency improvements repair the margin.

Hope is not a pricing strategy.

The strongest AI businesses will therefore be defined less by the ability to generate impressive demonstrations and more by the ability to control the relationship between usage and cost. That control can come from several sources:

• A workflow that produces measurable financial value for the customer
• Pricing tied to outcomes or consumption rather than vague access
• Smaller or specialized models that meet the task requirement at lower cost
• Infrastructure ownership or negotiated capacity advantages
• Distribution that lowers customer acquisition expense
• Product design that limits wasteful or redundant model calls

The weakest companies will confuse customer activity with economic traction. High request volume looks impressive in a presentation. It means little if each request carries thin or negative contribution margin.

This also changes the competitive landscape. The market may not consolidate around the company with the best model. It may consolidate around the company with the best cost-to-value architecture.

That is a less glamorous contest. It is also the one that determines who survives.

What Most Commentary Gets Wrong

Most commentary treats the AI risk as a question of whether the technology works. That is too shallow. The technology can work extremely well and still produce poor returns for investors.

The second lazy interpretation is that lower computing costs automatically solve the problem. They do not. Lower unit cost helps only if pricing remains stable, usage does not expand faster than efficiency, and suppliers do not capture the savings. None of those conditions is guaranteed.

The third mistake is to treat political resistance and higher interest rates as the main causes of market disappointment. Those forces expose weak economics, but they do not create them. Higher rates make long-duration growth claims less valuable. Political resistance may slow deployment. Neither explains why a successful AI product could have disappointing margins. That issue sits inside the operating model.

The fourth mistake is to compare AI directly with earlier software businesses. The comparison works at the interface level and fails at the cost level. Software can distribute an existing capability repeatedly at low incremental cost. Generative AI performs fresh computation repeatedly. The customer experience may look like software, but the cost structure behaves more like a managed utility with a volatile demand curve.

Finally, many investors focus on revenue growth before asking what kind of revenue is being purchased. A dollar of revenue generated by a low-usage enterprise account is not economically identical to a dollar generated by a customer whose requests consume substantial model capacity. The gross margin profile depends on the behavior behind the revenue line.

The correct diligence questions are operationally boring:

• What is the cost of serving the average account?
• How does that cost change as usage increases?
• Which customers are profitable after inference and support costs?
• How much pricing power exists when competitors offer similar capability?
• Who owns the infrastructure bottleneck?
• Can the company limit unprofitable usage without damaging the product?

Those questions are more useful than another comparison of benchmark scores.

The Hard Business Lesson

The hard lesson is that AI demand is not automatically valuable demand.

Investors have been trained to see engagement as an unqualified asset. In conventional digital businesses, more usage often improves retention while adding little cost. In generative AI, more usage can increase both retention and the bill required to support it.

That makes the business model unusually sensitive to pricing design, customer behavior, and infrastructure control. A company can win the product battle and lose the margin battle.

The real investment test is not whether AI becomes ubiquitous. It probably will in some form. The test is whether the provider can turn ubiquity into retained economic value rather than passing the value through to customers, cloud platforms, chip suppliers, and energy providers.

The market will eventually stop rewarding activity for its own sake. When it does, the winners will not be the companies with the loudest adoption claims. They will be the companies that can prove one additional customer, one additional workflow, and one additional unit of usage improves the economics instead of merely expanding the cost base.

That is the risk investors have overlooked. AI may be unstoppable as a technology while remaining disappointingly difficult as a business.