Does sustainable AI exist? What really makes a model more efficient?

If two artificial intelligences can perform the same task with similar quality, but one uses a fraction of the energy, memory, and infrastructure of the other, does it make sense to treat them as environmentally equivalent? Probably not. And this difference helps to understand why talking about "sustainable AI" requires more care than simply classifying certain models as green and others as polluting.

The environmental impact of artificial intelligence is not determined solely by the model's name or the size of the data center in which it operates. It results from a chain of factors involving... the requested task, the chosen model, the amount of computing power required, the hardware used, the efficiency of the infrastructure, the processing time, the electrical matrix, the cooling system, and the physical location of that infrastructure..

This means that there isn't necessarily a single "green AI" as an absolute category. Above all, there are..., more or less efficient ways of performing a given task with artificial intelligence.. This distinction matters because the demand for computing continues to grow. International Energy Agency It projects strong expansion in electricity consumption by data centers throughout this decade, with AI among the main drivers of this growth.

The solution, therefore, cannot depend solely on building more infrastructure or adding renewable energy to the equation. We also need to ask whether we are utilizing [the current energy source]. more computational intelligence than the task actually requires.. This is perhaps one of the most important environmental questions of the next phase of artificial intelligence.

In summary, what makes AI more sustainable?

An AI tends to be more environmentally efficient when it can deliver the necessary result using less computing power and when that computing power occurs in an efficient infrastructure powered by a lower-impact energy matrix. This can involve smaller models, more efficient architectures, quantization, distillation, better hardware utilization, reduced idle processing, shorter answers when long answers do not add value, selective use of reasoning, more efficient data centers, and cooling strategies appropriate to the local context.

However, there is one important condition: Efficiency cannot be measured in isolation from quality.. A model that uses half the energy but cannot adequately perform the task is not necessarily the most sustainable choice if the work needs to be repeated multiple times or later transferred to a larger model.

Therefore, the question shouldn't just be "which AI consumes less?", but “"What is the minimum amount of computation capable of solving this problem with the required quality?"” This is the idea that we can call computational proportionality.

What does Green AI mean?

the term Green AI It gained academic traction from a simple discussion: for many years, the advancement of artificial intelligence systems was evaluated almost exclusively by performance improvements, while the computational cost required to achieve those improvements received far less attention.

In a study published in 2019, researchers proposed that efficiency should also be considered a relevant dimension in the evaluation of AI systems, alongside traditional performance metrics. The proposal included making computational and financial costs more transparent to incentivize models that achieve good results without simply relying on ever-increasing amounts of computing power.

This logic remains especially relevant today. A model should not be considered better simply because it manages to gain a few extra points in a given benchmark if, to do so, it requires dozens of times more energy in an application where that performance difference has little practical value. The goal is not to make efficiency the sole criterion, but to ensure that it ceases to be an invisible criterion.

Bigger doesn't always mean better.

There's an intuitive perception that the most powerful model available will also be the best choice for any given task. This logic is understandable: if a model demonstrates superior capabilities in general benchmarks, why choose another?

Because real-world tasks are not general benchmarks. An analysis conducted in 2025 The study compared different models on specific questions constructed from reports by the IPCC, the World Bank, and the World Health Organization. In one assessment, a considerably smaller model came relatively close to the performance of a much larger system, using a fraction of the energy. In other tasks, the differences between the most and least efficient models exceeded... 200 times.

These results pertain to the tasks, models, and configurations evaluated and should not be taken as a universal rule that small models will always be better or good enough. Final consumption also depends on the hardware, processing type, and deployment configuration.

Nevertheless, the practical conclusion remains relevant: Automatically selecting the largest available model can mean wasting computing power without producing a proportional improvement in the result.. Instead of asking “which is the best model?”, perhaps we should ask: Which model is most efficient and achieves the required quality level for this task?

Smaller models can do much more than they seem.

Recent advances are also changing what it means to be a "small" model. More efficient architectures allow compact systems to perform tasks that previously required far more resources.

THE Gemma 3n, from Google, Gemma 3n is an example of this direction. Its architecture uses techniques that reduce the memory footprint and allow execution on relatively modest devices for the capacity offered. This doesn't mean that Gemma 3n is automatically a "green AI," nor that running a model locally is always environmentally superior to using a cloud service. The example demonstrates something more important: architecture matters.

