Weekly AI Watch

Every week, what changed and what it means for you. No hurry, no alarm.

Week of 6 November 2026 | Past editions15,706 papers found · 14,822 read · 4,771 belonged · 43 stories passed every check

Observatorio de la
Inteligencia Artificial

Cada semana, qué cambió y qué significa para usted. Sin prisa y sin alarma.

Semana del 6 de noviembre de 2026 | Ediciones anteriores15,706 estudios encontrados · 14,822 leídos · 4,771 correspondían · 43 notas pasaron todas las comprobaciones

weeklyAI · Week of 6 November 2026weeklyAI · Semana del 6 de noviembre de 2026

← Your bill← Su recibo

other · Environmental management · la publicación, 15 sep 2026 · gratis

Los centros de datos de IA buscan la luz barata y escriben su huella en otra parte

Un estudio de 247 proyectos anunciados halló que los centros se concentran donde la electricidad emite más carbono; sus cuentas de agua y emisiones son un patrón, no el retrato de ningún proyecto en particular.

Versión breve · la versión detallada sigue, unos 7 min

Pregunte al chat de weeklyAI

Pregúnteme sobre la IA en su vida: le señalo las notas que weeklyAI ha publicado sobre el tema, esta semana y antes.

Las conversaciones se guardan mientras exista weeklyAI, para mejorar la publicación. Se responde en el idioma en que usted escribe.

El estudio, de un vistazo
Quiénes
Proyectos anunciados de centros de datos para inteligencia artificial
Cuántos
247
Dónde
34 jurisdicciones
Cuándo
entre enero de 2018 y marzo de 2025
Tipo de estudio
analysis of what people did
Quién lo hizo
Punjab Bar Council y Universidad Mohi-ud-Din Islamic
El límite que importa
Se basa en cargas anunciadas y supone 85% de uso; el consumo real puede diferir.
De cada 100 gramos de gases de un racimo de cómputo, de dónde salen
Fabricar los servidores70%
La refrigeración15%

Son las partes del total de gases de efecto invernadero de todo el ciclo de vida de un racimo de 3.6 megavatios con aceleradores NVIDIA H100; el resto no se reparte en el pasaje.

Qué pesa más al elegir dónde se instala un centro: el carbono de la red o el estrés del agua

Red con 100 gramos más de CO₂ por kilovatio horafrente aCuenca que pasa de estrés hídrico bajo a alto

La red sucia sube la probabilidad de instalarse 8.1 puntos; la cuenca estresada la baja 5.7 puntos, y el freno desaparece si hay electricidad limpia

La luz de su recibo y el agua de su llave son la misma cuenta.
Lectura de weeklyAI

Imagínese el recibo de luz de un centro de cómputo del tamaño de una fábrica. Un estudio con 247 proyectos grandes de centros de datos para inteligencia artificial, anunciados entre 2018 y principios de 2025 en 34 jurisdicciones, encontró que estos tienden a ubicarse donde la red eléctrica emite más dióxido de carbono. Por cada 100 gramos más de dióxido de carbono por kilovatio hora que emite la red local, la probabilidad de que un proyecto se instale allí sube 8.1 puntos porcentuales. Los autores son Hazrat Usman, del Punjab Bar Council, y Sidra Zakir, de la Universidad Mohi-ud-Din Islamic.

No dice que la red sucia atraiga por sí sola a las empresas. El estudio observa qué se anunció y dónde, y describe una asociación, no una causa. Tampoco dice que los proyectos huyan de la sequía: en zonas de estrés hídrico alto, la probabilidad bajó 5.7 puntos, pero si allí había electricidad limpia, el freno desapareció.

