Why intelligence, energy efficiency, autonomy, ownership, and political power are not the same thing.
1. The Thermodynamic Starting Point — AI’s Extraordinary Cost
Artificial intelligence is increasingly becoming an industrial construction project rather than merely a software project.
The discussion around OpenAI’s reported “Doug” pre-training model and the rumored 10-trillion-parameter “Bel” successor illustrates where the frontier could be heading. At that scale, the problem is no longer simply writing better algorithms. It becomes a question of memory, silicon, networking, electricity, cooling, capital and physical infrastructure. A hypothetical 10-trillion-parameter model would require roughly 20 terabytes simply to store its weights at 16-bit precision, while the full training footprint could be vastly larger. The required accelerator fleet could run into tens of thousands of high-end processors, with billions of dollars committed to silicon alone.
This changes the meaning of “building an AI model.” The frontier is becoming increasingly tied to the ability to mobilize industrial resources.
The question is no longer only, How intelligent can we make the model? It is also, How much civilization can we afford to put behind it?
2. Beyond Scaling — When Intelligence Becomes a Compute Budget
For years, the dominant intuition was that more data, more parameters and more training compute would steadily produce more capable models.
The frontier is now moving toward another idea: spend more computation at the moment of thinking itself.
Instead of performing one inference and accepting the answer, an AI system can generate multiple approaches, criticize them, verify them, search for alternatives, backtrack and try again. Multiple agents can work on different parts of a problem and communicate their findings.
The result is a subtle change in the economics of intelligence.
Intelligence is no longer represented only by what has been permanently stored in a model’s weights. Some intelligence can be temporarily purchased with compute, time and electricity.
How much intelligence can we buy if we are willing to spend enough computation?
3. The Silicon Paradox — More Intelligence, Enormous Energy
There is, however, an extraordinary paradox hidden inside this progress.
The machines becoming increasingly capable are also extraordinarily expensive in physical terms.
The discussion estimates that a 10-trillion-parameter frontier system could involve 160–200+ terabytes of training memory, tens of thousands of accelerators and potentially 100–250+ megawatts of continuous power. The numbers are not merely engineering specifications. They reveal the physical scale of modern artificial intelligence.
A model may appear to be a piece of software when viewed through a chat window. Behind that interface, however, may be an industrial system occupying buildings, consuming electricity, moving enormous quantities of data and depending on sophisticated semiconductor supply chains.
This produces an uncomfortable comparison.
The most advanced artificial intelligence may require a power infrastructure measured in megawatts. The human brain operates on approximately 20 watts of metabolic energy.
That is not a minor efficiency difference. It is an entirely different physical regime.
4. The 20-Watt Counterargument — Biology Is Still an Astonishing Machine
The human brain does not achieve intelligence by simply throwing more computation at a problem.
Its architecture is fundamentally different. Memory and computation are intertwined at the synapse. Neural activity is sparse and event-driven. Learning occurs continuously rather than through periodic retraining of a frozen model. Biological neurons are also chemically and dynamically richer than the numerical operations represented by conventional artificial neural networks.
The result is extraordinary.
A roughly 1.4-kilogram organ consumes about 20 watts while allowing a human being to perceive the world, move through it, learn from experience, form relationships, imagine possibilities, pursue goals and construct abstractions.
The contrast raises a deeper question.
If general intelligence already exists in nature at such extraordinary energy efficiency, perhaps simply scaling silicon is not the final answer.
Perhaps the real challenge is architectural. Perhaps the question is not merely how to make computers bigger, but how to make computation more like the thing biology has already achieved.
5. But Humanity Is Not One Brain — It Is Billions of Autonomous Minds
There is another problem with the usual comparison between AI and the human brain.
We often talk about “the human brain” as though humanity were a single competing computational system.
It is not.
Humanity consists of billions of individual minds. Each person carries a separate history, separate experiences, separate goals and separate ways of looking at the world. These minds interact, compete, cooperate, teach one another and generate new ideas.
One human may discover something another never considered. One generation can inherit the discoveries of previous generations and extend them. A scientific idea can travel across continents. A technological invention can be reproduced millions of times.
Human intelligence is therefore not simply the computation occurring inside one skull. It is also a distributed social system of billions of autonomous agents.
That changes the comparison.
The relevant contest is not necessarily one AI system versus one human brain. It may be artificial intelligence infrastructure versus billions of biological agents embedded in civilization.
And those biological agents possess properties that are not captured by a parameter count.
6. Intelligence Is Not Agency
This is where the discussion moves from computation to power.
A system can be extremely intelligent without necessarily being autonomous. It can solve a problem without choosing which problem matters. It can generate a strategy without possessing an independent reason to pursue that strategy. It can calculate the fastest route without caring where it is going.
That distinction matters because we routinely collapse several different concepts into one word: intelligence.
But intelligence, capability, agency, autonomy and control are not identical.
A system may be extraordinarily capable while still operating inside objectives, infrastructure and institutions established by someone else.
Who decides which problem AI is solving?
7. The Ferrari Principle — Driving Is Not Manufacturing
Consider a Ferrari.
A person who has passed 10th standard can potentially learn to drive one. They can accelerate it, brake it, steer it and decide where it goes. They do not need to understand the thermodynamics of the engine, the metallurgy of its components, the aerodynamics of the body or the semiconductor technology inside its electronics. They certainly do not need to know how to design the Ferrari.
And yet they can control it.
Driving the Ferrari is not making the Ferrari.
Operating a sophisticated system does not require understanding every layer that makes the system possible.
The analogy matters for AI. Suppose an AI becomes vastly better than an individual human at mathematics, programming, scientific reasoning or information retrieval. That demonstrates extraordinary capability.
