The AI Bubble Test Is Wrong: The Real Moat Is a Living System

What if the most important question about the AI boom is not whether the machines are being used, but whether the system around them can keep learning?

During the telecom bubble, companies celebrated the miles of fiber they had buried. The physical infrastructure existed, but most of it was dark. At the peak, roughly 97 percent of the fiber laid in America was unused. The story was not that the internet was imaginary. The story was that capital had arrived faster than demand, software, customers, and economics could absorb it.

Today, the opposite signal is visible in much of the AI infrastructure stack. GPUs are heavily utilized. Data centers are being built for immediate demand. Major buyers are generating enormous cash flows, and several have seen returns on invested capital improve after increasing their spending. The machines are not waiting in a warehouse for a future that may never arrive. They are running hot.

That distinction matters, but it is not enough. Utilization can tell us whether a machine is busy. It cannot tell us whether the machine is becoming more valuable, whether its output will earn a return, or whether the system supporting it can adapt when the technology changes.

The deeper question is this: What separates productive technological overinvestment from wasteful technological overinvestment?

The answer is not simply demand. It is the presence of a learning system, one that converts infrastructure into better products, better products into more users, users into better data, and better data into more valuable infrastructure.

The durable AI advantage will not belong to whoever owns the most compute. It will belong to whoever turns compute into a self improving system fastest.

From dark fiber to living infrastructure

A useful way to understand the current moment is to distinguish between static infrastructure and living infrastructure.

Static infrastructure is built on a forecast. Fiber is laid, towers are erected, or factories are constructed in anticipation of future demand. If the forecast is wrong, the assets sit idle. They depreciate while their owners wait for the future to catch up.

Living infrastructure is different. It is not merely an asset. It is part of a feedback loop. Its use produces information that improves the next version of the product, which increases usage, which produces more information. The infrastructure can therefore become more productive even before demand expands dramatically.

This is why the comparison between dark fiber and modern AI compute is revealing but incomplete. The absence of dark GPUs is evidence that demand is real. It is not proof that every dollar of AI spending will earn an attractive return. A fully occupied factory can still manufacture products nobody wants. The crucial issue is whether utilization is connected to learning and economic value.

Consider four layers of the AI system:

1. Compute, which supplies the raw capacity for training and inference.
2. Software, which makes that capacity programmable and useful.
3. Distribution, which puts AI tools in front of millions or billions of users.
4. Feedback, which improves models, workflows, and products through usage.

A weak AI investment has only the first layer. It owns expensive machines.

A stronger investment adds software. It makes the machines easier to use and harder to replace.

A powerful investment adds distribution. It makes advanced capabilities available through an existing cloud, operating system, productivity suite, social network, or developer platform.

The strongest investment adds feedback. Every interaction helps the system improve, whether through explicit user preferences, verified outcomes, synthetic data, tool use, or the discovery of new failure modes.

This framework explains why two companies can spend the same amount on compute and produce radically different results. One is renting capacity. The other is building a compounding organism.

The hidden asset is not the chip, but the learning loop

The most valuable AI systems increasingly resemble businesses that learn rather than products that merely sell.

A model answers a question. The answer is evaluated. The system observes whether the user accepts it, edits it, rejects it, or uses it to complete a task. That signal can improve post training, tool selection, routing, and product design. Better performance attracts more users, and more users produce more feedback.

This is the classic consumer technology flywheel, but with a new engine. A product attracts users. Users generate data. Data improves the algorithm. The algorithm improves the product. The improved product attracts more users.

Reasoning systems make this loop more powerful because the feedback does not have to be limited to clicks or ratings. It can include verifiable results. Did the code compile? Did the customer issue get resolved on the first call? Did the booking match the traveler’s preferences? Did the financial report reconcile? Did the robot place the glass correctly in the dishwasher?

These are not vague signs of engagement. They are outcomes.

That changes the economics of software. For decades, many software companies sold access to tools while leaving customers responsible for the final result. AI makes it possible to move closer to outcome based pricing. A customer service system may be paid for resolved cases. A travel assistant may receive an affiliate fee when it completes a booking. A coding tool may be valued according to the amount of usable software it helps produce.

The shift creates both opportunity and danger. Companies that preserve a traditional subscription model may protect their margins while missing the larger market. Companies that spend heavily on inference may show lower gross margins while creating much more revenue and much stronger customer relationships.

This is why declining gross margins can sometimes be a sign of success. If a software company reports 90 percent margins because nobody uses its expensive AI features, that efficiency may be meaningless. Another company might earn far more revenue at 60 percent margins because customers rely on its AI every day.

The right question is not, “How high are the margins?” It is, “What does each dollar of lower margin purchase?”

Does it buy temporary usage that disappears when prices change? Or does it buy proprietary feedback, customer habit, workflow integration, and a larger installed base?

That is the difference between consumption and investment.

Why flexibility beats perfection in a changing environment

AI hardware and software are developing at different speeds. Model architectures can change within months, while data centers, networking systems, and specialized chips may take years to design and deploy.

This creates a dangerous mismatch. A system optimized for one model architecture may be obsolete before its physical infrastructure is fully depreciated. The winning design therefore cannot be the one with the highest peak performance under one narrow assumption. It must combine specialization with adaptability.

This principle explains the extraordinary importance of a software layer such as CUDA. Its value is not merely that it makes a chip fast. Its deeper value is that it gives developers a stable environment while the underlying algorithms continue to change.

The same logic applies to organizations. A company that builds only for today’s model may achieve impressive benchmarks and still lose. A company that builds an adaptable platform can absorb new architectures, new workloads, and new customer demands without asking its entire ecosystem to start over.

