Every AI CEO Is Warning About Danger. But What If the Real Problem Is Economics?

Almost every major AI leader is now talking about the risks of artificial intelligence. Sam Altman, Elon Musk, and Anthropic CEO Dario Amodei have all raised concerns about how powerful AI could become and what it might mean for humanity. The public conversation is therefore focused on safety, control, and the possibility that AI could become dangerous.

But I think there is another question we are not examining closely enough.

What if the deeper problem is not only that AI is becoming too powerful—but that it is becoming too expensive to operate at scale?

What if the industry’s biggest challenge is not simply controlling AI, but finding a sustainable economic model for running it? What if the enormous cost of chips, data centers, electricity, cooling, networking, land, and computation is growing faster than the value AI can realistically produce?

I cannot prove what motivates any individual CEO, and their public safety concerns may be completely genuine. But I believe the economic and infrastructure side of the AI boom deserves far more attention than it currently receives.

The real question may not be whether we can build more powerful AI.

The real question may be whether we can afford to run it.

AI Is Not Too Expensive to Build. It Is Too Expensive to Run.

The next breakthrough in artificial intelligence may not come from making models bigger or smarter. It may come from making intelligence affordable.

There is something interesting happening in the artificial intelligence industry.

Some of the most influential people in technology are warning about the dangers of AI and discussing whether its development should be slowed down. Sam Altman, Elon Musk, and Dario Amodei have all participated in public conversations about the risks of increasingly capable AI systems.

The public conversation is largely focused on safety.

What happens if AI becomes too powerful? What happens if it escapes human control? What happens if advanced systems cause catastrophic harm to humanity?

These are serious questions. I am not dismissing them.

But I think there is another question that deserves much more attention.

What if the problem with scaling AI is not only what it might do to humanity, but what it might cost humanity to operate?

My view is that the current AI infrastructure race may be running ahead of its economic reality.

We are investing enormous amounts of money into data centers, advanced chips, energy systems, and computing infrastructure. We are increasing computational power to build more capable models. We are asking AI systems to perform increasingly complicated tasks, often through long sequences of token generation, reasoning, and tool use.

But at what point does the cost of producing that intelligence become greater than the economic value of the work it performs?

That is the question I keep coming back to.

I believe AI has extraordinary potential. I believe it will transform industries. But I also believe that the industry may need a major hardware and energy-efficiency breakthrough before its full economic promise can be realized.

The next AI revolution may not be about making models dramatically smarter.

It may be about making them dramatically cheaper to run.

The Difference Between AI Capability and AI Economics

We often judge AI by what it can accomplish.

Can it write software? Can it analyze documents? Can it generate business plans? Can it operate computers? Can it solve complicated scientific problems?

The answer to many of these questions is increasingly yes, at least to varying degrees.

But capability is only one part of a business model.

A company does not make a profit simply because its technology can perform a task. It makes a profit when the value generated by that task exceeds the cost of delivering it, after accounting for its other expenses.

This distinction becomes especially important with AI.

Imagine an AI system that can perform a task that previously required a human employee. That sounds like an obvious economic win.

But what if the AI needs a large amount of computation to complete the task? What if it needs multiple attempts? What if it requires extensive context, repeated corrections, and expensive tool calls? What if the infrastructure supporting that AI requires billions of dollars in capital investment?

The calculation changes.

The question is no longer just whether AI can do the work.

The question is whether AI can do the work at a cost that makes the business sustainable.

A technology can be powerful, useful, and commercially valuable while still being too expensive to operate at the scale its developers envision.

That is the central issue I believe we need to examine.

The Supersonic Lesson: Faster Does Not Always Mean Better Business

One of the clearest examples of this principle comes from aviation.

For a period of time, the world developed commercial supersonic passenger aircraft. The most famous example was Concorde, which could travel at more than twice the speed of sound and dramatically reduce travel time across the Atlantic.

Compared with conventional passenger jets, it offered a remarkable improvement in speed.

A journey that could take around seven to eight hours on a conventional aircraft could be completed in roughly three and a half hours on Concorde, depending on the route and conditions.