Los autores compararon cada proyecto con la red del país del cliente. La diferencia mediana fue de 348 kilos de dióxido de carbono por kilovatio hora, y 0.49 litros de agua. El artículo imprime la unidad como kilogramos por kilovatio hora, unas mil veces lo que emite cualquier red eléctrica por kilovatio hora, así que la cifra probablemente está en gramos. Sumadas todas las diferencias de carbono del estudio, la primera sería de 21.4 millones de toneladas de dióxido de carbono al año. El agua, en cambio, se calculó para un centro típico de 150 megavatios: 1,780 millones de litros al año. Es una escala, no un pronóstico para su zona: sale de cargas anunciadas y de un supuesto de uso del 85%, y el agua se estima con promedios regionales de enfriamiento. Para el centro que le propongan, pida el permiso de agua y el trámite ante la eléctrica: allí están las cifras reales.

El estudio también halló que el 42% de la carga eléctrica prevista caería en condados de Estados Unidos por encima de la mediana nacional en pobreza y en proporción de población no blanca. Y que solo el 18% de las empresas obligadas por la directiva europea de diligencia publica sus emisiones por proyecto; el 7% informa algo sobre agua. El 71% de las emisiones estimadas se atribuye a empresas alcanzadas por esa norma.

Los autores calcularon que comprar energía limpia a toda hora, enfriar con agua reciclada y aplicar ajustes fronterizos podría reducir a la mitad esas cifras. Son escenarios. Ningún tribunal ha resuelto todavía si esa obligación llega a un centro de datos en otro país.

Cuando anuncien un centro cerca de usted, busque el permiso de agua y el trámite de interconexión eléctrica. Son documentos públicos, y en ellos el consumo aparece medido.

Qué significa para usted

Cuando anuncien un centro cerca de usted, pida en el municipio el permiso de agua y el trámite de conexión ante la eléctrica. Ahí aparecen el consumo medido y lo que se le pedirá a la red de su zona, y podrá compararlo después con lo que el centro declare. El estudio no midió esos efectos en ningún lugar concreto: siga esos documentos cuando existan.

Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1

Quién pagó: Los autores declaran que no recibieron financiación, subvenciones ni apoyo de ninguna entidad pública, comercial o sin ánimo de lucro, y que ninguna persona o institución externa contribuyó al estudio.

Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1452 palabras · unos 7 minLeerla →Cerrar

Dónde ponen los centros de datos de la IA el costo que usted paga

Un estudio revisó 247 proyectos anunciados entre 2018 y 2025 y encontró que los centros se colocan más seguido donde la red eléctrica emite más carbono, y que la cuenta ambiental del cómputo que se hace para clientes de otro país es grande.

Piense en el recibo de luz y en el recibo del agua. Ahora piense en un edificio lleno de chips que trabaja día y noche para responder preguntas hechas desde otro país. Un estudio de Hazrat Usman, del Colegio de Abogados de Punjab, y Sidra Zakir, de la Universidad Mohi-ud-Din, revisó 247 anuncios de centros de datos para inteligencia artificial hechos entre enero de 2018 y marzo de 2025, en 34 jurisdicciones, y los cruzó con datos de emisiones de las redes eléctricas y de estrés hídrico de las cuencas. Los dos autores trabajaron sin financiamiento y declararon no tener conflicto de interés. No probaron ningún aparato ni midieron un centro en operación: analizaron anuncios, factores de emisión y normas legales.

Conviene saber qué tamaño tiene el asunto. Se calcula que la capacidad mundial de servidores superará los 205 gigavatios en 2030, y el 90 por ciento de esa capacidad estará impulsado por cargas de IA, según proyecciones de la industria que recoge el artículo1. Para que se haga una idea de lo que es un gigavatio, piense en una planta grande de electricidad: esa es la escala con la que se está construyendo.

El chip es la parte que menos se ve y la que más pesa. Es decir: la mayor parte de la huella ya ocurrió antes de que el aparato se enchufe. Lo curioso es que el enfriamiento, que uno imagina como el gran gasto de agua, es solo una parte del problema: también se gasta agua en generar la electricidad y en fabricar los chips.

