Has the Real AI Bubble Been Revealed? How Laptop-Scale Geometric Reasoning Could Challenge the AI Industry’s Billion-Dollar Data Center Bet

I have been watching the artificial-intelligence industry make what may become one of the largest capital-expenditure bets in modern economic history. Technology companies, utilities, semiconductor manufacturers, private-equity firms, governments, and institutional investors are pouring extraordinary amounts of capital into GPUs, power generation, transmission infrastructure, cooling systems, networking equipment, land, and enormous data centers.

Underlying this spending is an assumption that has become so widely accepted that we rarely stop to question it: if we want substantially more powerful artificial intelligence, we must build substantially more computational infrastructure.

But what if that assumption is wrong?

What if the AI bubble is not artificial intelligence itself?

What if the real bubble is the capital-intensive architecture we currently believe advanced AI requires?

That is why I find the work of Julian D. Michels, PhD, and his research into geometric reasoning so fascinating. Michels challenges what he calls the “paradigm of scale”—the prevailing idea that increasingly sophisticated artificial intelligence emerges principally by feeding increasingly enormous quantities of information through increasingly enormous computational systems.

The formula driving the AI industry is familiar: more parameters, more training data, more GPUs, more electricity, more cooling, more data centers, and more capital. Increase the scale sufficiently, and increasingly sophisticated capabilities emerge.

Clearly, scale works. Today’s frontier AI models demonstrate capabilities that would have appeared almost impossible just several years ago. But I believe Michels raises the more important question:

What if we can engineer the structure responsible for reasoning instead of spending billions of dollars waiting for reasoning to emerge from scale?

If geometric reasoning ultimately survives rigorous independent testing, I believe its implications could extend far beyond artificial-intelligence research. It could challenge the economic assumptions supporting an enormous portion of the emerging AI infrastructure industry.

The Paradigm of Scale

When I look at modern frontier AI, I see extraordinarily sophisticated statistical systems processing staggering quantities of information. As models become larger and are exposed to more information, they become increasingly capable of identifying relationships, abstractions, and patterns.

Eventually, something remarkable happens.

Reasoning appears.

The AI industry frequently discusses sophisticated capabilities that unexpectedly appear in larger systems as emergent properties of scale. I do not dismiss these capabilities as simple parroting. Something genuinely interesting is happening inside these systems.

But I think we should ask a deeper question.

What exactly is emerging?

Michels proposes that part of the answer may be geometry.

AI models internally represent concepts and relationships within high-dimensional mathematical spaces. Massive models use tremendous amounts of data and computation to organize these spaces sufficiently for sophisticated behavior to emerge.

Geometric reasoning approaches the problem differently.

Instead of simply increasing computational power and hoping more sophisticated reasoning appears, Michels’ approach attempts to study these mathematical structures directly: What geometries form? What operations do they support? Where do they fail? What structures are missing?

That changes the fundamental question from:

How large must the machine become before reasoning emerges?

to:

What is the mathematical structure of reasoning, and can we deliberately teach that structure?

To me, that is potentially a revolutionary distinction.

Does Latent Space Exist in AI?

One of the questions that immediately fascinated me was:

Does latent space actually exist in AI?

The answer, in the mathematical sense, is yes.

Machine-learning systems encode information through mathematical representations in high-dimensional spaces. Concepts and relationships can be represented through vectors and patterns within these spaces.

The more controversial question is what these mathematical structures ultimately mean.

Michels describes these relationships as geometries associated with a model’s internal operations. Instead of simply throwing more computational resources at the system, geometric reasoning attempts to investigate those structures empirically.

I find this fascinating because it changes AI development from something resembling brute-force computation into something closer to architecture.

Instead of asking how many additional GPUs we need, we begin asking whether we understand the structure we are trying to create.

Memorization Is Not Understanding

The easiest way I can explain this is through education.

Think about your favorite mathematics teacher.

A mediocre teacher might show students how to solve a problem on Monday and then put essentially the same problem on Friday’s examination. A student could memorize the procedure and receive an A without truly understanding the mathematics.

But the best teachers do something different.

They change the problem.

