Why AI Economics Fail: Cost Structures, Billing Models, and Stalled Adoption

Evidence from Public Disclosures and Earnings Calls

Abstract

Despite rapid advances in model capability, large-scale adoption of generative artificial intelligence has stalled. This paper argues that the constraint is not technical performance but economic viability and user trust. On the supply side, current AI business models rely on pricing proxies—tokens and flat-rate subscriptions—that fail to reflect underlying compute costs or user value, leading to uncontrolled operating expenses and weak revenue alignment. On the demand side, adoption has plateaued because users do not trust systems that lack reliable memory and consistent, user-controlled behavior. Using publicly available disclosures, earnings call statements, and observable usage patterns, this paper presents a bounded, testable claim: absent fundamental changes to cost provisioning, billing, memory persistence, and governance, AI systems cannot scale profitably or achieve durable enterprise adoption. The paper concludes by outlining observable implications and explicitly inviting rebuttal.

1. Introduction and Claim

Before artificial intelligence can transform industries, it must function as a business.

This paper advances a narrow, testable claim:

Current AI economics fail because costs are uncontrolled, revenues are mispriced, and adoption has stalled due to lack of trust.

The claim does not rest on speculative future capabilities or moral arguments. It rests on observable economic mechanics:

Cost structures that scale with usage rather than declining with volume
Billing models that approximate value instead of measuring it
User behavior that treats AI primarily as enhanced search, not as a trusted collaborator

The result is a system that may impress technologically while failing commercially.

2. The Economic Failure: Cost and Revenue

2.1 Cost Structures: Execute First, Discover Cost Later

Most large-language-model (LLM) systems operate under an “execute first, reconcile cost later” model. Compute is allocated dynamically during inference, with limited ex-ante controls over how much computation a given task should consume.

This guarantees waste.

Inference cost is driven by:

Model size
Context length
Iterative reasoning loops
Re-execution due to hallucinations or incomplete outputs

Because cost is not provisioned before execution, systems routinely over-allocate compute relative to task value. This problem compounds as usage grows. Unlike traditional software, marginal cost does not asymptotically approach zero.

Public disclosures reflect this dynamic indirectly through:

Rising capital expenditures on AI infrastructure
Increasing operating expenses tied to inference and energy
Limited transparency on unit economics per user or per task

Absent structural cost controls, scale magnifies losses rather than efficiency.

2.2 Revenue Models: Pricing Proxies Instead of Measurement

On the revenue side, most AI providers rely on two models:

Flat-rate subscriptions
Token-based pricing

Both are proxies.

Tokens do not represent value. They approximate text length, not computational intensity or user outcome. Flat subscriptions decouple price from usage entirely, encouraging overconsumption while obscuring true cost.

The economic consequence is predictable:

Heavy users are subsidized by light users
Providers cannot align revenue with actual resource consumption
Margins degrade as adoption increases

This is not a pricing optimization problem. It is a measurement problem. Without metering actual compute usage, AI cannot be priced rationally.

3. Adoption Failure: Trust and User Behavior

Even if the economics were repaired, adoption would still face a second constraint: trust.

3.1 Stalled Adoption Is Observable

Public statements and third-party usage data indicate that AI adoption rates flattened in 2025 at both consumer and enterprise levels. While trial usage remains high, durable integration into workflows has lagged.

This distinction matters.

Using AI occasionally does not constitute adoption. Adoption implies:

Repeated use
Reliance without constant verification
Accumulating confidence over time

These conditions are largely absent.

3.2 Memory Loss Undermines Trust

Current LLMs do not remember reliably.

They forget past interactions, lose context across sessions, and reconstruct responses from incomplete internal representations. Retrieval-augmented generation (RAG) attempts to compensate through lossy compression and retrieval heuristics.

The pattern is familiar.

JPEG compression allowed images to traverse slow networks by discarding information. At low resolution, the loss was tolerable. At scale, artifacts became visible.

LLMs exhibit the same dynamic:

Missing facts
Confident but incorrect assertions
Inconsistent reasoning across sessions

What JPEGs discard are pixels.
What LLMs discard are facts.

Users adapt accordingly. They verify outputs, restate context, and treat systems as disposable tools rather than dependable collaborators. Trust never compounds.

Without loss-less memory, trust cannot accumulate. Without trust, adoption plateaus.

3.3 Governance and Personalization Are Missing

In addition to memory, users lack durable control over system behavior.

Preferences for tone, verbosity, analytical style, or formatting are rarely persistent. Each interaction effectively resets the system. This increases cognitive overhead and discourages long-term use.

Governance here does not mean removing platform-level safety controls. It means allowing users to define how the system behaves within those boundaries.

Tools that cannot be tuned are not adopted deeply. They are sampled and abandoned.

4. Why Capability Gains Do Not Solve These Problems

A common rebuttal is that newer models are “better” and therefore adoption will resume.

Capability improvements do not resolve:

Cost misalignment
Pricing proxies
Memory loss
Lack of user control

Indeed, larger models often exacerbate cost issues by increasing inference expense without proportional gains in trust or usability.

This explains an apparent paradox: models improve, but business outcomes stagnate.

5. Public Evidence and Testability

This paper relies exclusively on public evidence:

Earnings call statements acknowledging uneven or limited enterprise usage
Disclosures showing rising AI-related capital and operating expenses
Observed user behavior treating AI primarily as enhanced search
Data from five independent labs showing adoption stalled in Spring 2025

The claim is testable in multiple ways:

If the claim is false, we should observe:

Declining unit inference costs per user without structural changes
Revenue models converging on sustainable margins
Enterprise and consumer adoption accelerating without improvements in memory or governance

If the claim is true, we should observe:

Continued cost pressure despite scale
Persistent experimentation without durable adoption
Increasing emphasis on infrastructure sales rather than end-user profitability

6. Implications

If AI economics remain unchanged:

Subscription and advertising models will fail to produce durable profits
Infrastructure investment will outpace monetization
Market valuations tied to AI growth will face downward pressure

If AI economics change structurally:

Cost provisioning and revenue metering could align incentives
Trust could accumulate through reliable memory
Adoption could shift from experimentation to integration

The distinction is not between optimism and pessimism. It is between structure and narrative.

7. Invitation to Rebuttal

This paper advances a single bounded claim grounded in observable economics.

The author explicitly invites rebuttal, correction, or counter-evidence based on:

Disclosed unit economics
Demonstrated large-scale profitable adoption
Empirical evidence of trust accumulation without memory persistence