AI’s biggest risk is not progress, it’s energy. But recently, I came across one of the most fascinating pieces of journalism I’ve read about a decades-old theory that could come true in 2025 and help us fight this risk.

Published in the reputable IEEE Spectrum magazine, it introduces reversible computing.

The promise? 4,000x efficiency gains.

The concept? Magical.

By the end of this piece, you will have learned one of humanity’s best bets at solving the energy problem and how one of the most important laws in nature will play a crucial role in making AI more efficient.

I make AI simple, less sexy (no hype), but much more potent for you. Newsletter written by an AI expert — or at least pretends to be one — with 170k followers and hundreds of thousands of monthly readers with independent thoughts.

Click below and start today for free.

The Energy Problem

Everyone agrees that while AI seems unstoppable at the technological level, it appears much weaker at the energy level.

Its foundation is brittle.

In fact, many CEOs, such as Mark Zuckerberg and Sam Altman, agree that energy constraints are the biggest risk to AI and that energy breakthroughs are desperately needed.

While building more power plants is one way to combat this, it will likely not be enough. The speed at which AI progresses and increases energy demand will surely outpace new energy generation.

A Huge Energy Problem

With AI, the world is rushing into a new state of affairs for the Internet. We are going from a digital world ruled by Google queries to one where most interactions are done with/exploiting AIs, like using ChatGPT to search the Internet instead of the 10 horrible blue links.

And while we can agree this is progress, it has catastrophic energy consequences. As seen below, according to estimates by the Electric Power Research Institute, a standard Large Language Model query consumes around 2.9 Wh (Watts per hour) compared to the 0.3 Wh of a Google query.

Ten times more.

With AI overviews (when Google uses AIs to generate responses to your queries instead of you having to search for them), that number approaches 10 Wh, or 30 times the energy cost.

And if we think about tools like Gemini’s DeepResearch, which can search and generate tokens for up to 10 minutes or more, the energy consumption will make even the most diehard fans of AI dizzy.

Concerningly, these numbers are probably dwarfed by now due to Large Reasoner Models (LRMs) like OpenAI’s o1 and o3. Although estimates differ, LRMs generate twenty times more tokens per query than their LLM counterparts.

For example, just today, while interacting with DeepSeek-r1 distillation of qwen 2.5 7B, a new powerful LRM, an extremely simple question request made the LRM generate an insane amount of tokens, something an LLM would have solved in a few dozen (yes, LRMs aren’t always the answer).

And while this increase in tokens does not imply a twentyfold increase in costs, as LRMs are generally smaller (meaning the number of mathematical operations needed to make a prediction falls considerably), the models are run for much longer. Thus, a reasonably safe estimate would argue that the energy cost of LRMs will be around 15 times the cost of an LLM, or 45 Wh, and that’s in a best-case scenario.

The result?

We are moving into a world where the average Internet request consumes , at least, 150 times more energy than before. This leads to a necessary escalation of energy requirements, to the point that SemiAnalysis estimates that by 2026, the world’s data centers will require 96GW of power, accounting for 830 TWh of consumed energy.

If that were a country, that would be the world’s sixth-largest electricity consumer.

This has also led to an escalation in the rhetoric of these companies, with OpenAI just announcing, via Trump, a $500 billion data center investment alongside Oracle and SoftBank.

Of course, you would assume that energy efficiency would improve as fast as energy demand increases. But as Koomey’s Law (Moore’s Law’s cousin) suggests, the amount of computations we can perform per unit of consumed energy, originally doubling every 1.5 years, is slowing down to almost three years (2.6).

For a more detailed analysis of energy constraints, read here.

All in all, energy demand is increasing faster than what we can meet through simple net improvements in energy generation capacity. We need something else.

And that “something else” could be reversible computing, a fascinating area of computers that could finally see its first product after decades of research.

Where Thermodynamics and Compute Meet

In a nutshell, reversible computing refers to the emerging field that aims to make computations reversible, thereby preventing compute erasure and, thus, minimizing heat dissipation, leading to more energy-efficient computers.

That’s a whole lot of incomprehensible jargon, Nacho! I know, but bear with me.

The Value of Reversible Computing

When a computer resets itself to prepare for a new calculation, it ‘cleans up’ the results of the last one. This cleanup generates heat as a byproduct, which is essentially lost energy.

Thus, the goal of reversible computing is reusing this energy instead of wasting it by preventing it from becoming heat and using it instead to perform a recomputation that reverses the circuit back to its original state.

While we may never create entirely waste-free computers (100% energy reuse), the potential energy savings are staggering — up to 4,000 times more efficient than current systems, music to the ears of AI incumbents.

