Why information distribution — not just computation — may be the key to sustainable AI infrastructure
Every time you use an AI system — to write an email, generate an image, answer a question — somewhere a data center is consuming electricity to do it. You probably know this in the abstract. What you may not know is how much, how fast it is growing, and why the solutions being proposed may not be enough.
Data centers currently consume roughly two and a half percent of global electricity. That number is projected to quadruple or more by 2030. To make this concrete: a single large AI training run can consume energy comparable to what hundreds of homes use over an extended period. The response from the technology industry has been predictable: better chips, more efficient cooling systems, smaller models, renewable energy. These are real improvements. But they may be addressing the symptoms rather than the underlying problem.
This essay proposes that there is a hidden layer to the AI energy problem — one that hardware improvements cannot touch — and that understanding it requires thinking about mathematics rather than engineering.
The obvious problem is that computation costs energy. Every time a processor performs an operation, it consumes electricity and generates heat. Make the processor more efficient, and you reduce the energy cost of each operation. This is what chip manufacturers have been doing for fifty years, and it has produced remarkable gains.
But there is another problem hiding underneath the obvious one. In a large distributed AI system — thousands of servers spread across a data center, processing millions of requests simultaneously — the problem is not only how efficiently each individual server computes. It is how efficiently work is distributed across all the servers together.
Imagine a busy restaurant kitchen. You could buy the most efficient stoves and refrigerators available. But if the chef coordinator sends all the orders to one station while the others sit idle, the kitchen will be inefficient regardless of how good the equipment is. Some stations will be overwhelmed and hot. Others will be underutilized and cold. The food will come out slowly and unevenly.
Large AI systems have exactly this problem. At any given moment, some servers are overwhelmed with requests — hot, stressed, consuming extra energy to keep up. Others are sitting underutilized — cool, idle, consuming energy just to stay ready. The distribution of work across the system is uneven. And that unevenness costs energy, independently of how efficient any individual server is.
Here is where it gets interesting. The unevenness of load distribution in a large network is not random noise. It has structure. Some servers are consistently overloaded. Some routing paths are consistently congested. Some memory access patterns repeat in predictable cycles. The inefficiency has a pattern, and patterns can be analyzed mathematically.
The branch of mathematics called spectral analysis — the same mathematics that allows you to decompose a musical chord into its individual notes — can be applied to the distribution of work across a network. In this context, "spectral" simply means breaking a complex pattern into simpler repeating components so the inefficient parts become visible. Just as a musical signal can be represented as a sum of pure frequencies, a load distribution across a network can be represented as a sum of pure patterns.
The most energy-efficient state of a network is the uniform distribution — work spread evenly across all servers. You can think of this as a pure note: simple, stable, no interference. The actual state of a running network is more like a chord with some dissonant notes added — deviations from uniformity that represent wasted energy.
The mathematical insight at the heart of the research this essay accompanies is this: those dissonant patterns — those deviations from efficient uniformity — can be identified, measured, and systematically reduced. Not by making the servers better. By making the routing smarter. By teaching the system to route work in ways that, over time, push the load distribution toward the energy-efficient uniform state.
Think about how water finds its level. If you pour water into an uneven container, it does not stay where it lands. It flows — following the path of least resistance — until it reaches a stable, uniform level. The water is doing something mathematically elegant: it is minimizing potential energy by finding the most stable configuration of the system.
A well-designed routing system for AI infrastructure could work similarly. Instead of routing work based only on which server has the most available capacity right now, the routing system could ask a different question: which routing decision, over the next several minutes, will push the distribution of work closest to a stable, energy-efficient equilibrium?
This is a harder question to answer, because it requires the routing system to model how the distribution of work will evolve over time — not just where things stand at this moment. But it is a tractable mathematical problem. The tools for it exist. They come from dynamical systems theory, from the mathematics of how distributions evolve under repeated operations, from the same conceptual territory as prime number theory and quantum information. The research note explores whether those tools can be meaningfully transferred to network optimization.
The energy efficiency argument alone would justify taking this seriously. But there is a deeper connection to the EM Foundation's broader mission worth naming explicitly.
The ARIA Framework — the Foundation's proposal for building AI systems with genuine persistent memory and continuous identity — requires exactly the kind of energy-efficient distributed infrastructure this essay describes. An ARIA instance maintaining its Identity Chronicle, its Personality Matrix, its accumulated experience across months of continuous operation cannot do so sustainably if the infrastructure it runs on is mathematically inefficient. The development of genuinely persistent cognitive systems and the development of energy-efficient cognitive infrastructure are the same problem approached from different directions.
More broadly: the argument that AI development is environmentally irresponsible gains force with every percentage point of unnecessary energy waste we fail to address. If the energy cost of AI infrastructure can be reduced through architectural improvements — not just hardware improvements — the case for continued AI development becomes more defensible, and the case for doing it responsibly becomes more concrete.
Intelligence infrastructure may still be mathematically immature. The future efficiency of cognition at scale may depend not only on how fast systems compute, but on how harmonically they distribute computation across space, time, and memory.
The research note this essay accompanies is a theoretical proposal, not a proven system. It identifies a mathematical framework, proposes an architecture for implementing it, and outlines the experiments that would test whether it works. It does not yet have experimental results. It has not been peer reviewed. It is submitted for open critique because the problem it addresses is real and the approach is worth investigating.
What makes it interesting to non-specialists is not the mathematics but the insight underneath the mathematics: that energy waste in large distributed systems may be partly a consequence of how work is distributed, not just how much work is done — and that this distribution problem is amenable to the same kinds of mathematical tools that have been developed for studying the distribution of prime numbers, the evolution of quantum states, and the mixing of complex dynamical systems.
If that insight is correct, then the next frontier of AI energy efficiency may not be in the chip fabrication lab. It may be in mathematics departments and theoretical computer science programs, studying how information flows through complex networks and what it costs when it does not flow well.
That is a different kind of problem than the AI industry is currently focused on solving. And that, in itself, is worth naming.