Thermodynamic Limits of Physical AI: Paper “Thermodynamic Limits of Physical Intelligence” Published

Koichi Takahashi (ALIGN, RIKEN) and Yusuke Hayashi (ALIGN) have published a new paper on the energy efficiency of physical intelligence (arXiv:2602.05463). This blog post explains the contents of that paper.

Introduction: Measuring the “Fuel Efficiency” of Intelligence

Large AI models such as ChatGPT, Gemini, and Claude consume massive amounts of electric power while demonstrating astonishing capabilities. Training cutting-edge models requires tens of megawatt-hours of power, raising societal interest in the “AI power issue.” In contrast, the human brain achieves advanced thinking on only around 20 watts—roughly equivalent to a single LED light bulb.

What does this orders-of-magnitude gap mean? What, in the first place, is the “energy efficiency of intelligence”? Just as automobiles have fuel efficiency (km/L or mpg), can we define a metric like “fuel efficiency” for intelligence?

This paper provides an answer grounded in the fundamental laws of physics. It proposes quantifying the efficiency of intelligence through a “bits per joule (bits/J)” metric, representing how many bits of information can be processed per joule of energy.

What Is “Physical Intelligence”?

This paper introduces the concept of “Physical Intelligence.” This refers to all intelligent systems that possess a physical embodiment and operate while consuming energy.

Physical intelligence encompasses both AI and biological intelligence:

  • Artificial Intelligence (AI): Language models running in data centers, robots, autonomous vehicles, etc.
  • Biological Intelligence: The human brain, animal nervous systems (and hypothetically, extraterrestrial intelligence)

In the AI industry in recent years, the term “physical AI” has mainly been used to refer to embodied AI that directly interacts with the physical world, such as robotics and self-driving cars. Physical AI in this sense is, of course, part of physical intelligence. However, the scope of this paper is broader: even a cloud-hosted language model is regarded as physical intelligence as long as it operates on physical servers, consuming power and generating heat.

Why is such a broad perspective necessary? Because no matter what form intelligence takes, as long as it operates in reality, it cannot escape the constraints of physical laws—thermodynamics in particular. This paper explores the ultimate limits of energy efficiency from this viewpoint of “intelligence as a physical entity.”

Paper Contents

Two Facets of Intelligence: “Learning” and “Acting”

The core of this paper lies in capturing physical intelligence along two complementary axes.

Perception and Learning: The Ability to Understand the World

An intelligent system learns the structure and regularities of the world through interaction with its environment. For example, language models learn grammar and knowledge from massive text corpora, and robots learn environmental physics from sensor inputs.

The paper captures this “ability to understand the world” using the concept of epiplexity. Epiplexity is a metric that measures how much “structural information” in data—information reflecting the regularities of the environment, excluding noise and redundancy—is encoded in the system’s internal state. This concept was formulated by Finzi et al. (2026) based on computational-resource-bounded Minimum Description Length (MDL) theory (arXiv:2601.03220). Unlike conventional information measures or entropy, it aims to extract only “meaningful structure” under constraints on compute.

Connecting this epiplexity to thermodynamics, the paper defines thermodynamic epiplexity per joule as the energy efficiency of learning. Put simply, it is an indicator of “how much you can learn about the world using 1 joule of energy.”

Control and Action: The Ability to Influence the World

Intelligent systems not only learn, but can also act upon the world. Examples include a robot moving objects or an AI assisting in decision-making.

The paper captures this “ability to influence the world” using the concept of empowerment. This measures “how diversely an agent can change future states through its own actions,” defined as the “channel capacity” in information theory. For instance, if a robotic arm can place an object into 10 distinguishable positions, its empowerment is log₂(10) ≈ 3.3 bits. If an action exerts no influence, it is 0 bits; the greater the influence, the higher the bit count.

Note: Regarding the theoretical background of empowerment, Hayashi and Takahashi (2025) showed that Bayes-optimal universal AI agents tend to maximize empowerment (ALIGN blog commentary).

The paper defines empowerment per joule as the energy efficiency of control. What is crucial here is proper energy cost “accounting.” For instance, even if an agent merely “waits without doing anything,” if the environment changes autonomously, an apparent information transfer can arise, potentially causing a “free control” artifact where infinite efficiency is achieved at zero cost. To avoid such issues, the paper proposes reporting the baseline energy of the “do-nothing” state separately from the additional energy consumed for control.

Why Two Axes Are Necessary

Learning and control are the two wheels of intelligence. No matter how well you understand the world, if you cannot act, you cannot change anything. Conversely, even with the ability to act, without understanding the world you cannot select appropriate actions. By evaluating these two separately, one can grasp system characteristics such as “efficient at learning, but inefficient at control.” Furthermore, in closed-loop agents (systems that alternate between learning and acting), both share an energy budget, enabling analysis of the trade-offs involved.