That was an extraordinary technological achievement.

But technological achievement and commercial sustainability are not the same thing.

Supersonic passenger travel came with major economic challenges. The aircraft were expensive to develop and operate. Fuel consumption was high. Maintenance and operating costs were substantial. There were also restrictions associated with sonic booms, which limited where supersonic aircraft could fly over land.

For many travelers, the benefit of saving several hours did not justify the much higher ticket prices.

The technology could deliver speed.

But the business model struggled to deliver that speed at a cost that made sense for widespread commercial travel.

Concorde’s commercial service eventually ended in 2003.

The world did not stop flying.

Instead, conventional subsonic jet aircraft continued to dominate commercial passenger aviation because they offered a more practical balance between speed, operating costs, capacity, and ticket prices.

The lesson is not that supersonic technology was useless.

The lesson is that a faster machine is not automatically a better commercial product.

If the cost of achieving greater speed rises too much relative to the value of the time saved, the economics can stop making sense.

This is the same distinction I see emerging in AI.

The AI Version of the Supersonic Problem

Imagine an AI model that is significantly more capable than its predecessor.

It can reason through more complicated problems. It can generate longer responses. It can use more tools. It can perform more sophisticated tasks.

That sounds like progress.

But suppose achieving those improvements requires substantially more computation, more expensive hardware, and more electricity.

Now imagine that the business wants to offer this capability to millions of users.

The company must pay for the infrastructure required to produce the additional intelligence.

The customer may appreciate the improvement. But will the customer pay enough to cover the additional cost?

That is the economic question.

A model can become more intelligent while becoming less economical to operate.

And if the cost of the additional intelligence exceeds the value it creates for customers, the improvement may not be commercially attractive.

This is not an argument against developing more capable AI.

It is an argument that capability must be accompanied by efficiency.

The aviation industry discovered that speed alone was not enough to define the best commercial aircraft.

The AI industry may discover that intelligence alone is not enough to define the best commercial model.

The Token Problem

One of the most important parts of AI economics is the cost of tokens.

Tokens are units used by many AI models to process information. They are not exactly the same as words, but they provide a useful way to understand the amount of information being processed.

For simple questions, the number of tokens may be relatively small.

But increasingly capable AI agents often perform much longer workflows.

They read instructions. They inspect files. They generate plans. They call tools. They evaluate results. They encounter errors. They revise their approach. They try again.

A task that appears simple to the user may involve a substantial amount of computation behind the scenes.

This creates a problem.

The amount of work performed by an AI system is not necessarily proportional to the value of the final result.

An agent might spend thousands of tokens trying to accomplish something that a skilled human can complete in a few minutes.

That does not mean the AI is useless. It may still provide value through availability, scale, or the ability to handle tasks that are difficult for humans to perform continuously.

But it means the economics must be measured carefully.

The question should be:

How much does it cost to produce a useful, reliable result?

Not merely: How impressive is the result?

If AI companies continue improving model capability by increasing computational requirements, they may eventually face a situation where the cost of additional intelligence rises faster than the value it creates.

That is where the business model becomes vulnerable.

A Simple HTML Website Can Reveal the Problem

Let us take a basic example.

Suppose you want to create a simple text-only HTML website.

No complicated animations. No advanced application logic. Just a clean website with headings, paragraphs, and a few sections.

A competent developer who understands HTML and CSS may be able to build it directly, make the necessary adjustments, and deliver the result.

Now ask an AI agent to do the same thing.

The agent may produce the first version quickly. It may even look impressive.

But then you notice problems.

The spacing is wrong. The layout does not match your instructions. The text is arranged incorrectly. The code contains unnecessary elements. A correction to one section affects another.

You ask the agent to fix it.

It makes a change.

You check again.

Another issue appears.

You ask for another correction.

The process continues.

Of course, human developers make mistakes too. AI can be an excellent coding assistant, and in many workflows it genuinely saves time.

My point is not that AI always performs worse than a human.

My point is that the economic value of AI depends on the complete process required to achieve the desired outcome.