The sheer number of parameters is no longer sufficient to understand how many resources a system actually requires during its operation. Activation method, numerical precision, architecture, hardware, and memory can also significantly alter this equation.

Quantization, distillation, and other ways of doing more with less.

Much of the efficiency gains from AI come from techniques that are rarely visible to the end user. One of these is... quantization, This reduces the numerical precision used to represent parameters and perform certain calculations when this can be done without significantly compromising quality. In simpler terms, it questions whether all calculations really need to be performed with the same level of precision.

Another important technique is... distillation, In this context, a larger model can act as a "teacher" for a smaller system. The idea is to transfer some of the capabilities of the larger model to a more compact architecture, allowing certain routine tasks to be performed with fewer resources.

There are also architectures such as Mixture of Experts, In this model, not all parts need to be activated simultaneously for each request. Only specific specialized components are used as needed based on the input, reducing the amount of computation required compared to an architecture that would activate the entire system all the time.

There are also gains from the hardware's own utilization. Accelerators consume resources even when underutilized, and techniques such as batching, better workload distribution, and more efficient infrastructure management can increase the amount of work accomplished with the same installed capacity.

Efficiency, therefore, is not just within the model. It exists between model, software, hardware and infrastructure.

Reasoning improves responses – but it can also multiply consumption.

One of the most important changes in recent years has been the popularization of models capable of using additional computation during inference to solve more complex problems. This feature is often described as reasoning, or reasoning.

From a quality standpoint, it's an extremely relevant tool. A system that dedicates more processing steps to a mathematical, scientific, strategic, or programming problem can achieve results that would be difficult with an immediate response. From an energy standpoint, however, this additional computing power is not free.

THE AI Energy Score v2 The study compared models with and without reasoning and found very significant differences in consumption under the evaluated conditions. In some experiments, enabling this type of processing greatly increased the volume of computation and tokens generated.

This doesn't mean we should avoid reasoning. It means that perhaps we shouldn't use it. when there is no reason to reason so much. It makes no sense to mobilize a complex computational process to correct a comma, classify a simple text, or answer a basic factual question if a significantly more economical system can perform the same task with equivalent quality.

On the other hand, reducing computation in a complex scientific analysis, an engineering problem, or a high-risk application simply to consume less energy may be a false economy. Sustainability doesn't simply mean doing less; it means... Use the appropriate amount of resources for the importance and complexity of the problem..

How much do tokens matter?

Tokens are the units of information used by models during processing and language generation. In simplified terms, the more content that needs to be analyzed or produced, the greater the amount of computational work tends to be.

A short question that receives a short answer is not equivalent to providing dozens of documents for analysis and requesting an extensive investigation accompanied by thousands of tokens. This effect becomes even more important in reasoning systems, agents, and workflows that perform several steps internally.

This doesn't mean users should obsessively start counting every word they type for an AI. The relevant point is structural: Long and complex tasks have different costs than simple interactions., Well-designed systems can recognize this difference instead of sending all requests to the same amount of computation.

The future may lie in model routers.

Imagine a system that receives a request and first assesses its complexity. If the task is simple, it can be sent to a small model; if it requires more capacity, an intermediate system is used; and only truly difficult problems are forwarded to frontier models or intensive reasoning processes.

This approach is commonly called model routing. This is interesting because it shifts some of the responsibility away from the user. Instead of waiting for millions of people to understand parameters, GPUs, quantization, and energy benchmarks before asking a question, the infrastructure itself can automatically choose. the minimum amount of computation suitable for that problem.

It's computational proportionality transformed into architecture.

But the model's efficiency is only half the problem.

Even an extremely efficient model needs to operate somewhere, and that place matters. Two servers running exactly the same workload can produce different environmental impacts depending on the efficiency of the installation, the electricity used, the climate, the cooling system, and the water availability in the region.

Node article about AI's water consumption, we saw that a study by Lawrence Berkeley National Laboratory found variations greater than 10,000 times the water consumption per workload among different possible configurations. Among the main determinants were server efficiency, water intensity of the electrical grid, equipment utilization, type of cooling, infrastructure, and climate.