The Ultimate Limits Imposed by Physical Laws

What Is the Landauer Limit?

There is a fundamental lower bound on computer power consumption dictated by the laws of physics. According to the principle demonstrated by Rolf Landauer in 1961, erasing 1 bit of information irreversibly requires at least k_B T ln 2 joules of energy (where k_B is the Boltzmann constant and T is temperature). At room temperature, this is approximately 3×10^-21 joules, meaning the theoretical limit is roughly 3.5×10^20 bits/J.

The switching energy of modern computers is around 10^-17 to 10^-18 J, leaving about 3 to 4 orders of magnitude of room before reaching the Landauer limit. On the other hand, according to some estimates, the human brain may be only 2 to 3 orders of magnitude away from the Landauer limit (though specific figures depend on the estimation method). If these estimates hold true, it suggests just how efficient a form of physical intelligence biological evolution has produced.

Limits on Learning Efficiency in Closed Cycles

An important theoretical contribution of the paper is clarifying the conditions under which the Landauer limit applies to learning efficiency, using insights from stochastic thermodynamics.

Specifically, under “closed-cycle” conditions where finite memory is reused repeatedly, acquiring new knowledge requires erasing old information, at which point the Landauer limit comes into play. The paper mathematically derives the upper bound on learning efficiency under these conditions.

The Role of Quantum and Reversible Computing

Can quantum computing or reversible computing break through this limit? The paper presents a clear conclusion on this point.

Quantum computing is not a means to “violate” thermodynamic limits. The advantages of quantum computing lie in (i) being able to execute the vast majority of computation along reversible (unitary) trajectories, and (ii) reducing the number of logical steps required for specific problem classes. However, when the accounting boundary includes the reset/refresh steps required for qubit initialization, measurement, and fault tolerance, the Landauer limit still applies.

In other words, quantum computing can approach thermodynamic limits, but it cannot break through them.

The Importance of “Accounting”

There is, however, an important caveat here. If initialized memory (akin to a pristine hard drive) is supplied unconstrained from “outside” the system, the apparent efficiency could be infinite unless the cost of that supply is accounted for (Proposition 1 in the paper). The paper shows that with a simple setup of copying environmental information onto externally supplied blank memory via XOR (exclusive OR) operations, one can acquire information while making the energy consumed within the boundary arbitrarily small. Of course, the cost of preparing that “blank memory” is paid somewhere—it simply was not accounted for.

This is analogous to financial accounting. When evaluating a company’s profits, if you do not account for raw material costs, apparent profits will look massive, but that does not reflect reality. Similarly, when evaluating the energy efficiency of physical intelligence, it is essential to make explicit what is included in the computation and what is omitted. The paper clarifies how to set this “accounting boundary” and provides guidelines for fair comparisons.

A Unified Framework for Practical Use

Checklist of Items to Report

To fairly compare the energy efficiency of different systems, the paper recommends reporting the following items:

  • Accounting boundary: What range of energy is included, such as compute, sensors, actuators, cooling, etc.
  • Baseline: How the power consumption of the “do-nothing” state is handled
  • Resolution: The granularity when measuring information quantity
  • Time horizon: The evaluation period and how resets are handled
  • Estimation method: How the information quantity was calculated or estimated

Making these explicit enables comparisons among researchers and allows objective tracking of technological progress.

Relationship to Scaling Laws

In current AI development, “scaling laws”—where increasing model size, data, and compute improves performance—are widely recognized (see also this article (Part 1) for trends in the scaling race). Under this law, however, energy consumption per unit of performance gain increases at an accelerating rate.

The paper demonstrates how to analyze these scaling laws from the perspective of energy efficiency. The performance gain per additional unit of energy input (marginal efficiency) declines as scaling proceeds. This is the law of diminishing returns, suggesting that a different approach will eventually become necessary.


Implications and Impact

Quantitative Refinement of Intelligence Explosion Arguments: Connection to the 2018 Paper

This paper provides a more quantitative refinement of the arguments presented in the 2018 paper published in the Journal of the Japanese Society for Artificial Intelligence, “Scenarios and Branching Points for Future Machine Intelligence” (English translation).

The 2018 paper categorized future machine intelligence scenarios into four types and examined the factors determining the branching into each:

  • Upper-Bound Scenario: Fundamental limits exist on engineered intelligence, preventing it from reaching transformative levels.
  • Ecosystem Scenario: Multiple agents coexist as an interdependent network, exhibiting blurred boundaries and emergent behaviors (similar to internet infrastructure or biological ecosystems).
  • Multipolar Scenario: Multiple superintelligent agents with distinct boundaries exist in equilibrium, giving rise to strategic dynamics such as negotiation, deterrence, and arms races (similar to international relations).
  • Singleton Scenario: A single agent achieves a decisive strategic advantage.

