A technology developed in the Global North will deepen resource stress in the Global South.
En un momento en que la humanidad deposita esperanzas crecientes en la inteligencia artificial, un informe de las Naciones Unidas nos recuerda que ninguna herramienta existe fuera del mundo físico. Para 2030, la infraestructura que sostiene a la IA consumirá agua equivalente a las necesidades domésticas básicas de 1.300 millones de personas en el África subsahariana y electricidad casi tres veces superior al consumo combinado de Pakistán, Bangladesh y Nigeria. El verdadero peso de esta advertencia no es técnico sino moral: los beneficios se concentran en el Norte Global mientras los costos ambientales se acumulan silenciosamente en las regiones con menor capacidad para soportarlos.
- Un informe de la ONU publicado el 3 de junio revela que los análisis ambientales sobre la IA han ignorado sistemáticamente el agua, el suelo y los recursos, centrándose solo en las emisiones de carbono.
- Para 2030, los centros de datos podrían consumir 945 teravatios-hora de electricidad, una cifra que convertiría a la industria en la undécima economía energética del mundo si fuera un país.
- El 80-90% del consumo energético de la IA no proviene del entrenamiento de modelos, sino de la inferencia diaria: cada consulta, imagen generada o video creado por algoritmos multiplica la presión sobre los recursos.
- Generar una imagen con IA requiere 1.450 veces más recursos que clasificar correos electrónicos; crear un video puede demandar hasta 200.000 veces más, revelando una disparidad brutal entre tipos de uso.
- Las regiones con escasez hídrica y menor desarrollo económico absorberán los costos ambientales de una tecnología cuyos beneficios fluyen principalmente hacia naciones y corporaciones ricas.
- Los investigadores no piden frenar la IA, sino usarla con responsabilidad y anticipar sus consecuencias antes de que la inequidad se vuelva irreversible.
Un instituto de investigación de las Naciones Unidas publicó esta semana un informe que obliga a replantear el verdadero costo ambiental de la inteligencia artificial. La conclusión es contundente: para 2030, la infraestructura física necesaria para sostener los sistemas de IA consumirá agua equivalente a las necesidades domésticas básicas de 1.300 millones de personas en el África subsahariana, y electricidad casi tres veces superior al consumo anual combinado de Pakistán, Bangladesh y Nigeria.
El informe, publicado el 3 de junio por el Instituto de la ONU para el Agua, el Medio Ambiente y la Salud, señala que los análisis ambientales sobre la IA han sido incompletos. La mayoría se limita a medir emisiones de carbono, ignorando el agua que consumen los centros de datos, la energía de los sistemas de refrigeración y la intensidad de recursos de las redes eléctricas que los alimentan. Al incorporar todos estos factores, el panorama cambia radicalmente. Kaveh Madani, director del instituto, fue claro: el informe no es un argumento contra la IA, sino un llamado a usarla de forma responsable y a anticipar sus consecuencias no deseadas.
Lo que complica aún más el cuadro es dónde se concentra realmente el consumo energético. Contrario a la intuición común, entrenar grandes modelos ya no es el principal gasto. Entre el 80 y el 90 por ciento de la electricidad asociada a la IA se destina a la inferencia: las millones de consultas diarias que los usuarios realizan en tiempo real. Y no todas las consultas son iguales. Una pregunta conversacional consume unas 200 veces más energía que clasificar correos electrónicos. Generar una imagen requiere 1.450 veces más recursos. Crear un video con IA puede demandar hasta 200.000 veces más que esa misma tarea de clasificación.
La preocupación más profunda del informe es la inequidad. Los costos ambientales —el agua extraída de acuíferos, el territorio ocupado por plantas de energía, el calor liberado a la atmósfera— no se distribuirán de manera uniforme. Las regiones que ya enfrentan escasez hídrica serán las más afectadas, mientras que los beneficios de la tecnología fluyen hacia naciones y corporaciones con mayores recursos. Es esa asimetría la que los investigadores buscan visibilizar: una tecnología desarrollada y desplegada principalmente en el Norte Global que profundizará la presión sobre los recursos del Sur Global.