Therefore, An efficient model in an inefficient infrastructure does not necessarily equate to efficient AI.. It is necessary to observe the entire system.

The power grid changes the footprint of the same AI.

A kilowatt-hour used in two different locations does not necessarily have the same environmental footprint. The composition of the electricity grid matters because different matrices have different carbon intensities and may also be associated with different levels of water use.

This means that performing the same amount of computing on two different electrical systems can generate different impacts. It also means that stating that a particular service "uses renewable energy" needs to be interpreted carefully, since there is a difference between the electricity physically present in the grid at a given moment and the contractual mechanisms used by companies to acquire renewable energy.

To understand the impact of AI, therefore, Where and when computation happens can be as relevant as which model is being used..

Data centers are also becoming more efficient.

Physical infrastructure has its own race for efficiency.

THE Google reported in 2025 that a typical text prompt in Gemini Apps used, according to the company's methodology, approximately 0.24 Wh of electricity, 0.03 g of CO₂ equivalent, and 0.26 ml of water.. More interesting than the absolute number was the reported evolution: the company estimated a reduction of approximately 33 times the energy and 44 times the carbon footprint of the median prompt. in about twelve months, while claiming to have improved the quality of the responses.

These data need to be treated with due caution. They are corporate, represent a specific methodology, infrastructure, and time period, and should not be transformed into a universal measure of the impact of AI.

Nevertheless, they demonstrate something important: Efficiency per unit of service can change very quickly.. If a system can reduce the energy consumption required to complete a task by dozens of times while maintaining or improving its outcome, then innovation in efficiency has real environmental potential.

The problem is that there is another variable: scale.

If each prompt becomes more efficient, is the problem solved?

Not necessarily. Imagine that the energy required for a task drops by 80%, but the number of tasks performed increases tenfold. Total consumption may still increase.

That's why efficiency gains per unit don't guarantee an absolute reduction in demand. As AI becomes cheaper, faster, and more accessible, more users, more applications, more automations, more image and video generation, and more agents working continuously emerge.

This dynamic is close to what economics describes as rebound effectMore efficient technology reduces usage costs and, precisely for that reason, can be used on a larger scale.

This is a fundamental point for any serious discussion about "green AI": Efficiency matters, but scale matters too.. It's not enough to measure how much energy a model saved per query. We also need to look at how many new queries and applications came into existence because this computation became cheaper.

And what about water? Is it possible to cool data centers without continuous consumption through evaporation?

New technologies are attempting to reduce precisely this dependence.

THE Microsoft announced an architecture A data center with closed-loop liquid cooling eliminates the need for continuous water evaporation to remove heat from equipment. According to the company, this design can prevent more than... 125 million liters of water per year per data center, taking certain previous projects as a reference.

This doesn't mean the entire data center will have a "zero water footprint." Water can still be used for other purposes, the manufacturing of equipment has its own impacts, and the electricity consumed can also be associated with indirect water use.

Furthermore, there is an important trade-off: systems that reduce direct water consumption may, under certain conditions, use slightly more electricity than highly efficient evaporative technologies.

This example shows why sustainability cannot be reduced to a single metric. We can decrease water consumption and increase energy; decrease energy consumption and use a more carbon-intensive electricity grid; or reduce operational emissions while increasing impacts related to hardware manufacturing.

The goal needs to be to optimize. the system, It's not simply about producing the best possible number in a column of a report.

There's no such thing as a free lunch: sustainability is a matter of trade-offs.

It's tempting to look for a technology that definitively solves the environmental impact of AI: a small model, a more efficient GPU, solar power, liquid cooling, or a data center built in a specific region.

In practice, none of these decisions exist in isolation. Research on water consumption by workloads shows precisely that there is no single recipe capable of minimizing the impact in all situations. The result depends on the combination of server efficiency, electricity, usage, cooling, climate, and infrastructure characteristics.

This logic can be extended to computational sustainability more broadly. We don't need to find "sustainable AI." We need to build it. more sustainable decisions about AI.

A simple model of computational proportionality

The table below does not represent a physical measurement of consumption. It serves as a decision framework to show that different problems justify different levels of computation.