Among these, the paper argued that in addition to cognitive architectures, physical constraints such as the thermodynamic efficiency of computation and the speed of light play crucial roles as conditions for the emergence of an intelligence explosion (rapid capability gain through recursive self-improvement).

Recursive self-improvement proceeds in two directions:

- Inward improvement: Computing more efficiently using existing resources (improvements in hardware and algorithms)

- Outward improvement: Acquiring more resources (compute, energy, spatial expansion)

Inward improvement is constrained by thermodynamic limits (i.e., the Landauer limit), while outward improvement is constrained by locality limits (coordination delays due to the speed of light). Physical barriers stand in both directions.

Regarding thermodynamic constraints in particular, while the 2018 paper pointed out that the Landauer limit could impose a fundamental ceiling on capability gains, it left open the possibility that quantum or reversible computing might bypass this limit. The 2018 paper stated: “This limit does not apply locally when using types of quantum computing where pure observation does not destroy quantum states, or when using reversible processes with molecular machines where information is not lost.”

The current paper develops thermodynamic constraints in a more rigorous formalism, reaching the conclusion that quantum computing can approach thermodynamic limits, but cannot exceed them. This clarifies that an in-principle upper bound exists for inward improvement.

(Locality constraints on outward improvement are discussed in detail in the 2018 paper.)

Intersections with AI Alignment Research

From the perspective of AI alignment research, understanding the limits of physical intelligence is also vital. As discussed in the 2018 paper, the feasibility of the Singleton scenario (where a single AI acquires a decisive strategic advantage) depends on what physical upper bounds exist on AI capability gains.

The thermodynamic limits demonstrated in this paper provide quantitative bounds on the pace and endpoint of an intelligence explosion. This serves as critical input when estimating worst-case scenarios in AI safety discussions.

Furthermore, the paper’s two-axis evaluation (learning and control) aligns with the conditions for acquiring a “decisive strategic advantage” discussed in the 2018 paper. A decisive strategic advantage requires both learning capability (predicting the environment and other agents) and control capability (exerting influence through action), each of which faces its own limits on energy efficiency.

Contributions to Sustainable AI Development

AI power consumption is surging rapidly. Some estimates suggest that AI-related power demand could reach several times its current level by 2030. Amid the need for climate change countermeasures, how to address the trade-off between AI advancement and environmental impact has become a major policy issue.

Currently, however, there is no standardized method for fairly comparing the energy efficiency of different AI systems. Even when someone says “this model is efficient,” the criteria for what constitutes efficiency remain vague. The efficiency metrics and reporting checklist proposed in this paper offer an answer to this problem.

Implications for Policy and Society

The framework in this paper provides a foundation for reporting the energy efficiency of AI systems in a standardized manner. This makes possible policy initiatives such as:

- Energy efficiency labeling for AI products and services

- Formulation of efficiency standards for data centers

- Prioritization of R&D investments

Just as home appliances have energy-saving labels, we can envision a future where AI services carry efficiency labels as well.


Conclusion

This paper has provided a rigorous, physics-grounded framework addressing the question of the “energy efficiency of physical intelligence.”

The proposed two-axis evaluation metrics—learning efficiency (epiplexity per joule) and control efficiency (empowerment per joule)—form a theoretical foundation for measuring the “fuel efficiency” of intelligence. Furthermore, the connection to the Landauer limit under closed-cycle conditions provides an ultimate benchmark for how far technology can advance.

The discussion on the conditions for the emergence of an intelligence explosion raised in the 2018 paper has been made more quantitatively rigorous by this work. In particular, the conclusion that no computing technology—including quantum computing—can bypass thermodynamic limits places clear constraints on debates regarding the upper bound of future machine intelligence capabilities.

Current AI systems remain orders of magnitude away from this theoretical limit. While this fact shows that there is vast room for improving energy efficiency, it also means there is an ultimate wall that cannot be breached by any technological innovation.

When considering the sustainable development of AI and the coexistence of AI and humanity, the perspective of physical intelligence’s energy efficiency will only grow in importance. We hope this paper serves as a common foundation for that discussion.


If you are interested, please take a look at the original paper.

Links:


This article is reprinted from the ALIGN Blog (February 6, 2026). Original text: https://www.aialign.net/blog/thermodynamic-limits

Koichi Takahashi

tomola is a place to exchange ideas across languages. Every post can be read in your language.

Log in to reply or mark a post as Insightful. Membership is by invitation. Log in