The Model Is Not the Asset

AI is making intelligence cheaper. A model can lead for three months, attract millions of users, and lose much of its advantage once similar capabilities become widely available. This essay asks what still compounds when intelligence itself is no longer scarce.

Part I · Technical Leads Fade

Frontier-model leads rarely last. Benchmark gaps close, inference gets cheaper, and once-scarce capabilities become available through competing APIs or open weights.

The same happens in applications. Perplexity and Cursor helped define new ways to search and code, but similar features soon appeared across many products. In AI, being early creates a window; it does not mean the company will capture the lasting value.

Solar-panel manufacturing is a useful comparison for the LLM industry. Demand grew rapidly and the technology created enormous social value, but falling prices, weak differentiation, and heavy capital needs made it difficult for many manufacturers to retain the value they created. Frontier models could face a similar economic pattern as model quality converges and intelligence becomes cheaper. A technology can transform the world without its suppliers becoming good assets.

DeepSeek illustrates the tension. Open weights can accelerate adoption, but they also allow users to deploy and adapt the model independently. Broad use therefore does not automatically create a durable asset for DeepSeek. The question is whether that adoption ultimately strengthens something the company itself can retain and monetize after the model lead narrows.

If a model lead is temporary, durable value must come from what remains after it fades. That includes the research and engineering system that produces the next model, as well as knowledge gained from real use: where the model fails and how experts correct it.

The Weights Are Not the Whole System

Two companies can run the same model weights at very different costs and levels of reliability. The harder-to-copy advantage may lie in quantization, compilers, kernels, inference serving, cluster scheduling, and hardware utilization. If the same toolchain and operating knowledge make each new model cheaper and faster to deploy, the production system, rather than the static weights, can compound.

Model companies may reinforce that system in both directions. Partnerships with chip, cloud, or data-center providers can secure reliable access to compute and lower costs, while moving downstream into applications can capture usage and feedback. These moves create a moat only if they produce a lasting advantage in cost, access to compute, or proprietary data. Otherwise, they simply add capital and complexity.

Part II · Value Moves to What Remains Scarce

Cheaper intelligence does not determine who captures the value it creates. Value accrues to whoever owns the customer relationship, becomes part of the workflow, and learns from real-world outcomes.

These advantages do not automatically belong to the model provider. They may remain with the enterprise that deploys the model, the application through which users access it, or the managed provider that keeps the resulting usage data and feedback.

Who Captures the Value?

DeepSeek and Anthropic distribute model capability in different ways. DeepSeek gives customers greater control through open weights and self-deployment. This can spread adoption, but much of the usage data and resulting value may remain with the customer.

Anthropic keeps the model behind an API and sells managed reliability and accountability. This allows it to stay closer to customer usage and feedback, but only if customers are willing to pay for the service.

Training Data Must Be Continually Renewed

Training data has an unusual life cycle. Once a set of expert answers teaches a model a task, the marginal value of the same data falls.

Consider Mercor. It began as a recruiting marketplace, and its early advantage was not a proprietary dataset but a system for finding, screening, and matching skilled people. As frontier labs began buying expert reasoning and evaluation data, the same system became useful for identifying the right experts and organizing their judgment into training tasks.

A business like Mercor cannot rely on one collection of expert work. Its potential asset is not the last dataset, but the system that finds new capability gaps, recruits the right experts, and turns their judgment into tasks, rubrics, and verifiers. Verification therefore does more than catch errors. It shows what the model still needs to learn.

Verification Becomes the Bottleneck

AI lowers the cost of producing answers faster than the cost of knowing whether they are right. As output grows, companies need systems that define success, trace sources, test results, and assign responsibility when things fail. Answers do not compound; a system that learns from errors can.

For enterprise customers, correctness alone is not enough. They also need to know what data entered the system, how an output was produced, and who approved it. The relevant metric is not token price, but the total cost of a reliable outcome.

Not every task, however, produces feedback at the same speed.

Automation Depends on Fast, Reliable Feedback

Automation works best when success and failure can be observed quickly. In coding, compilers and unit tests provide immediate feedback.

In investing, medicine, or contract design, outcomes arrive late and can be misleading. A profitable investment may still result from poor reasoning, so experts must also evaluate the process: whether the evidence was sound, the risks were considered, and the decision was well reasoned.

These reviews create positive and negative examples before the final outcome is known. The next challenge is giving experts a reason to keep producing them.

Reusable Judgment Requires Human Incentives

Industrialization made physical skills reusable through machines. AI may do something similar with parts of human judgment: experts' edits, exceptions, review comments, and scoring rules can be recorded and used again.

But recording expert feedback is easier than persuading experts to keep contributing their best judgment. Contributors need to be recognized for improving the system, understand how their work is being used, and share in the value they help create. Otherwise, they may provide only superficial feedback or withhold the tacit knowledge the system needs most.

Rewards should reflect verified improvement, not simply the volume of feedback produced. The asset is not the dataset alone, but a system that can repeatedly elicit, verify, recognize, reward, and lawfully reuse high-quality judgment.

Part III · How Investors Can Analyze AI Transformation

What Exists, What Can Improve, and What Might Grow

A company can be analyzed in three layers. The core asset is the business that exists today and would remain valuable even if its AI initiatives failed. Execution is the use of AI to improve that business by lowering costs, increasing speed, or improving quality. Growth is the additional demand or revenue created by new products or use cases that customers adopt and pay for.

Each layer requires different evidence. Core assets need existing customers, cash flow, and a defensible competitive position. Execution needs measured improvements in cost, quality, or margins. Growth needs adoption, retention, pricing, and contribution profit.

Take Tripadvisor as an example. Its brand, traffic, review data, and Viator are core assets; the agreement to sell TheFork crystallizes value from one asset but does not improve ongoing operations. Execution means improving Viator's margins and lowering Brand Tripadvisor's cost base, with AI counting only when it measurably reduces costs or improves conversion. AI travel planning, data licensing, agentic search, and commerce belong to growth. Until these products generate adoption, retention, and contribution profit, they remain optionality.

AI strategy is not a fourth layer. It must strengthen the core asset, measurably improve execution, or create a new product that customers adopt and pay for.

What Still Compounds?

Models will become cheaper and easier to replace. For investors, the question is who captures the value they create, and what remains when the underlying model changes.

A model lead creates a window. The asset is what still compounds after it closes.