The Illusion of Immaterial Intelligence
Artificial intelligence feels like pure software. You type a prompt, an answer appears, and the experience seems detached from the physical world. That perception is precisely why the headlines about AI consuming enormous energy and water feel shocking. But AI is not “immaterial intelligence.” It is industrial-scale computation. Every output is the result of electricity being converted into heat inside physical hardware, operating in buildings that must remove that heat continuously. The controversy exists because AI has crossed a threshold where its footprint is no longer a rounding error inside the broader economy, and because its growth is faster than the infrastructure and policies designed to manage it.
Why AI Uses So Much Electricity: Chips, Scale, and Constant Demand
The primary driver of AI electricity consumption is the hardware. Modern AI is trained and served on GPUs and specialized accelerators that draw hundreds of watts per chip. At the level of a single machine this seems manageable. At the scale of thousands to tens of thousands of chips running around the clock, it becomes a permanent industrial load on electrical grids.
Training large models makes the issue visible because it can take weeks or months of continuous high utilization, and because computation does not scale linearly with model size. As models become larger, the compute requirements often grow much faster than the size alone suggests. However, the persistent cost is frequently inference rather than training. Even after a model is trained, serving millions of users means that every query triggers heavy real-time matrix operations. Popular systems cannot “sleep.” They must be available constantly, across global time zones, with bursts of demand that push systems toward always-on capacity.
The deeper structural reason AI burns so much electricity is that most modern computing architectures waste energy moving data. Memory and computation are separated. Enormous energy is spent shuttling values between storage and processing rather than doing arithmetic. Nearly all electricity used eventually becomes heat. That heat then becomes a second-order problem that drives additional energy use.
Why AI Uses Water: Cooling Is the Hidden Bill
AI does not drink water. Data centers use water because cooling high-density compute is unavoidable, and water is one of the most efficient, scalable coolants available. GPUs generate far more heat per rack than traditional servers—often by an order of magnitude or more—so cooling requirements rise sharply as AI deployment increases.
Many facilities rely on evaporative cooling, cooling towers, or water-based loops that move heat away from hot components. When scaled across millions or billions of daily interactions, the resulting water consumption can become substantial. The controversy intensifies when data centers operate in regions already under water stress, or when they compete with municipal or agricultural needs.
Headlines sometimes exaggerate by presenting simplified “per query” water numbers as if they were universal constants. Those numbers vary widely by data center design, local climate, and cooling method, and not all facilities use potable freshwater. Yet the core claim is still real: AI’s growth is increasing electricity demand, and cooling that electricity use can translate into large water footprints.
A key dynamic sits behind many sustainability debates: efficiency gains do not automatically reduce total consumption. When systems become cheaper per operation, people use them more. This rebound effect, commonly described as Jevons paradox, means that improved efficiency can coincide with rising total usage.
Why “Just Use the Sea” Is Harder Than It Sounds
A natural response is to ask why we do not use the sea as a near-infinite heat sink. In limited cases, seawater cooling can work. Some industrial systems already use ocean water through heat exchangers. But it is not widely adopted for AI data centers because the constraints are severe.
The first constraint is geography. Many hyperscale data centers are built inland, optimized for power availability, network connectivity, and land costs. Moving infrastructure to coastlines changes latency, shifts fiber requirements, and concentrates risk. The second constraint is materials. Seawater is corrosive. Saltwater systems degrade pipes, pumps, and exchangers rapidly unless expensive alloys and maintenance regimes are used. The third constraint is environmental regulation. Dumping warmed water back into the ocean can disrupt local ecosystems, and many jurisdictions tightly regulate thermal discharge. Finally, seawater cooling does not eliminate operational complexity: cleaning, treatment, and backup systems remain necessary. It can reduce freshwater stress in some places, but it does not globally solve the cooling problem.
Why Liquid Nitrogen Is Not a “Free Coolant,” but an Energy Trap
Liquid nitrogen sounds like an obvious solution because it is extremely cold. The physics, however, reverses the intuition. Liquid nitrogen is not a cheap cooling resource; it is energy-intensive to produce. Liquefying nitrogen requires compressing, cooling, and separating air using industrial processes that consume large amounts of electricity. You would often spend more energy producing and handling the nitrogen than you would save by cooling hardware with it.