Flexibility is not the opposite of optimization. It is optimization across time.

The strongest platforms are therefore not defined by a single technical advantage. They combine several forms of compounding:

• An installed base that developers already understand.
• A trusted software environment that preserves prior work.
• A supply chain capable of producing increasingly complex systems.
• A feedback network that reveals what customers and researchers will need next.
• A reputation for continued investment, so partners are willing to commit before demand is obvious.

This last point is easy to underestimate. The future of AI depends on memory, packaging, networking, cooling, power generation, construction, and robotics. A company building at this scale cannot simply order the future from a catalog. It must persuade suppliers to invest billions in capacity that may not become profitable for years.

That is not ordinary procurement. It is a form of market creation.

When a leader explains a likely future to employees, suppliers, customers, and investors, the explanation itself can change the probability of that future. Suppliers build capacity. Developers learn the platform. Customers design their workflows around it. The expected future becomes partially self fulfilling because people coordinate around a credible forecast.

This does not make every forecast correct. It does mean that strategy is not passive prediction. At sufficient scale, strategy is the construction of an ecosystem that makes a prediction more likely to come true.

The next bottleneck is coordination, not intelligence

The popular image of AI progress focuses on model intelligence. But the practical bottlenecks are increasingly systemic: power, cooling, networking, memory, data movement, latency, software compatibility, and the ability to deploy models where users actually need them.

A single GPU is no longer the relevant unit of analysis. The meaningful unit is an AI factory: a coordinated system of chips, racks, networks, software, energy, cooling, and operations. The question is not whether one component is efficient. The question is whether the whole system produces useful tokens, decisions, or completed tasks per unit of energy and capital.

This systems perspective also changes how to think about the energy problem. Data centers are often designed around perfect availability, even though electrical grids operate near their maximum only during relatively rare periods. If AI systems can shift workloads across locations, reduce speed temporarily, or prioritize critical tasks while delaying less important ones, they can consume excess grid capacity without demanding that the grid be rebuilt for the worst hour of the year.

That is an example of graceful degradation. A system does not need to operate at full performance under every condition. It needs to fail intelligently.

The same principle applies to business. A company does not need every product to be immediately profitable. It needs enough financial stability to run new products near breakeven while they accumulate users, data, and distribution. This is why an established software company may be able to challenge a fast moving startup if it is willing to treat its new AI product as a strategic investment rather than a margin center.

The startup often has speed. The incumbent has customers, data, trust, and cash flow. The contest is decided by whether the incumbent can redeploy those advantages before the startup’s feedback loop becomes too strong to overcome.

There is a narrow window in which an established company can use its old business to finance the new one. But the window closes when management becomes more committed to protecting current metrics than to building future capabilities.

The greatest threat to an incumbent is not a competitor with better technology. It is the incumbent’s refusal to let the new technology change its economics.

What this means for individuals and organizations

If AI is a living system, then the practical goal is not merely to “adopt AI.” It is to place AI inside a loop where each use creates a better next use.

For an individual, this means avoiding the shallow pattern of asking a chatbot occasional questions. A stronger pattern is to build repeatable workflows. Use AI to draft a report, inspect the result, record the errors, improve the prompt or process, and reuse the workflow. The value comes from the accumulated system, not from any isolated answer.

For a company, the same principle requires three changes.

First, measure outcomes rather than activity. Count resolved customer issues, completed analyses, shipped features, successful sales conversations, and hours returned to skilled employees. Token volume is an input. Business value is an output.

Second, treat internal data as a feedback asset. If employees correct AI outputs but those corrections disappear into private documents and disconnected conversations, the organization is paying for learning without retaining it.

Third, build for reversibility. New models will arrive quickly. Workflows should be modular enough that the underlying model can change without forcing the entire operation to be rebuilt.

There is also a human consequence. As functional intelligence becomes cheaper, distinctly human qualities become more economically important: judgment, responsibility, taste, courage, compassion, and the ability to coordinate people around a meaningful goal.

A machine may reason through a customer complaint, but an organization still has to decide what kind of relationship it wants with customers. An agent may optimize a hiring funnel, but leaders still decide what fairness and dignity mean in practice. Intelligence can be commoditized without making humanity interchangeable.

Key Takeaways

• Do not use utilization as the only bubble test. Ask whether infrastructure is connected to a feedback loop that improves products and economics.
• Measure what each dollar of AI spending purchases. Look for proprietary data, user habit, workflow integration, and better outcomes, not only short term margins.
• Build for adaptation. Prefer systems, software, and teams that can absorb new models and tools without requiring a complete restart.
• Run strategic products near breakeven when necessary. Existing cash flows can finance learning, distribution, and ecosystem formation.
• Make AI usage cumulative. Turn repeated interactions into documented workflows, evaluations, and retained organizational knowledge.

The AI boom may still contain overvaluation, bad forecasts, and projects that will fail. But the most important dividing line is not between a boom and a bust. It is between systems that merely consume capital and systems that convert capital into faster learning.

Dark fiber was a warning about infrastructure built ahead of use. Today’s warning should be more subtle: busy machines can still support weak businesses, while modest machines can support powerful learning loops.

The decisive asset of the AI era will therefore not be intelligence alone, and not even compute alone. It will be the ability to coordinate machines, people, software, energy, distribution, and feedback into a system that improves faster than its environment changes.

The future will not simply belong to the companies with the smartest models. It will belong to the companies that make every user, supplier, employee, and experiment increase the value of the next one.