If a human can complete the task accurately in one pass, while the AI requires multiple rounds of generation, correction, and verification, the expected cost advantage may disappear.

The AI may still be valuable as a collaborator. But the assumption that automation automatically means lower cost is not always justified.

This small example reveals the larger issue.

The AI industry often highlights the final capability of a system. But businesses must pay for the entire process that produces that capability.

And when the process requires extensive computation, the token bill becomes part of the business model.

Are We Building AI Infrastructure Before the Economics Make Sense?

Look at the direction of the industry.

Companies are investing in larger data centers. They are acquiring advanced accelerators. They are expanding computing capacity. They are competing to train increasingly capable models.

The logic is understandable.

More computing power can enable more sophisticated AI systems. More sophisticated systems may create more valuable products. More valuable products may generate more revenue.

But this chain of reasoning contains an assumption that deserves scrutiny.

It assumes that the economic value of improved AI will grow sufficiently to justify the cost of producing it.

That may happen. But it is not guaranteed.

Consider what is required to operate a large AI infrastructure system.

There is the cost of purchasing and replacing computing hardware. There is electricity, cooling, networking, buildings, maintenance, staffing, financing, and the cost of keeping the infrastructure available to customers.

These costs do not disappear simply because the AI model is impressive.

Every answer, every generated piece of code, every image, every reasoning step, and every agent action consumes resources somewhere in the system.

The customer sees the result.

The company pays for the machinery that produces it.

If the cost of producing those results remains high, the business must either charge more, reduce its expenses, increase efficiency, or find a way to generate substantially more value from each interaction.

Otherwise, the business model becomes difficult to sustain.

This is why I question whether the current infrastructure race is economically mature.

We may be building the foundations of a transformative industry. But we may also be spending ahead of the efficiency improvements needed to make that industry profitable.

The Electricity Constraint

The economic problem becomes even more serious when we consider electricity.

AI computation requires energy. Training and operating models at scale depend on large amounts of computing infrastructure, and the energy required varies with the hardware, workload, and efficiency of the systems involved.

If the industry continues increasing computation without achieving comparable improvements in efficiency, electricity demand could become a major constraint.

We cannot build unlimited data centers without considering where their electricity will come from.

Power generation takes time. Transmission infrastructure takes time. Grid upgrades take time. New energy projects require investment and construction.

Even if the industry can secure enough electricity, the cost of supplying that power remains part of the economics.

There is also a broader question about sustainability.

If the world keeps increasing AI computation at a rapid pace, while each additional improvement requires more energy, how far can that model continue?

At some point, the industry must find ways to produce more useful AI work without requiring proportionally more electricity.

Otherwise, the physical infrastructure needed to support AI may become a limiting factor in its expansion.

This is why I think energy efficiency is not merely an environmental concern.

It is a business concern.

It is an infrastructure concern.

It is a question of whether AI can become a technology that operates economically across the world, rather than primarily inside the largest companies with access to enormous capital.

The Problem With Using Today’s Infrastructure Forever

We are currently building AI using technologies that have been developed for modern computing.

Silicon-based chips have enabled extraordinary progress. They have transformed the world and made modern computing possible.

But that does not mean the same hardware architecture will necessarily be the most efficient foundation for every future AI workload.

AI computation has particular demands. It requires substantial processing power, memory bandwidth, and data movement. The energy consumed by a system depends not only on how many calculations it performs, but also on how efficiently the hardware and software move and process information.

If we continue increasing AI capability primarily by adding more conventional computing resources, we may encounter increasingly difficult economic and physical constraints.

We could end up in a situation where the cost of adding more intelligence becomes too high for many applications.

That is why I believe we need to look beyond simply adding more chips.

We need to develop better ways of computing.

Perhaps that means more specialized processors. Perhaps it means new memory technologies. Perhaps it means different architectures that can perform certain AI operations more efficiently.

Perhaps the future will involve technologies that are fundamentally different from the dominant approaches we use today.

I do not know exactly which technology will win.

But I believe the direction is clear.

We need more computing power for less energy.

We need more useful output for less infrastructure.

