The Real Reason Why True AGI Will Never Happen

AI development is about to hit the hardest wall since Pink Floyd.

It’s already been four years since tech giants started pouring billions into advancing AI, and it shows no signs of slowing down. Better and smarter models mean ever-increasing energy consumption, along with a highly complex hardware and services supply chain to support them. The irony is that most actions performed by regular users can be accomplished with a lower-tier model like GPT-4o. So why the rush?

This race is clearly not about getting better meeting summaries after your Zoom calls. It’s about reaching human- or even above-human-level cognition—what is generally referred to as AGI (Artificial General Intelligence). At that stage, AI could be put to real work.

What work you ask?

Any work humans are presently doing. Every desk job, consulting job, remote job, gig job and eventually—through integration with robotics—physical jobs. This would be an industrial‑revolution type of event, the kind that would potentially benefit only the ones controlling the technology. A post‑AGI economy is something that needs to be debated, just in case the AI overlords crack the puzzle anytime soon.

But this evolutionary step might not be as easy as just throwing money at the problem. Just like with Moore’s Law, sooner or later you hit a wall—and the wall here is good old physics.

The way AI works right now is by leveraging models created through training on data. In simple terms, you feed the system large amounts of data and it starts identifying correlations.

For example, if you want to build a machine learning model that recognizes cats, you feed it thousands of cat images. The model learns the structural patterns that make a cat look like a cat rather than a dog: eye shape and position, ear structure, fur patterns, whiskers, tail, and so on.

If you want to build a machine learning model that supports natural language conversation, you feed it vast amounts of text so it can learn the statistical relationships between words and determine what makes a sentence coherent when words are arranged in a certain order.

What this means is that initially you have some random data that does not make any sense. Through training, the randomness of the data is reduced and meaning is gained.

In other words, AI is the reduction of entropy through training. Doing that takes a lot of power, so in essence, we expend energy to reduce entropy.

Training AI is the process of spending energy in order to reduce entropy.

This may be true for narrow AI—the “dumb” AI we use every day to research various subjects, make summaries or generate pictures. But AGI is a different beast. Could it be more than an entropy‑reducing system?

Meet Landauer, the physicist who put thermodynamics and information processing in the same sentence.

Landauer’s principle is a physical principle pertaining to the lower theoretical limit of energy consumption of computation. It holds that an irreversible change in information stored in a computer, such as merging two computational paths, dissipates a minimum amount of heat to its surroundings.

According to Landauer’s principle, every data operation dissipates an amount of energy through heat. Training an AI model and operating it involves a lot of data operations. There are energetic bounds to computational processes. The second law shapes what hardware and algorithms can achieve efficiently. Meaning, we could go further, but the energy cost grows exponentially.

At this point in time, we don’t know for sure if AGI is about quantity or quality . There might be two approaches:

1. Continue spending billions on power, expanding horizontally, hoping for a breakthrough—the quantitative approach, or…

2. Spark an innovation that would enable advanced cognition in low‑energy devices—the qualitative approach.

It’s not that we’re not capable of eventually achieving AGI; it’s just that it might not be feasible in terms of cost to operate it effectively.

Think of the human brain—the most complex system in the known universe, the result of millions of years of evolution. Its energy consumption is around 20 watts per hour. This is incredibly low. Compare that to ~18 Wh per query in GPT‑5 and ~3.8 kW of hardware to run it. And these numbers are conservative.

At room temperature, the Landauer limit represents an energy of approximately 0.018 eV (2.9×10⁻²¹ J). As of 2012, modern computers use about a billion times as much energy per operation.

Source: Wikipedia

According to calculations based on the Landauer principle, we’re using hugely more power than necessary to perform data operations. Not necessarily because we like guzzling energy, but because it’s the only way we currently know how to do it. Perhaps this limitation stems from the use of silicon chips and we need to look towards biology for solutions, or from using a binary system and we need to look to quantum computing for solutions. Either way, the math suggests it is not economically feasible to run human‑level cognition on current tech.

That, however, will not stop certain individuals from moving forward with this quest. Whether this will be in humanity’s best interest or not, is probably something our future selves will get to see very soon.