Artificial intelligence has an energy problem, but describing it only as a demand for more electricity may obscure a more interesting question.
As AI systems expand, the infrastructure supporting them requires electricity, cooling equipment, networks, processors, water in some facilities, and increasingly large data centers. Much of the current conversation understandably focuses on supplying those resources more cleanly and using them more efficiently.
Renewable energy, efficient cooling, and more efficient hardware are important. Though, there is another layer worth exploring. Why does computation use resources in the ways it does in the first place?
Some of the energy consumed by computing systems is not spent on the calculation we ultimately care about. Energy is also used moving data between memory and processors, keeping hardware available, transmitting information across networks, cooling equipment, and supporting the infrastructure surrounding computation.
That makes AI’s resource problem partly a question of architecture. If we want computation to become substantially more sustainable, perhaps we should look not only at how its energy is generated, but at how information moves, when computation occurs, where it occurs, and what happens to the resources and outputs involved.
Living systems offer an interesting place to look for different design principles.
Life Is Not a Closed System
It is tempting to describe nature as perfectly circular: everything recycled, nothing wasted, each organism somehow generating what it needs internally. That isn’t quite how life works.
Living systems depend continuously on energy from outside themselves. Plants capture solar energy. Animals obtain chemical energy through food. Organisms exchange matter with their environments and release heat as they metabolize energy. Life does not escape thermodynamics.
What living systems often do remarkably well is operate within energetic constraints. Activity changes with conditions. Resources move through networks. Some processes happen continuously while others activate only when necessary. Matter is reused through larger ecological cycles. Organisms regulate temperature, maintain reserves, respond to signals, repair damage, and allocate resources differently depending on need.
A forest and a computer are not equivalent systems, and biological metaphors become misleading when taken too literally. Still, billions of years of evolution have produced countless examples of systems functioning under constraints of energy, material, space, and time. That makes biology interesting as a source of questions rather than a blueprint.
What would computation look like if resource availability were treated as a fundamental design condition rather than something infrastructure must continuously provide around it?
Computation Does Not Always Need to Be Constant
Biological nervous systems do not process every possible signal at maximum intensity all the time.
Neurons communicate through changes in activity. Sensory systems prioritize some information while filtering enormous amounts of other input. Attention shifts. Metabolic resources are distributed differently depending on what an organism is doing.
Some emerging computing architectures already explore related principles. Neuromorphic computing, for example, takes inspiration from neural systems and often uses event-driven processing. Instead of performing the same kind of continuous operation regardless of whether relevant information has changed, computation can be triggered by events.
Event-based sensors apply a similar idea to perception. A conventional camera repeatedly records complete frames. An event camera can instead report changes in brightness at individual pixels as those changes occur.
These systems are not tiny artificial brains, nor are they universally more efficient for every task. Their advantages depend on the workload. What they demonstrate is more fundamental: continuous computation is not the only architecture available to us. Sometimes the efficient response to nothing happening is to do less.
Moving Information Has a Cost
Another major source of computational energy use is surprisingly mundane, moving information around. Traditional computer architectures often separate memory from processing. Data must travel repeatedly between locations while calculations are performed. As models and datasets grow, that movement becomes increasingly significant.
This has led researchers to explore approaches such as in-memory and near-memory computing, where some computation occurs closer to where data is stored.
The biological comparison is imperfect, but useful. Brains do not contain a central processor repeatedly retrieving every piece of information from a separate memory warehouse. Memory and processing emerge through deeply interconnected networks. Again, the lesson is not that computers should become brains. It is that architecture determines resource use.
If information repeatedly travels enormous distances within a system because that is how the system was historically designed, improving the efficiency of each trip may not be the only option. We can also ask whether the trip was necessary.
What Needs to Happen Where?
The same question extends beyond individual chips. Modern AI increasingly exists across several layers of computation. Some processing happens on personal devices. Some happens at the network edge. Large workloads may be sent to centralized data centers.
Each arrangement has advantages. Centralized infrastructure can provide enormous computational capacity and allow expensive hardware to be shared. Local computation can reduce network traffic, improve responsiveness, preserve some forms of privacy, and allow smaller tasks to occur without sending every interaction to distant infrastructure.
The most sustainable architecture may not be entirely centralized or entirely distributed. It may be adaptive. A small local model might handle an ordinary task while a larger remote system is used only when additional capability is genuinely necessary. Frequently needed information might remain close to where it is used. Expensive computation could be reserved for problems that actually require it.
This resembles a principle found throughout living systems: resources are not necessarily distributed equally at every moment. They are allocated in response to conditions. For AI, that raises a deceptively simple design question. How much computation does this particular problem actually need?
Memory, Context, and the Cost of Repetition
There is another resource question that receives less public attention. AI systems frequently reconstruct context. Information is retrieved, transmitted, processed, summarized, or supplied again so that a model can understand what is relevant to the current interaction. In long-running systems, maintaining useful continuity without repeatedly processing unnecessary information becomes both a design problem and a resource problem.