Task type	Typical complexity	Computational strategy	Advanced reasoning?	Relative intensity
Simple classification	Low	Small or specialized model	Usually not	Low
Short revision or rewrite	Low	Small or medium model	Usually not	Low
Document summary	Low average	Small or medium depending on size.	Limited	Low average
Content creation and analysis	Average	Intermediate model	When it adds value	Average
Complex programming	Medium to high	Capable model + tools	Often	Medium to high
Multi-stage research	High	Advanced model + search and tools	Yes	High
Complex scientific analysis	High	Advanced model + validation	Yes	High
long-term self-employed agent	Very high	Multiple calls and tools	Often	Very high

The idea is not to prevent complex tasks, but to avoid that... all tasks should be treated as if they were complex. If a smaller model correctly solves a classification, using a frontier system with intensive reasoning does not make that classification better simply because we use more computation. It may only mean that we use more infrastructure.

Do users need to start feeling guilty about using AI?

Turning sustainability into individual guilt tends to obscure the most relevant decisions, which occur in architecture, infrastructure, energy procurement, product design, and model selection.

This doesn't mean that usage behavior is completely irrelevant. Generating dozens of unnecessary variations, using extremely advanced models for trivial tasks, or continuously running agents without a clear objective naturally adds computation.

But it would be unproductive to turn the environmental discussion of AI into a list of minor digital sins. The most important responsibility lies in to design systems where the efficient option is also the default option..

Platforms can automatically select models, interrupt redundant processes, reuse results, optimize contexts, reduce unnecessary tokens, and provide efficiency metrics for companies that need to make purchasing decisions. Users shouldn't need to know GPU power to make an environmentally sound choice.

Should companies also compare models based on energy consumption?

Whenever sufficiently comparable data is available, yes.

Today, however, there is a problem: suppliers and studies still use different methodologies. Benchmarks such as AI Energy Score They are trying precisely to establish more standardized conditions for comparing models and to allow energy consumption to become a more visible dimension of technological choice.

This points to an important evolution. In the future, selecting a corporate model may involve an equation similar to that used for other technologies: quality + cost + latency + privacy + security + energy consumption.

Instead of environmental efficiency appearing only in the annual sustainability report, it can become part of the technical decision-making process itself. When buyers begin to demand efficiency, suppliers gain an additional economic incentive to measure and improve it.

How to choose a more efficient AI in practice

An environmentally responsible choice can begin by verifying whether the task truly requires a boundary model, whether reasoning significantly improves the outcome, what the usage volume will be, and whether there is a benchmark for the model in the specific task to be performed.

It is also useful to assess whether the supplier discloses indicators such as PUE, WUE, energy and emissions, where the processing takes place, and whether there is the possibility of using smaller or specialized models for part of the workload.

In large-scale automated applications, even seemingly small elements can become important. An unnecessarily long response, a context repeated thousands of times, or an agent executing redundant steps may represent little in an individual interaction, but a great deal when multiplied by millions of operations.

The principle that connects all these decisions is simple: do not use more computing power than the problem requires..

So, does truly green AI exist?

Perhaps “green” is the wrong word, because it suggests a permanent property, as if we could put an environmental label on a certain model and consider the matter resolved.

The same model can produce different impacts depending on the hardware it runs on, the power supply, the duration of the task, the region, the cooling, and the number of times it is used. Similarly, a small model might be an excellent choice for one task and an inefficient choice for another if it needs to repeatedly perform a job that a more capable system would solve on the first try.

Therefore, a more useful definition would be: Sustainable AI is AI that seeks to deliver the necessary benefit using the least reasonable amount of resources, while simultaneously considering performance, energy, water, carbon, and infrastructure..

The important word in this definition is necessary. Sustainability doesn't require always using the smallest model. It requires justifying when we use a larger one.

Efficiency may be one of the greatest environmental solutions for AI itself.

There is an understandable tendency to imagine that the environmental impact of artificial intelligence will be addressed primarily in the energy sector, through more solar, wind, and nuclear power, storage, and larger power grids. All of this will likely be part of the answer.

But there is another source of energy that doesn't need to be produced: The energy we've managed to save because computing has become more efficient..