A United Nations research institute released a report this week that reframes how we should think about artificial intelligence's true environmental footprint. The finding is stark: by 2030, the physical infrastructure required to run AI systems will consume water in quantities equal to what 1.3 billion people in sub-Saharan Africa need for basic household use. The electricity demand will be nearly three times what Pakistan, Bangladesh, and Nigeria consume combined in a year.
The report, published by the UN Institute for Water, Environment and Health on June 3rd, argues that the environmental reckoning for AI has been incomplete. Most analyses focus narrowly on carbon emissions—the greenhouse gases released when power plants generate electricity. But that misses the fuller picture. The researchers examined water consumption at data centers, the energy demands of cooling systems that prevent servers from overheating, and the resource intensity of the electrical grids that feed these operations. When you account for all of that, the environmental cost looks substantially different.
Kaveh Madani, who directs the institute and led the research, was careful about framing. "This report is not an argument against artificial intelligence," he said. "It is a call to use it responsibly and to address its unintended consequences proactively, to make it sustainable and fair." The distinction matters. The point is not that AI should not exist, but that its expansion is happening without adequate consideration of where the burden falls.
To put the electricity numbers in perspective: data centers alone consumed 448 terawatt-hours of electricity in 2025. If they were a country, that consumption would rank eleventh globally—ahead of Saudi Arabia, behind France. And that was before the projections for the next few years.
What makes the energy picture more complex is where that power actually goes. Most people assume that training large AI models—the computationally intensive process of teaching a system to recognize patterns—is the main energy drain. It is not anymore. Between 80 and 90 percent of AI-related electricity now goes to "inference," which is the technical term for what happens when millions of users ask the system questions every day. Every time someone types a prompt into a chatbot, every time an image is generated, every time a video is created by an algorithm, that is inference happening at scale.
But not all queries are equal. A standard conversational question consumes roughly 200 times more energy than the simple task of sorting emails into folders. Image generation requires about 1,450 times more resources than email sorting. And creating a video with AI can demand up to 200,000 times more energy than that same email classification task. The complexity of what you are asking the system to do directly determines how much of the world's electricity it consumes.
The report's deeper concern is about inequality. These environmental costs—the water drawn from aquifers, the land used for power generation, the heat released into the atmosphere—will not be distributed evenly. Regions already facing water scarcity will feel the pressure most acutely. The benefits of AI technology flow toward wealthy nations and corporations. The environmental consequences will accumulate in places with the least capacity to absorb them. That is the inequity the researchers are flagging: a technology developed and deployed primarily in the Global North will deepen resource stress in the Global South.
Citas Notables
This report is not an argument against artificial intelligence. It is a call to use it responsibly and to address its unintended consequences proactively, to make it sustainable and fair.— Kaveh Madani, director of the UN Institute for Water, Environment and Health
La Conversación del Hearth Otra perspectiva de la historia
So the report is saying AI uses a lot of water and electricity. But why does that matter more than, say, manufacturing or agriculture?
Because it's growing exponentially and we're not accounting for it. Agriculture has been around for millennia—we know its footprint. AI infrastructure is doubling every few years, and most policy discussions still treat it as a carbon problem only.
The inference thing—that's the part that surprised me. I thought training models was the energy hog.
That was true five years ago. But once a model is trained, millions of people use it every day. That daily use is now the dominant energy cost. And it's accelerating because more people are using these tools.
The water number is almost abstract. 1.3 billion people. How do you even visualize that?
You don't need to visualize it. You need to understand that sub-Saharan Africa already has severe water stress. Adding another massive demand on top of that—demand that serves people elsewhere—is a concrete harm.
But couldn't we just build data centers in places with abundant water and cheap electricity?
That's what companies are already doing. But "abundant" is relative, and it often means taking water from agricultural regions or communities that depend on it. You're just moving the problem, not solving it.
So what does "responsible use" actually mean, in practical terms?
That's the question the report leaves open. It could mean limiting certain kinds of queries, or building more efficient systems, or requiring companies to offset water use. But right now, there's no mechanism for any of that.