But capability inside a technological system is not automatically equivalent to the ability to independently create, maintain, reproduce or control the entire civilization that supports that system.
A Ferrari can be vastly more sophisticated than its driver. The sophistication of the machine does not automatically transfer ownership of the road to the machine.
8. The Business-Owner Principle — The Smartest Person Need Not Be in Control
There is another analogy that may be even more revealing.
Imagine a large company. The owner may not be the best engineer in the organization. The chief engineer may know more engineering. The accountant may know more accounting. The lawyer may understand the law better. A scientist may understand the company’s technology at a depth the owner could never match.
Yet the owner can still control the organization.
Why? Because control does not necessarily come from being the smartest person in the room. It can come from ownership, resources, authority and decision rights.
The business owner decides what the organization is trying to accomplish. Other people may possess greater expertise in individual domains, but their expertise operates within an organizational structure.
Being the most intelligent participant is not the same as being the controlling participant.
If AI eventually becomes more intelligent than humans in many domains, that fact alone does not tell us who controls civilization. We have to ask who controls the resources, infrastructure, institutions and decisions through which that intelligence is deployed.
9. The Missing Variables — Autonomy, Mobility, Desire and Drive
This is where five variables from the discussion become important: autonomy, mobility, desire, drive and political power.
A human being is not merely an information-processing system. A human can decide to leave the room. A human can walk somewhere. A human can start a company. A human can refuse an instruction. A human can desire something that nobody asked them to desire. A human can spend decades pursuing a goal simply because they want to.
These characteristics are not reducible to benchmark scores.
If AI is eventually to become an independent actor rather than an extremely powerful instrument, the relevant transition would involve much more than intelligence.
It would involve the acquisition of persistent agency. The machine would need not merely to know what could be done, but to independently determine what it wants to accomplish, acquire or influence the resources required to accomplish it, and continue acting toward that objective.
That is a very different threshold.
10. From Intelligence to Power — The Real Threshold
We can therefore imagine a progression:
Intelligence → Capability → Autonomy → Mobility → Resource Acquisition → Economic Power → Political Power → Control
The first step is cognitive. The later steps are institutional and physical.
An AI could become extremely intelligent while remaining dependent on humans for electricity, hardware, maintenance, physical access, legal authorization and deployment.
That does not mean such a system would necessarily remain harmless or controllable. It means that intelligence alone does not establish the outcome.
The real question is where the transition occurs.
At what point does superior intelligence become sufficient to acquire the other forms of power?
That is the more consequential AI-control problem.
11. The Open-Source Problem — Intelligence May Become Cheap to Copy
There is another paradox in the economics of frontier AI.
The leading laboratories can spend enormous sums discovering new capabilities. But once an architectural idea, training technique or capability becomes understood, it can potentially spread.
The discussion compares this dynamic with the history of AlphaZero and open-source chess engines, and argues that techniques developed through expensive proprietary systems can eventually be reproduced or distilled into much cheaper systems.
This creates a strange economic situation.
The frontier laboratory may spend billions to discover a capability. The rest of the ecosystem may then spend dramatically less to reproduce something close to it.
If this pattern continues, raw compute expenditure may provide only a temporary advantage. The frontier keeps moving forward, but yesterday’s frontier gradually becomes infrastructure.
12. The Copyability Advantage — AI Does Not Need Eight Billion Bodies
There is, however, one enormous advantage on the artificial side.
A human brain cannot simply be copied. If one human becomes exceptionally capable, there is still only one physical human being with that particular brain, experience and personality.
Software is different.
If an artificial intelligence can be instantiated efficiently, one capable system can potentially become many instances.
This creates an extraordinary asymmetry.
Humanity has billions of autonomous minds, but each biological mind is physically embodied. AI may eventually have the ability to reproduce a successful cognitive architecture across machines.
The contest could therefore evolve from billions of human minds versus one artificial system into something more complicated: billions of autonomous biological agents versus potentially enormous populations of artificial agents.
At that point, the copyability of intelligence could become more important than the intelligence of any single system.
13. The Central Question — Who Is Actually Driving the Ferrari?
The Ferrari analogy now becomes more interesting.
Imagine a machine that is vastly more capable than its driver. It can calculate faster. It can perceive more information. It can predict trajectories better. It can perhaps even recommend better decisions than the driver.
But there remains a difference between capability and authority.
Who owns the car? Who controls the keys? Who determines the destination? Who controls the roads? Who controls the fuel? Who decides whether the car is allowed to operate?
These questions sound mundane when applied to a Ferrari. They become profound when applied to artificial intelligence.
AI may eventually become vastly more capable than the human beings operating it. But capability does not automatically answer the question of control.
14. The Final Paradox — Intelligence May Not Be the Same as Power
The most important conclusion may therefore be surprisingly simple.
AI becoming smarter than humans is not the same event as AI becoming more powerful than humans.
Intelligence is one component of power. But power also involves agency, resources, physical capability, institutions, ownership, social coordination and political authority.
The human brain itself remains an astonishing counterexample to the idea that intelligence requires enormous computational infrastructure. A roughly 20-watt biological system can produce flexible intelligence while billions of such systems cooperate and compete across a civilization.
AI may eventually overcome the brain in raw cognitive performance. It may become faster. It may become more knowledgeable. It may reason across more information. It may even become dramatically more capable than an individual human.
But there is still a question that intelligence alone cannot answer:
Who is driving the Ferrari?
And perhaps the deepest AI question is therefore not, Will AI become smarter than us? but, If AI becomes smarter than us, what else must happen before it becomes powerful enough to control where the Ferrari goes?