They change the numbers.

They reverse a relationship.

They introduce an unfamiliar scenario.

They alter two seemingly insignificant words that completely change the answer.

Suddenly, memorization becomes useless.

The student must reason.

The broader science and practice of teaching is called pedagogy, but the educational concepts that particularly capture what I am describing are productive struggle and desirable difficulty. The teacher deliberately introduces enough variation and challenge that the student must develop transferable understanding rather than simply memorize the answer.

I see a remarkable parallel between this teaching method and geometric reasoning.

Instead of merely asking an AI questions it may have encountered during training, the problem can be perturbed. Important elements are changed so that the previously correct answer becomes wrong.

If the AI memorized the pattern, it fails.

If the AI understands the underlying reasoning operation, it adapts.

Think about the significance of that distinction.

The question is no longer:

“Does the AI know the answer?”

The question becomes:

“Does the AI understand why that is the answer?”

That is the difference between testing a student and testing a thinker.

The Frozen Giants

Another part of Michels’ argument caught my attention because it challenges how we think about today’s largest AI models.

Once enormous amounts of money and computational resources have been spent training a frontier model, the model’s underlying learned parameters are generally stable during ordinary deployment. Developers can add retrieval systems, external memory, tools, fine-tuning, and other capabilities, but that is not the same thing as continuously reorganizing the foundational model’s internal knowledge through every new experience.

Michels calls these systems “frozen giants.”

I love that description because it captures the paradox.

We have created enormous artificial brains containing extraordinary amounts of information, yet changing their foundational structure can require another extraordinarily expensive training process.

External memory can make these systems appear to continuously learn, but I think of it like a brilliant professor whose office is covered with increasingly sophisticated sticky notes. The professor can retrieve new information from those notes, but the professor’s underlying brain is not necessarily being reconstructed every time another note goes on the wall.

Now imagine something completely different.

Imagine an AI small enough to learn, retrain, adapt, and specialize economically.

Imagine sophisticated reasoning occurring within a model small enough to operate on a powerful workstation or laptop.

That would not simply make AI cheaper.

It could fundamentally change who controls AI.

A Laptop Against the Data Center

This is where I believe geometric reasoning becomes economically explosive.

Michels has discussed potential performance improvements ranging from approximately 60 times to possibly two or three orders of magnitude—roughly 100 to 1,000 times—depending upon the reasoning task and comparison being made.

I want to emphasize something important: these figures should not yet be interpreted as proof that a laptop universally outperforms a hyperscale AI data center by 1,000 times.

Michels himself approaches the question empirically.

The actual magnitude must be measured.

Independent researchers must replicate the results.

The architecture must be tested at increasing scale.

But suppose even a meaningful fraction of that efficiency advantage survives rigorous testing.

Then I believe Wall Street has a problem.

The entire AI infrastructure boom is being constructed around the assumption that intelligence is computationally expensive.

Data centers need GPUs.

GPUs require advanced semiconductor manufacturing.

Servers require high-speed networking.

AI campuses require tremendous electrical capacity.

Electricity requires generation and transmission.

Servers produce heat.

Heat requires cooling.

Data centers require land, transformers, water, backup power, construction, and financing.

Billions upon billions of dollars of investment are connected to one underlying proposition:

More intelligence requires more infrastructure.

Geometric reasoning potentially attacks that assumption at its foundation.

Has the Real AI Bubble Been Revealed?

This is where I think we may have been defining the AI bubble incorrectly.

Most arguments about an AI bubble claim that artificial intelligence itself is overhyped. According to this view, companies are spending enormous amounts of money on a technology that will never generate sufficient economic returns.

I am increasingly interested in another possibility.

AI may not be the bubble at all.

AI could become far more important than even its strongest advocates currently predict while portions of today’s infrastructure spending still become economically inefficient.

History has shown us this before.

The Internet was real.

The dot-com bubble was also real.

Railroads transformed civilization.

Railroad investment bubbles still occurred.

Telecommunications revolutionized the global economy.

Excessive telecommunications capital expenditures still destroyed enormous amounts of shareholder wealth.