More context can improve performance, but more is not always better. A system that can identify what information matters now may not need to process everything it has ever encountered. Memory architecture, retrieval, caching, persistent state, model routing, and selective context all become part of the sustainability conversation.
The question is similar to the one we encounter elsewhere. What genuinely needs to move through the system each time, and what can remain where it already is?
Efficiency may come not only from making computation cheaper, but from avoiding computation that does not need to happen.
Heat Is an Output, Not Just a Problem
Whatever computation occurs eventually produces heat. Processors transform electrical energy, and some of that energy leaves the computing system as thermal energy. Data centers therefore require thermal-management systems to keep equipment within operating temperatures.
There are many ways to improve this process, including liquid cooling, direct-to-chip cooling, immersion systems, improved facility design, and more efficient heat transfer.
It is tempting to go one step further and imagine simply turning all of that heat back into electricity. Thermoelectric technologies can convert temperature differences into electrical energy, but recovering useful electricity from relatively low-temperature waste heat can be difficult. The amount of recoverable energy depends heavily on temperature differences, system design, economics, and the available technology.
This points toward a broader lesson about circular systems.= Returning something does not necessarily mean turning it back into exactly what it was before. Sometimes the better question is whether the output remains useful in its current form.
Heat that is inefficient to convert back into electricity may still be useful for buildings, industrial processes, district heating, or controlled-environment agriculture under appropriate conditions.
That is the question we explored in What If Data Centers Could Give Something Back?: one system may produce something another system needs.
At the scale of computation, the same principle encourages us to follow every resource beyond the point where the primary task considers it finished.
A Metabolic Lens for Computing
Our earlier work at Resofield used the phrase metabolic intelligence to describe the possibility of computing systems that regulate resources more like living systems. Computers do not metabolize in the biological sense. AI does not need food, heal tissue, maintain cellular homeostasis, or participate in ecosystems in the way organisms do.
A metabolic lens can nevertheless help us ask useful questions about flows. What enters the system? Where does it travel? What work does it perform? What does that movement cost? Which processes need to remain active? Which can become responsive to demand? What outputs remain useful? What must be dissipated? What can enter another system rather than terminating as waste? How should activity change when conditions change?
Seen this way, sustainable computation becomes more than a problem of reducing watts per operation. It becomes a problem of resource relationships.
Sustainability Across Nested Systems
No single technological breakthrough is likely to solve AI’s environmental footprint. The opportunity exists across several nested layers. At the chip level, architectures can reduce unnecessary data movement and explore more selective forms of processing. At the model level, developers can investigate smaller models, efficient architectures, quantization, routing, sparsity, and other methods that reduce the computation required for useful results. At the device level, appropriate tasks can happen locally rather than automatically traveling to centralized infrastructure. At the data-center level, cooling, workload scheduling, energy sourcing, water use, heat recovery, and facility design can reduce environmental pressure.
Beyond the data center, infrastructure can be planned in relationship with electrical grids, water systems, communities, land use, and potential uses for recovered heat. Each layer changes the possibilities available to the next.
This is why improving one efficiency metric is useful but insufficient as a complete strategy. A highly efficient chip can still operate inside poorly designed infrastructure. A sustainable facility can still host unnecessarily expensive computation. A locally efficient system can still shift environmental costs elsewhere. The whole chain is important.
What Computation Can Learn From Life
Nature does not provide a perfect engineering manual. Evolution produces redundancy, competition, failure, excess, and strange compromises alongside extraordinary efficiency. Living systems optimize for survival and reproduction under particular conditions, not for the goals of human technology. That makes biomimicry most useful when it inspires investigation rather than reverence.
Living systems invite us to notice certain possibilities, selective activity rather than constant activity, distributed processes alongside centralized coordination, responsiveness to changing conditions, resource sharing, feedback, repair, redundancy, and relationships in which the output of one process becomes input to another.
Some of those principles may translate beautifully into computation while others may not. Finding the boundary is part of the work.
Rethinking the Energy Problem
Perhaps sustainable AI will not emerge from a single dramatic invention. It may develop through many changes occurring at once: chips that move less data, models that use computation more selectively, memory systems that avoid unnecessary repetition, devices that handle appropriate work locally, data centers that recover useful outputs, electrical grids that provide cleaner energy, and infrastructure designed with the surrounding community and environment in mind.
None of this makes artificial intelligence a living organism, though it does suggest a different design philosophy. Use what is necessary. Move information intentionally. Respond to conditions. Preserve useful state. Recover what remains valuable. Understand where outputs go. Consider the larger system before optimizing one component inside it.
The energy question surrounding AI is therefore larger than how we generate enough electricity to support increasing computation.
We can also ask whether every computation needs to happen, whether it needs to happen there, whether it needs to happen now, and whether the architecture surrounding it was designed for the resource constraints we actually face.
Those questions ask us to reconsider the system itself. At Resofield, this is part of what we mean by The Return Phase, following resources and information beyond the point where a conventional process considers them finished, then asking whether another relationship is possible.
AI will always require energy. The more interesting challenge may be learning how to build computation that understands its place within the larger energetic systems that sustain it.
Leave a Reply