While this summarises reversible computing, it does not help us understand the beautiful physics behind it.

Why could it work? Let’s solve that.

AI and Thermodynamics

To fully understand this, we need to revisit the second law of thermodynamics, which states that the entropy of an isolated system never decreases over time.

But what is ‘entropy’?

When we talk about ‘entropy’ in the context of thermodynamics (the field that studies the relationship between heat, energy, and work), we always think of it as a measurement of ‘disorder,’ or ‘chaos.’

Therefore, this law states that an isolated system’s entropy, or the amount of disorder, always tends to increase (or stay constant).

But what does that even mean? And why entropy always tends to increase?

Imagine you have a puzzle, with all individual pieces separated and mixed inside a box. For a puzzle to have low entropy or low disorder, all pieces must be precisely positioned so that the puzzle is solved, as we have exactly one unique correct configuration of the pieces.

On the other hand, there are potentially infinite ways to assemble the pieces while simultaneously being an incorrect puzzle solution.

Simply put, chaos is more probable than order.

To drive this home, let’s say we close our eyes and try to solve the puzzle. Here, solving it becomes a game of probabilities. If our chances are good enough, by trial and error, we will eventually arrange the pieces in a way that solves the puzzle.

However, this is far from possible, as the number of wrong solutions is infinite compared to the number of correct solutions (one). Again, chaos is more probable than order. And most processes in nature behave exactly the same way.

Therefore:

As high-entropy states (disordered states) are much more statistically probable, all systems tend to high-entropy configurations, proving the second law of thermodynamics.

This law also has a second consequence fundamental to understanding reversible computing: irreversibility.

Not only do most processes tend to high-entropy states, but these processes are most often irreversible. For instance, if we mix sugar with our cup of coffee, we move into high entropy (all the particles of both elements are now mixed). Now, try to reverse that process by separating the sugar from the coffee.

You can’t, right?

Well, in fact, you theoretically could, but the likelihood of you achieving such a miraculous reversion is so unlikely that the process is considered irreversible, as the chances of arriving at the low-entropy configuration (managing to separate every sugar particle from every coffee particle) are so statistically remote compared to all other high-entropy possibilities (remaining mixed) that the system is almost guaranteed to remain in a high-entropy state for good.

All things considered, what the second law of thermodynamics implies is that most processes in nature are irreversible as they tend to high-entropy states (much more statistically probable), meaning that entropy never decreases, it can only remain constant or ever-growing.

Fine, but what does all this have to do with reversible computing a nd energy efficiency?

Circumventing Clausius

Knowing all this, the question arises: What if we can decrease a system’s entropy? Well, this law never said we couldn’t; it simply states that it happens at a cost.

Nothing is Free

In other words, if you force an isolated system’s entropy to decrease, entropy is increasing elsewhere because, as the second law enforces, the entropy of the overall system always increases.

In most cases, this means that if work is done to force a system into a statistically improbable state, entropy is released into the environment, usually through heat, to not violate the second law (the entropy of the overall system cannot decrease).

If we think about a refrigerator, it cools items below the environment’s temperature, creating a highly improbable state that would not occur naturally.

This is achieved by performing work, using a compressor that extracts heat from the enclosure and releases it into the room. This demonstrates what we just said: forcing systems into improbable states (low entropy) requires work, which releases energy into the environment (the environment around the refrigerator actually heats up. As a result, the entropy of the whole system — comprising the refrigerator and its surroundings — still increases.

Now, finally, we can talk about what all this means to computers and AI.

The Erasure Problem

Computers use circuits comprised of logical gates that allow them to perform calculations, such as the XOR Gate below, which outputs a one if both inputs differ.

These calculations are irreversible, as the inputs aren’t stored in memory (which they aren’t because that would be too taxing on the limited computer memory).

Thus, without the inputs, one cannot reconstruct the original calculation just with the output (using XOR’s example, if the output is one, we know A or B are one, but we don’t know which), making this an irreversible calculation.

As it’s irreversible, we need to erase the previous result to perform new calculations. This directly links with entropy: We have automatically decreased entropy by adding order to the system (reset for new computations), which has the byproduct of information loss.

However, as you may have guessed by now, this implies that, in the overall system, entropy must “appear” elsewhere so that the overall system’s entropy does not decrease. Thus, this erasure causes heat to be dissipated into the environment to compensate the localized entropy decrease.

This is the main reason computers heat up (while also factoring in hardware inefficiencies).

This brings us to the idea of reversible computing, in which we store some of the inputs to reverse the process, preventing information loss from erasure and, thus, minimizing heat dissipation.

As you can see below, by storing one of the XOR inputs, we can reconstruct the original state with that input and the output, making this computation theoretically reversible.