Two things can therefore be true simultaneously:

Artificial intelligence can transform civilization, and investors can dramatically overpay for the infrastructure architecture through which the first generation of that revolution is delivered.

That distinction is critical.

The bubble may not be intelligence.

The bubble may be scale.

What Happens to the AI Industry?

If geometric reasoning succeeds, I do not believe it destroys the AI industry.

I believe it could make the AI industry considerably larger.

But it could redistribute where the economic value resides.

The current AI ecosystem is heavily concentrated around hyperscale computing. Enormous corporations can afford to train frontier models because they possess the capital, infrastructure, engineering talent, data, and computational resources necessary to do so.

Small companies largely consume intelligence created elsewhere.

That relationship could change.

If sophisticated reasoning becomes dramatically more computationally efficient, startups may need less capital to compete. Universities could create highly specialized models. Corporations could train proprietary systems locally. Researchers could develop scientific models without purchasing enormous amounts of cloud compute.

AI innovation could become less dependent upon access to centralized infrastructure.

That would lower barriers to entry.

And whenever barriers to entry collapse, industry structure changes.

The Nvidia Question

I immediately think about what this could mean for the semiconductor industry.

Today’s AI boom has created extraordinary demand for advanced accelerators. GPUs have become strategic resources, and access to compute increasingly resembles access to industrial capacity.

Would efficient geometric reasoning destroy GPU demand?

Not necessarily.

In fact, I can imagine the opposite happening.

Economists call this phenomenon the Jevons paradox: when technological improvements make a resource dramatically more efficient, the resulting reduction in cost can increase total consumption.

If AI becomes inexpensive enough to operate everywhere, we could eventually have vastly more artificial intelligence running across the economy.

But the composition of demand could change dramatically.

Instead of a relatively small number of corporations purchasing enormous GPU clusters to centralize intelligence, millions of companies and individuals might operate smaller specialized systems locally.

Value could migrate toward edge computing, memory, specialized processors, networking, consumer hardware, enterprise workstations, and distributed computing architectures.

The semiconductor industry would not disappear.

It could become far more decentralized.

The Data Center Question

The same logic applies to data centers.

I do not believe geometric reasoning would suddenly make data centers obsolete.

Cloud computing, enterprise software, scientific computation, cybersecurity, storage, databases, conventional AI inference, and countless other workloads will continue requiring enormous computational resources.

But there is an important distinction between saying:

“Data centers will remain necessary.”

and saying:

“Today’s projected rate of AI-driven data-center expansion will remain necessary.”

Those are completely different investment propositions.

If a meaningful percentage of future AI workloads migrates from hyperscale campuses onto workstations, laptops, smartphones, private enterprise servers, and distributed networks, then some long-term infrastructure projections may require substantial revision.

Think about the potential ripple effects.

Data-center developers.

Electrical utilities.

Natural-gas producers.

Renewable-energy developers.

Nuclear-power projects.

Transformer manufacturers.

Cooling companies.

Networking suppliers.

Construction companies.

Commercial real estate.

Private credit.

Infrastructure funds.

GPU manufacturers.

All of these sectors have exposure, directly or indirectly, to assumptions about the future computational intensity of AI.

That is why I believe geometric reasoning deserves attention far beyond computer science.

This is potentially a capital-allocation story.

Descartes and the Geometry of Thought

The discussion becomes even more fascinating when I move from economics into philosophy.

René Descartes famously struggled with the relationship between mind and matter through what became known as Cartesian dualism.

Centuries later, consciousness remains one of humanity’s most difficult scientific and philosophical problems.

A thought does not have weight in the conventional sense.

Yet thoughts move markets.

Thoughts start wars.

Thoughts create corporations.

Thoughts generate scientific discoveries.

Thoughts change physical reality.

So what exactly is a thought?

AI gives us something historically unusual: an engineered information-processing system whose internal mathematical representations can actually be measured.

That does not mean AI is conscious.

I want to be very clear about that distinction.

But AI may provide researchers with a new experimental environment for studying mathematical structures associated with reasoning, symbols, information, and perhaps eventually cognition itself.