We need better economics per token.

Could Nuclear Energy Become Part of the Solution?

There is another possibility worth considering: nuclear energy.

As AI infrastructure grows, companies may need access to large amounts of reliable electricity.

Nuclear power could become part of the long-term energy mix supporting data centers. Small modular reactors, often called SMRs, are one technology being explored for potentially providing reliable, high-output electricity in a more modular format.

However, nuclear energy is not a simple or immediate solution.

Building nuclear facilities involves substantial capital costs, regulatory requirements, engineering challenges, safety considerations, and long development timelines. Small modular reactors may offer potential advantages in deployment and standardization, but their economics and commercial availability still need to be demonstrated at scale.

The important point is not that nuclear power is automatically cheap today.

It is that future improvements in nuclear technology could potentially provide reliable electricity for energy-intensive industries.

If reactor designs become cheaper to manufacture, faster to deploy, and more economically competitive, they could play a role in supporting the energy demands of future AI infrastructure.

But even abundant electricity would not solve every AI problem.

If the chips themselves are inefficient, or if the models require excessive computation, the cost of electricity alone will not make the system economical.

The industry would still need better hardware and better software.

Nuclear energy could help address the power supply challenge.

More efficient computing could help address the power consumption challenge.

Both may be important.

The Real AI Race May Be About Efficiency

The public conversation often describes the AI race as a competition to build the smartest model.

I think that description is incomplete.

The long-term competition may be about who can produce the most useful intelligence at the lowest sustainable cost.

Consider two systems.

The first is slightly more capable, but it requires substantially more computation, energy, and infrastructure.

The second is somewhat less capable, but it can perform the required task reliably at a fraction of the cost.

For many businesses, the second system may be the better product.

The same principle applies to AI agents.

An agent that completes a task reliably with fewer model calls may be more commercially valuable than an agent that performs more sophisticated reasoning but requires many expensive iterations.

Efficiency is not a secondary feature.

It is part of the product.

It is part of the business model.

And ultimately, it is part of what determines whether AI can scale.

The Future May Depend on Doing More With Less

I do not think the answer is to stop improving AI.

I think the answer is to improve the underlying economics of AI alongside its capabilities.

We should continue developing better models. We should continue exploring new applications. We should continue investing in research.

But we should also recognize that the current approach to scaling computation may not be sustainable indefinitely.

The industry needs better hardware.

It needs more efficient inference.

It needs better model architectures.

It needs systems that reduce unnecessary computation.

It needs energy infrastructure that can support growth without creating unsustainable costs.

And it needs business models that can convert AI capability into durable economic value.

The companies that solve these problems may become some of the most important technology companies of the future.

Not necessarily because they build the largest models.

But because they make intelligence affordable enough to use everywhere.

Conclusion: AI’s Next Breakthrough May Be Economic

When I look at the AI industry today, I see a technology with enormous potential operating within an infrastructure that is still evolving.

We are building powerful models. We are investing heavily in data centers. We are increasing computing capacity. We are discovering what AI can accomplish.

But the cost of producing that intelligence remains a central challenge.

The industry cannot assume that more computation will always lead to proportionally more economic value.

It cannot assume that increasing electricity consumption indefinitely is a sustainable strategy.

And it cannot assume that every technically impressive AI application will automatically become a profitable business.

The technology may need time to mature.

Just as computing became more accessible through advances in hardware and manufacturing, AI may require its own efficiency revolution before it can reach its full potential.

That revolution could come from better chips, more efficient architectures, improved software, or entirely new approaches to computation.

It may also involve new energy technologies that can supply reliable electricity at competitive costs.

Whatever form it takes, the goal should be the same:

To produce more useful intelligence with fewer resources.

I remain optimistic about AI. I believe it can change the world.

But I think we should be willing to question the economics of the current infrastructure race.

The most important question may not be whether AI can become more intelligent.

It may be whether we can make that intelligence affordable, reliable, and sustainable enough to support the future we are trying to build.

The future of AI will not be determined by intelligence alone. It will also be determined by the cost of producing it.