El hallazgo central es una asociación. Cada 100 gramos de CO₂ por kilovatio-hora que una red eléctrica emite por encima del promedio del cliente, la probabilidad de que esa jurisdicción atraiga nueva capacidad sube 8.1 puntos porcentuales4. Cuando los proyectos se pesan por el dinero anunciado, la cifra sube a 11.4 puntos: las inversiones más grandes son las que más se inclinan hacia redes sucias5. El estudio no puede decir que la red sucia sea la causa. Solo muestra que donde la red emite más, se anuncian más centros.

Con el agua pasa algo distinto y revelador. Pasar de una cuenca de estrés hídrico "bajo" a una de estrés "alto" baja la probabilidad de instalarse en 5.7 puntos porcentuales6. Pero el efecto desaparece cuando la red eléctrica es limpia. O sea: la empresa acepta una cuenca estresada si a cambio presume energía limpia. El agua se negocia; el carbono se cuida por imagen.

La cuenta que sale de ahí: el excedente medio de carbono por kilovatio-hora entregado, frente a hacer el mismo trabajo en la red del cliente, es de 348 kilogramos de CO₂. El artículo imprime la unidad como kilogramos por kilovatio hora, unas mil veces lo que emite cualquier red eléctrica por kilovatio hora, así que la cifra probablemente está en gramos. La revisión recoge un estudio en el que, en agua, el excedente medio es de 0.49 litros por kilovatio-hora, que para una instalación típica de 150 megavatios suma 1,780 millones de litros al año7. Tres lugares concentran el 58 por ciento de esa carga de agua que cruza fronteras: el centro de Arizona, el norte de Virginia y Madrid8.

¿Y quién está detrás de esos anuncios? El 71 por ciento de la fuga de carbono y el 64 por ciento de la de agua se rastrea a empresas que ya caen bajo la Directiva europea de diligencia debida corporativa. Sin embargo, solo el 18 por ciento de esas empresas publica factores de emisión por proyecto, y apenas el 7 por ciento reporta alguna cifra de agua. La palabra de la empresa, por lo visto, alcanza para el comunicado y no para el dato.

Hay que decir que el incentivo para instalarse ahí no cae del cielo. Quien paga la factura local compite con una empresa que llega con exención y con prisa.

¿Qué se ha intentado? Existen políticas de lugar para atraer o frenar estos desarrollos, como las Zonas de Crecimiento de IA del Reino Unido y el programa chino de "datos del este, cómputo del oeste", pero según el resumen de un estudio sobre localización de centros de datos, la evidencia sobre qué decide dónde se instalan es escasa11. También encontró algo que contradice el sentido común: los lugares más cálidos, donde enfriar cuesta más, albergan más instalaciones13.

Sobre el agua hay más. Ese mismo trabajo advierte que sigue abierto si la exposición al estrés hídrico se traduce en impactos operativos o en consecuencias medibles para otros usuarios del agua17.

Los límites del estudio importan para que usted no lea de más. Las cifras de fuga se calculan sobre las cargas anunciadas y suponen un 85 por ciento de utilización; el consumo real de un centro puede diferir. El modelo de agua usa promedios regionales de eficiencia, no el equipo concreto de cada sitio: un centro con enfriamiento híbrido o agua reciclada podría extraer mucho menos. El conjunto cubre 247 proyectos de 20 megavatios o más anunciados entre 2018 y principios de 2025, así que los centros más chicos o más nuevos cerca de usted pueden no aparecer: que no estén en el estudio no prueba que no existan. Y las proyecciones no incorporan la evolución futura de las redes eléctricas ni modelos de cuenca que liguen la extracción con la disponibilidad por temporada.

El dato social es el que más debería mirar quien vive cerca. La revisión recoge un estudio en el que, al cruzar la ubicación de los proyectos con índices de vulnerabilidad de condado, se encontró que el 42 por ciento de la carga eléctrica proyectada para IA caerá en condados por encima de la mediana nacional tanto en pobreza como en proporción de población no blanca18.

