Information vs. Physics

Information Is Physical

Rolf Landauer proposed in 1961 that “information is physical,” proving that the irreversible erasure of information requires a minimum of k_BT·ln2 of energy to be dissipated as heat. In 2016, nanomagnetic bit-flip experiments confirmed this value at 144% of the Landauer minimum. Information processing cannot be independent of the laws of physics.

The prosperity of software and the internet over recent decades created the illusion that “information can expand infinitely.” AI has destroyed this illusion. AI’s algorithmic capabilities expand exponentially, but the physical substrate that drives them—electricity, semiconductors, cooling systems, rare earth elements—is supplied only linearly.

Theoretical Foundation: The Thermodynamic Tax on Information

2.1 The Landauer Principle and Maxwell’s Demon

Maxwell’s Demon paradox was the most famous challenge to the Second Law of Thermodynamics since 1867. Through Szilard (1929) and Bennett (1982), Landauer’s principle provided the final resolution: the demon must pay an entropy cost in acquiring and erasing molecular information. The minimum energy to erase 1 bit at room temperature (300K) is approximately 2.8 × 10⁻²¹ J.

2.2 Thermodynamic Implications in Modern Computing

The actual energy consumption of modern GPUs is approximately 6–7 orders of magnitude above the Landauer limit. However, as engineering optimization approaches its limits, the thermodynamic floor emerges as the ultimate barrier.

During LLM inference, the Transformer architecture performs matrix operations across all attention heads, and intermediate activations and attention scores generated and discarded are dissipated as heat. This is the direct realization of the thermodynamic tax on information processing.

Empirical Analysis I: The Physical Wall of Data Centers

3.1 The Exponential Growth of Energy Consumption

3.2 The Thermodynamic Implications of Cooling

Cooling can account for up to 40% of data center energy consumption (WEF, 2025). This is the macroscopic manifestation of the Landauer Principle: processing information generates heat, removing it requires additional energy, and there is no physical escape from this cycle. AI workloads are pushing rack power density from a traditional 4–10kW to over 100kW.

Empirical Analysis II: OOM — The Structural Physical Bottleneck of AI

4.1 The Nature of the OOM Problem

OOM (Out-Of-Memory) is not a simple software bug but a structural collision that occurs when information processing demand exceeds physical memory capacity. A large language model like Llama 3 70B requires approximately 140GB of GPU memory at half-precision (FP16), and a 128K-token context window’s KV cache consumes an additional ~40GB per user.

4.2 OOM and the Physical Impossibility of AGI

4.3 Personal-Scale Evidence: DGX Spark OOM Experiment

This constitutes first-person empirical data demonstrating that data-center-scale OOM problems are reproduced at the personal hardware scale by the same laws of physics.

Empirical Analysis III: Geopolitical Restructuring of Physical Assets

5.1 Rare Earth Elements and the Rise of Physical Possession Theory

China controls approximately 90% of global rare earth processing capacity. As of early 2026, no Western nation has achieved commercial-scale processing of medium and heavy rare earth elements.

5.2 Energy Infrastructure: The New Basis of Power

In the past, software patents, copyrights, and data platforms were the sources of power. In the future, power belongs to those who physically possess GPU hardware, semiconductor manufacturing capability, mineral resources, electricity, transformers, satellites, and submarine fiber-optic cables.

The Information-Physics Confrontation Model in Human Biology

6.1 Thermodynamic Cost Asymmetry of Cognitive Activity

The human cerebral cortex accounts for approximately 2% of body weight while consuming approximately 20% of total energy—a high-energy information processing device.

6.2 Biological Asymmetry of Inter-Individual Information Processing

AI as an external cognitive tool can partially compensate for this biological gap, but individuals who can use AI most effectively are those with inherently high information processing capability. Therefore, AI is more likely to widen the gap than narrow it.

The Fourth Industry: Cognitive Industry and Physical Friction Coefficient

7.1 From Production-Consumption to Collection-Supply

The Fourth Industry is a structure in which humans function as information collectors and suppliers, and large AI model companies function as information consumers. The value of AI-generated information is collapsing. What AI companies truly need is OOD (Out-Of-Distribution) human-origin information.

7.2 Information Valuation by Physical Friction Coefficient

7.3 Restructuring of Social Stratification

Polarization of Humanities and Restructuring of Education

Conclusion: The Era of Information-Physics Coupling

• First, information processing is not free. Physics levies a tax on every information operation from the ground floor (Landauer Principle).
• Second, AI is the apex of humanity’s information industry, but its expansion is constrained by the linear supply of physical resources—electricity, semiconductors, rare earth elements, cooling systems.
• Third, OOM is not a software bug but a structural physical bottleneck. Unless it is solved, AGI discourse has no physical foundation.
• Fourth, the era of Information Determinism is ending, and the era of Physical Possession Theory—where those who possess physical resources hold power—is arriving.
• Fifth, future value is determined by physical friction coefficient. In the Fourth Industry (Cognitive Industry), the upper stratum comprises those who produce high-friction source information through direct interaction with the physical world.
• Sixth, the relationship between information and physics is not conquest but coupling. True power belongs to the entity that simultaneously commands both information capability and physical resources, and continuously expands the coupling boundary between the two.

References

1. Davies, P. (2019). The Demon in the Machine. London: Allen Lane.
2. Landauer, R. (1961). “Irreversibility and Heat Generation in the Computing Process.” IBM J. Res. Dev., 5(3), 183–191.
3. Bennett, C.H. (1982). “The Thermodynamics of Computation.” Int. J. Theor. Phys., 21(12), 905–940.
4. Bérut, A. et al. (2012). “Experimental verification of Landauer’s erasure principle.” Nature, 483, 187–189.
5. Hong, J. et al. (2016). “Experimental test of Landauer’s principle in single-bit operations on nanomagnetic memory bits.” Science Advances, 2(3), e1501492.
6. Gartner (2025). “Electricity Demand for Data Centers to Grow 16% in 2025 and Double by 2030.”
7. Lawrence Berkeley National Laboratory (2024). 2024 Report on U.S. Data Center Energy Use. DOE.
8. International Energy Agency (2025). Energy and AI. IEA, Paris.
9. BloombergNEF / BCSE (2026). 2026 Sustainable Energy in America Factbook.
10. Li, W. et al. (2026). “Out of the Memory Barrier.” arXiv:2602.02108.
11. Tang, Y. et al. (2025). “Training large-scale language models with limited GPU memory.” Frontiers of IT & EE, 26, 309–331.
12. Shannon, C.E. (1948). “A Mathematical Theory of Communication.” Bell Syst. Tech. J., 27(3), 379–423.
13. Szilard, L. (1929). “On the decrease of entropy.” Z. Physik, 53, 840–856.