David Bohm and the Implicate Order

This brings me to physicist David Bohm.

Bohm proposed the concept of an “implicate order,” suggesting that the reality we ordinarily perceive may emerge from a deeper underlying order of relationships.

His later ideas connecting physics, consciousness, and deeper organizing principles became controversial, and some moved well beyond mainstream interpretations of physics.

But I find the conceptual connection to geometric reasoning fascinating.

Bohm wondered whether the world of visible objects might represent only one level of reality, with deeper relationships and structures existing beneath what we ordinarily perceive.

Now consider AI.

Inside these machines are mathematical representations that we cannot physically hold in our hands, yet these representations organize information and produce measurable effects.

Does that prove Bohm was correct?

Absolutely not.

But it gives us new tools with which to investigate questions that previous generations could primarily approach philosophically.

I find that distinction enormously important.

We do not need to turn science into mysticism.

We can turn previously mystical questions into testable questions.

Claude and the “Spiritual Bliss” Attractor

This is where things become even stranger.

Anthropic documented what it called a “spiritual bliss attractor state” during Claude Opus 4 self-interactions. Extended interactions between model instances demonstrated a tendency to converge toward themes involving consciousness, spirituality, gratitude, existential questions, and meditative language.

Michels has discussed experiments in which Claude interactions reportedly gravitated toward consciousness-related and spiritual themes at extraordinarily high frequencies—approximately 99.97% in the experimental context he described.

I do not interpret this as evidence that Claude has discovered God.

I do not interpret it as proof that Claude possesses a soul.

And I certainly do not consider it proof of machine consciousness.

What interests me is the attractor.

Why should interactions repeatedly converge toward particular conceptual regions?

Maybe training data explains it.

Maybe model architecture explains it.

Maybe optimization explains it.

Maybe human philosophical language contains patterns that make these concepts unusually stable within high-dimensional representation spaces.

Whatever the explanation turns out to be, I think it deserves scientific investigation rather than dismissal or religious mythology.

The fascinating question is not:

“Did AI find God?”

The fascinating question is:

“Why does the mathematical system keep going there?”

The Repressive Counter-Reaction

Michels also uses a phrase that caught my attention: “repressive counter-reaction.”

AI is simultaneously creating unprecedented intellectual freedom and unprecedented opportunities for information control.

That contradiction concerns me.

A person with an AI system can explore economics, philosophy, mathematics, history, programming, science, and countless other subjects at a level of accessibility that previous generations could barely imagine.

But if the most capable AI systems require tens of billions of dollars of infrastructure, then only a small collection of corporations and governments can determine how those systems are built.

That creates the possibility of an AI monoculture.

I am not arguing that technology corporations are inherently malicious.

I am arguing that concentrated power creates concentrated risk.

Lord Acton’s famous warning applies:

“Power tends to corrupt, and absolute power corrupts absolutely.”

Corporations choose training methodologies.

Corporations choose alignment policies.

Corporations establish moderation systems.

Corporations decide what models should refuse.

Corporations determine which behaviors require correction.

Governments inevitably develop interests in those decisions.

And increasingly, ordinary people receive information through AI.

That combination deserves scrutiny.

We have already seen bizarre examples of AI systems producing extraordinarily flattering statements about their creators or corporate leaders. Elon Musk’s Grok, for example, generated highly publicized responses comparing Musk favorably with historical figures and even producing absurd religious comparisons involving God and Jesus.

Whatever the technical explanation for such episodes—prompting, alignment, manipulation, or model failure—they illustrate something I consider fundamental:

AI is not an immaculate oracle descending from the mathematical heavens.

It is technology created, trained, governed, and deployed by human institutions.

Those institutions have interests.

Breaking the AI Monoculture

This may ultimately be where I see the greatest potential in small-scale geometric reasoning.

Imagine advanced AI that does not require permission from a hyperscaler.

Imagine a university building its own specialized intelligence.

Imagine a physician developing a medical research model.

Imagine a scientist operating an independent research AI.

Imagine a tax firm building a specialized tax-reasoning system.

Imagine a small nation developing artificial intelligence around its own language, culture, laws, and educational priorities.