At scale, additional issues appear. Nitrogen boils rapidly at ambient temperatures, implying continuous resupply. It introduces safety risks due to oxygen displacement. Materials can become brittle under cryogenic conditions. The overall system becomes expensive, complex, and poorly suited to continuous industrial operation. Liquid nitrogen is valuable for niche scientific and engineering applications, not for cooling vast AI compute fleets.
Why Space Is Not a Practical Heat Dump
Space seems cold, so it is tempting to imagine sending heat away “into space.” The key misunderstanding is that cooling is not about temperature alone. In a vacuum, you cannot remove heat through conduction or convection. You can only remove heat through radiation, which is far less effective at the temperatures and surface areas practical on Earth.
To radiate the heat produced by a large AI data center, you would require enormous radiator surface area—on the order of square kilometers—plus robust thermal coupling and structural support. Putting data centers in orbit makes the problem worse: the cost of launching hardware is extreme, repairs are difficult, radiation damages electronics, and power supply becomes a constraint. Space cooling is relevant for satellites because they have no alternative, but their heat loads are tiny compared with terrestrial AI computing.
The Most Radical Alternative: Brain Cells and “Wetware” Computing
If the goal is energy efficiency, the human brain becomes the most obvious comparison. It operates on roughly twenty watts while exhibiting capabilities that challenge even the largest AI systems. It runs on biochemical energy—glucose and oxygen—through electrochemical gradients, with memory and computation tightly integrated. From a pure energy perspective, biology appears unbeatable.
This has inspired real research into biological computing, neuron cultures, brain organoids, and hybrid neuron–silicon interfaces. In controlled environments, neuron cultures can be stimulated and measured, and simple adaptive behaviors can be induced. The concept is not science fiction.
Yet the obstacles are fundamental. Neurons are not logic gates. They are noisy, adaptive, and non-deterministic. You cannot precisely address and set states like you can with transistors. Two “identical” cultures behave differently and drift over time. Training methods like backpropagation and gradient descent do not map cleanly onto living tissue. The biggest bottleneck is input/output: reading from neurons requires electrodes or imaging; writing requires electrical, chemical, or optical stimulation. These interfaces are expensive, slow, and difficult to scale.
The “sugar is cheaper than electricity” idea also breaks when you include life-support requirements. Living computation demands sterility, temperature control, oxygenation, pH regulation, nutrient delivery, and waste removal. Maintaining these at scale resembles running pharmaceutical-grade bioreactors. The neurons may be efficient, but the system is fragile and operationally expensive. Biological computation could become useful in narrow niches—especially hybrid systems or ultra-low-power adaptive sensing—but it is not a realistic replacement for cloud-scale AI, nor for deterministic services that require repeatability.
A deeper reason sits beneath all of this: the brain’s efficiency comes with tradeoffs. It is slow, approximate, redundant, and embodied. Engineering often demands speed, deterministic behavior, and exact reproducibility. You cannot maximize all of these simultaneously.
Why Quantum Computing Does Not Solve AI’s Energy and Scale Problem
Quantum computing is another popular “future fix” people cite for AI energy concerns. In reality, it is misaligned with the core workload and imposes extreme operational requirements.
Many quantum platforms require millikelvin temperatures and dilution refrigeration. Maintaining those conditions can consume significant energy, often orders of magnitude more than the energy used by the quantum chip itself. More importantly, quantum computers are not general-purpose accelerators for everything. They are specialized machines that excel at narrow classes of problems such as certain factoring, quantum simulation, and specific optimization or sampling tasks.
Modern AI workloads are dominated by dense linear algebra and deterministic numerical computation. These map poorly to today’s quantum hardware. The practical barrier of error correction is enormous: one reliable logical qubit may require thousands of physical qubits plus continuous error-correction overhead and extensive classical control electronics. The gap between current quantum machines and what would be required for large-scale AI is not a small engineering jump. It is a very large structural gap. As a result, quantum computing is better understood as a scientific instrument for specific domains, not a path to replace GPUs for mainstream AI training and inference.
The One Alternative That Truly Helps: Photonic (Light-Based) Computing
Using light for AI computing is real and promising, but the correct framing is important. Photonic chips do not remove electricity. They replace certain operations with optical physics while still relying on electronics for control, memory, and general logic.
AI workloads are dominated by matrix multiplication. Photonics can perform forms of multiplication and addition through interference and phase shifts in waveguides and optical circuits, delivering massive parallelism with low resistive loss and reduced heat generation for the math itself. For specific operations, photonic accelerators can deliver significant improvements in throughput per watt.