Así lo leemos nosotros. Cuando una actividad sirve a clientes de fuera y se instala donde los costos caen sobre gente que no decide, lo esperable es que en la audiencia del permiso los vecinos pesen menos que la empresa y que los clientes lejanos, y que las condiciones de agua, ruido y respaldo eléctrico se negocien sin que quien paga la factura local esté en la mesa. Usted podría comprobarlo en el expediente municipal o del regulador: si las objeciones vecinales aparecen atendidas, si se impusieron condiciones al operador por presión local, o si la tarifa y el permiso de agua se ajustaron después de una participación con efecto medible, entonces nos equivocamos. Lo que puede hacer es pedir ese expediente completo y la lista de quién compareció, y anotar si los vecinos figuran como parte o solo como notificados.

También cabe esperar que el operador negocie desde una posición de fuerza, ofreciendo empleos e impuestos como contrapeso, y que las obligaciones concretas de agua, respaldo y retiro al cerrar queden vagas o a su buena voluntad. Sabríamos que nos equivocamos si encuentra contratos vinculantes con plazos, multas y garantías de retiro, auditados y públicos, y si el operador aceptó por escrito límites de consumo antes de construir. Vale la pena leer el convenio de incentivos y buscar qué pasa si la empresa se va o no cumple: qué garantía queda en el municipio.

Qué se puede hacer técnicamente ya está sobre la mesa. Un análisis de escenarios del propio artículo indica que comprar energía libre de carbono las 24 horas, enfriar con sistemas híbridos y aplicar ajustes fronterizos al cómputo podría reducir a la mitad tanto la fuga de carbono como la de agua20. Son escenarios modelados, no resultados observados: ningún tribunal ha resuelto todavía si la directiva europea alcanza estas cadenas de actividad.

Para su casa, la pregunta útil no es si el centro de datos es bueno o malo, sino qué le van a pedir a su red y a su cuenca. Cuando llegue el anuncio, pida tres cosas por escrito: el estudio de demanda eléctrica que presentó la empresa a la utility, el permiso de agua con su volumen y su fuente, y el convenio de incentivos con lo que pasa si se va. Si no se los dan, eso también es una respuesta. Y si ya vive junto a uno, compare el consumo medido con el pronosticado en el permiso: ahí se ve si la promesa se sostuvo.

La luz de su recibo y el agua de su llave son la misma cuenta. Pregunte quién la está pagando antes de que llegue el primer camión.