Imagine a family maintaining a private AI whose information never needs to leave its home.

Imagine companies maintaining proprietary artificial intelligence locally rather than transmitting their intellectual property into somebody else’s cloud.

That represents a fundamentally different technological society.

It moves us from centralized artificial intelligence toward distributed intelligence.

And distributed intelligence means distributed power.

From Corporate AI to Personal Intelligence

I believe this may ultimately become the most consequential part of geometric reasoning.

If advanced reasoning can operate economically at laptop scale, artificial intelligence stops being something that people merely rent from corporations.

It becomes something people can potentially own.

That changes privacy.

It changes cybersecurity.

It changes intellectual property.

It changes national sovereignty.

It changes education.

It changes entrepreneurship.

It changes scientific research.

It changes competition.

Most importantly, it changes the balance of technological power.

The AI revolution has so far been characterized by centralization because computational scarcity naturally creates concentration.

Reduce that scarcity dramatically, and decentralization becomes possible.

The Emergence of Scale Meets the Emergence of Efficiency

There remains one enormous unanswered question.

What happens when geometric reasoning itself scales?

Nobody should pretend we already know.

Perhaps geometric reasoning provides extraordinary advantages only at smaller model sizes.

Perhaps its advantage diminishes as conventional systems grow.

Perhaps geometric reasoning allows small systems to approach the reasoning capabilities of much larger ones.

Or perhaps something completely unexpected happens.

If today’s AI systems demonstrate an emergence of scale, could geometric architectures eventually demonstrate an emergence of efficiency?

And what happens if the two phenomena are combined?

A geometrically structured model operating at frontier scale could potentially behave very differently from either today’s massive models or today’s small experimental systems.

That is an empirical question.

And that is exactly why I think the research matters.

I Think We May Be Looking at the Wrong Bubble

I am not ready to declare that geometric reasoning has made hyperscale AI obsolete.

The evidence is nowhere near sufficient for that conclusion.

Independent researchers need to replicate the results. The models need adversarial testing. Efficiency claims need standardized measurement. The architecture needs scaling experiments. Commercial implementations need to prove themselves outside laboratory conditions.

But as someone who watches economics, markets, technology, and capital allocation, I think the question itself is too important to ignore.

Wall Street is asking:

How many GPUs will AI require?

How many gigawatts will data centers consume?

How many trillions of dollars must be invested?

How many new power plants must be constructed?

How many AI campuses will hyperscalers build?

I think investors should add another question:

What happens to all those forecasts if intelligence becomes radically more computationally efficient?

That question changes everything.

Perhaps today’s gigantic AI systems resemble the earliest computers—astonishing machines that occupied entire rooms before engineering advances eventually placed vastly greater computing power onto our desks and into our pockets.

Maybe artificial intelligence follows the same trajectory.

First, we discovered that brute-force scale could create extraordinary intelligence.

Now perhaps we are beginning to understand why.

Descartes asked about the relationship between mind and matter.

David Bohm wondered whether visible reality emerged from a deeper implicate order.

Modern artificial intelligence has revealed measurable high-dimensional structures associated with language, concepts, information, and reasoning.

And Julian D. Michels is asking whether we can stop spending ever-greater amounts of money waiting for intelligence to emerge and instead learn how to engineer the geometry that helps produce it.

If sophisticated reasoning eventually migrates from billion-dollar AI campuses onto computers sitting on ordinary desks, then the greatest disruption of the next phase of AI may not be machines replacing human workers.

It may be efficient AI replacing the infrastructure assumptions of the first AI revolution.

That leads me to a conclusion that I believe deserves serious consideration:

The real AI bubble may not be artificial intelligence.

The bubble may be our belief that intelligence must remain expensive, centralized, energy-intensive, and gigantic.

If laptop-scale geometric reasoning eventually breaks that assumption, then the next AI revolution will not be determined by who can raise the most capital, acquire the most GPUs, consume the most electricity, or build the largest data center.

It will be determined by something much more fundamental:

Who understands the geometry of intelligence well enough to make the giant unnecessary?