However, photonic computing is typically analog, which introduces noise and precision challenges. It struggles with state storage and conditional logic. It also faces conversion overhead: data must be encoded into light and then converted back to electrical signals, and those conversions cost energy and add complexity. For these reasons, photonics is more suitable for inference than for training, and it is best deployed as a co-processor or accelerator rather than as a wholesale replacement for electronic computing. It is a major efficiency improvement in the right place, not a silver bullet that ends data centers.
The rebound effect still applies: if inference becomes cheaper, total usage often rises.
The Most Sustainable Direction: Decentralized, Connected Local AI
After exploring cooling tricks and exotic computing paradigms, the most plausible systemic solution is architectural rather than magical. Decentralized local AI resembles volunteer or grid computing from earlier decades—where idle compute was shared globally—except that today’s hardware is dramatically more capable and includes dedicated AI accelerators in consumer devices.
The sustainability advantage is profound. Local devices are already powered on, already cooled, and already geographically distributed. Using idle capacity avoids the concentration of heat that forces water-heavy cooling. It reduces the need to move data across networks, and it can keep sensitive data on-device, improving privacy and lowering bandwidth and energy costs.
This approach previously struggled because of heterogeneity, coordination overhead, unpredictability, and lack of incentives. Devices varied widely in architecture and availability, and many distributed tasks required tight synchronization. Additionally, participants paid the electricity bill and received little benefit, so the economic model was weak.
The landscape has changed. Many AI tasks—especially inference—are loosely coupled and parallel. Federated learning enables local training with centralized aggregation without uploading raw data. Incentive systems can be built through credits, reduced subscription costs, priority access, or reciprocal compute markets. Decentralization does not replace centralized frontier models entirely; it reduces their necessity by making most routine intelligence local.
There is a political economy dimension as well. Centralized AI centralizes data, control, and profit. Decentralized AI distributes capability and weakens lock-in. The resistance to decentralization is not always technical.
What We Should Expect Next, and What We Should Not
In the near future, we should expect optimization rather than revolution. Smaller, task-specific models will become mainstream. Hybrid architectures will dominate, where local AI handles routine tasks and centralized AI is called only when needed. Cooling improvements will continue, including more liquid cooling and better closed-loop systems, and waste-heat reuse will expand where infrastructure allows. Photonic accelerators will enter production in niche inference-heavy deployments. Regulation and transparency pressures will increase, forcing clearer accounting of energy and water footprints.
What we should not expect soon is equally important. GPUs will not be replaced outright. Quantum computing will not become a general AI engine. Biological computing will not scale into industrial AI infrastructure. Data centers will not disappear.
In the medium term, we should expect a more visible architectural shift. Decentralized and local AI will become economically attractive as centralized compute costs rise and privacy regulation pushes data locality. Neuromorphic hardware—brain-inspired but not biological—will appear in real products for low-power event-driven tasks, such as always-on sensing and robotics. AI usage will likely become tiered, resembling an energy market: local cheap intelligence for everyday tasks, regional shared intelligence for heavier workloads, and expensive centralized frontier models for rare, high-value applications. Regulatory frameworks will mature from vague principles into concrete reporting requirements and incentives that reward efficiency and penalize waste.
In the long term, AI will normalize into infrastructure. Intelligence will be embedded, contextual, and mostly invisible. Frontier models will exist but will be treated as strategic resources for science, engineering, medicine, and national-scale modeling rather than as default engines for trivial everyday usage. Society will increasingly accept the physical reality that intelligence has a cost, and the “bigger is always better” assumption will weaken under energy and environmental constraints.
The Pattern Under Everything We Discussed
Across every step of this path—electricity, water, seawater cooling, liquid nitrogen, space, brain cells, quantum computing, photonics, decentralization—a single pattern repeats. Centralized approaches eventually collide with physical limits. Exotic alternatives either fail thermodynamically, fail economically, or fail operationally at scale. The only direction that consistently reduces systemic pressure is distributing intelligence across many smaller nodes and reserving centralized compute for tasks that truly require it.
The AI energy debate is not fundamentally about whether we can invent a clever coolant or a magical chip. It is about whether we will keep building intelligence as an always-on centralized utility, or redesign it as a distributed capability that respects physical reality. The future will not be decided by one breakthrough. It will be decided by constraints forcing architecture to evolve—and by whether society accepts limits instead of assuming intelligence can grow without cost.
By Abu Adam Al-Kiswany