In an industry that operates on a global scale, improvements in hardware, software, and infrastructure capable of significantly reducing the amount of electricity needed to deliver the same amount of service can make a huge difference.

This shows that sustainability is not necessarily the enemy of innovation. In many cases, efficiency means precisely making the system work better: using less memory, obtaining more performance per watt, reducing latency, lowering costs, making better use of hardware, and making models capable of working in smaller devices.

The same improvements that reduce environmental impact can also expand access and reduce expenses.

The discussion is not about "using or not using AI".“

Upon discovering that artificial intelligence has an environmental footprint, we can react in two extreme ways. The first is to conclude that using the technology is environmentally irresponsible. The second is to believe that future efficiency gains will automatically resolve any impact.

Neither position is sufficient.

Artificial intelligence infrastructure is growing rapidly. At the same time, there is concrete evidence that models, hardware, and data centers can become much more efficient, and that different choices already produce enormous differences in the resources required for similar tasks.

Therefore, the most useful question is not "should we use AI?", but “"What is the least resource-intensive way to obtain the benefit we truly need from AI?"”

This logic replaces blame with engineering, green marketing with measurement, and polarization with continuous improvement.

Frequently asked questions about sustainable AI

What is Green AI?

Green AI is an approach that considers computational efficiency and resource cost as relevant dimensions in the development and evaluation of artificial intelligence systems, rather than solely analyzing performance and accuracy.

Do smaller models always use less energy?

Not necessarily. Size is an important factor, but architecture, hardware, number of tokens, reasoning, utilization, and type of processing also influence consumption. A small model used inefficiently can consume more than a larger system well-suited for a given task.

Does reasoning consume more energy?

In general, processes that use more inference steps and generate a larger number of tokens require more computation. Recent benchmarks have found very significant differences between versions with and without reasoning under some conditions, but these numbers should not be treated as universal constants.

What is quantization in AI?

It is a technique that reduces the numerical precision used to represent and process model parameters, seeking to decrease memory and computational cost without producing a significant loss of quality for a given application.

What is model distillation?

It is a technique in which a larger model helps train a smaller model so that some of its capabilities can be reproduced with fewer computational resources.

Can a data center cool servers without continuously consuming water?

Yes. There are closed-loop systems that prevent continuous water evaporation in server cooling. This doesn't mean the entire water footprint disappears, as electricity, equipment manufacturing, and other operations also have impacts.

Does renewable energy make AI sustainable?

It helps reduce the carbon intensity of electricity, but it doesn't solve all the impacts on its own. Hardware, water, materials, location, efficiency, and total processing volume also need to be considered.

Is using smaller AI always the environmentally correct choice?

No. The model needs to achieve the required quality. If a small system produces inadequate results and requires repetition or further processing by a larger model, the initial savings may disappear.

How we researched this article

This article uses academic research, efficiency benchmarks, data from energy institutions, and technical information released by infrastructure operators. Key references include the academic concept of... Green AI, the experiments and benchmarks of AI Energy Score, data and projections of International Energy Agency, research from Lawrence Berkeley National Laboratory and technical documentation on architectures and infrastructure released by companies such as Google and Microsoft.

Data provided directly by technology companies is treated as corporate data, and not automatically as independent measurements. This is especially important for figures related to prompt consumption, the efficiency of the data centers themselves, and the stated benefits of new technologies.

We also avoid using model size as an automatic synonym for energy consumption. Architecture, reasoning, number of tokens, hardware, and deployment conditions can significantly modify this relationship.

Limitation: There is still not enough standardized transparency to compare all business models under equivalent conditions. Efficiency also changes rapidly as models, chips, and infrastructure evolve. Therefore, specific numbers should always be interpreted within their methodology and reference date.

Sustainable AI in one sentence.

The most sustainable AI isn't necessarily the smallest, the newest, or the one hosted in the data center with the best environmental pitch; it's the one that delivers the necessary result using a proportionate amount of computing power and the most efficient infrastructure available for that task.

We don't need to choose between intelligence and sustainability. We need to abandon the idea that... More computing is always better computing..

If different models can produce adequate results using radically different amounts of energy in specific contexts, choosing the right model ceases to be merely a technical or financial decision. It also becomes an environmental decision.

Guilherme Godoy