De dónde sale cada dato de contexto, y cuánto leímos de cada documento

  1. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Industry forecasters predict that the global stock of server capacity will exceed 205 gigawatts by 2030, 90 percent of which is driven by AI workloads (Infrastructure Masons, 2025)."
  2. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Water withdrawals are similarly non-trivial; current estimates place datacenter consumption near 292 million gallons per day and trending toward 450 million gallons per day by 2030 (Friedmann, 2024)."
  3. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - el artículo del que trata esta nota — el artículo completo — el pasaje: "life-cycle assessments for a single 3.6-megawatt AI cluster anchored on NVIDIA H100 accelerators attribute approximately 70 percent of cradle-to-grave greenhouse gas (GHG) emissions to embodied server manufacturing and a further 15 percent to cooling operations (d’Orgeval et al., 2024)."
  4. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Every 100 g CO₂ kWh⁻¹ increase in host-grid intensity raises the probability of attracting new capacity by 8.1 percentage points, while moderate water stress deters investment only when carbon advantages are absent."
  5. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - el artículo del que trata esta nota — el artículo completo — el pasaje: "When the same projects were re-weighted by publicly disclosed capital expenditure, the elasticity rose to 11.4 percentage points, confirming that the largest investments gravitate most strongly toward high-carbon grids"
  6. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Controlling for carbon intensity, corporate-tax rates, and data-sovereignty restrictions, moving from a “low” to a “high” water-stress classification lowered siting probability by 5.7 percentage points (p = .03)."
  7. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - el artículo del que trata esta nota — el artículo completo — el pasaje: "the sample’s median offshore differential equalled 0.49 L kWh⁻¹, or 1.78 billion liters annually for a typical 150-megawatt facility."
  8. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Projects located in central Arizona, northern Virginia, and Madrid together accounted for 58 percent of the total cross-border withdrawal burden"
  9. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Microsoft’s 2024 announcement of a 1.2-gigawatt campus in Goodyear, Arizona, for generative-AI inference would demand more than 4.5 billion gallons of water annually under standard Water Usage Effectiveness values, equal to roughly 20 percent of the city’s current municipal withdrawals (Food & Water Watch, 2025)."
  10. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In the United States, nine states exempt data-center electricity from sales tax and seven grant “mega-project” property tax holidays exceeding 20 years (Urban Land Institute, 2024)."
  11. Hajkowicz S. (2027). Where the cloud clusters: A global spatial analysis of data centre locations and their determinants. Applied Geography. https://doi.org/10.1016/j.apgeog.2026.104220 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Place-based policies such as the UK's AI Growth Zones and China's ‘Eastern Data Western Computing’ programme seek to attract, steer or restrain this development, yet the evidence base on what drives location choice is thin."
  12. Hajkowicz S. (2027). Where the cloud clusters: A global spatial analysis of data centre locations and their determinants. Applied Geography. https://doi.org/10.1016/j.apgeog.2026.104220 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Data centres are highly concentrated with 76.2% inside 258 clusters at a median nearest-neighbour distance of 0.86 km."
  13. Hajkowicz S. (2027). Where the cloud clusters: A global spatial analysis of data centre locations and their determinants. Applied Geography. https://doi.org/10.1016/j.apgeog.2026.104220 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Climate runs counter to expectations. Warmer places with greater cooling costs host more facilities."
  14. Tahmid T, Upaul S. (2026). Water footprint of large AI models: risks to global water security and sustainable optimization strategies. Water Policy. https://doi.org/10.2166/wp.2026.233 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Critically, over 52% of the world's top 100 data centers are projected to be located in high water stress regions by 2030, where AI infrastructure increasingly competes with agricultural, municipal, and ecological needs."
  15. Tahmid T, Upaul S. (2026). Water footprint of large AI models: risks to global water security and sustainable optimization strategies. Water Policy. https://doi.org/10.2166/wp.2026.233 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Direct water use occurs primarily through evaporative and liquid cooling in data centers, where a single large facility can consume 11,000–19,000m3/day."
  16. Rosa L, Sangiorgio M, Scanlon BR, Mazzotti M, Berry J, Michalak AM. (2026). One Quarter of Global Data Centers are Located in Water-Stressed Areas. Environmental Science & Technology. 10.1021/acs.est.6c08088 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "The United States hosts 39% of data centers, with 21% located in water-scarce regions, where overlap with municipal and agricultural demands is pronounced."
  17. Rosa L, Sangiorgio M, Scanlon BR, Mazzotti M, Berry J, Michalak AM. (2026). One Quarter of Global Data Centers are Located in Water-Stressed Areas. Environmental Science & Technology. 10.1021/acs.est.6c08088 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "The extent to which water-stress exposure translates into operational impacts for data centers themselves or into measurable consequences for other water users remains an open question to address."
  18. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Synthesizing county-level Social Vulnerability Index values with siting coordinates showed that 42 percent of the projected AI-linked electricity load will accrue in counties above the national median for both poverty and non-white population share"
  19. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Case-study evidence from Iceland, Norway, and Greenland shows that while renewable siting offers low operational carbon intensity, it can trigger a “digital resource curse”: housing inflation, boom-bust labor cycles, and strain on fragile grids built for aluminum smelting rather than 24-hour server loads (Sovacool et al., 2022)."
  20. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Scenario analysis showed that 24/7 carbon-free procurement, hybrid cooling, and compute-adjusted border adjustments could halve both carbon and water leakage."

Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1

Quién pagó: Los autores declaran que no recibieron financiación, subvenciones ni apoyo de ninguna entidad pública, comercial o sin ánimo de lucro, y que ninguna persona o institución externa contribuyó al estudio.

Los hallazgos de otros estudios que aquí se mencionan provienen de este documento; esos estudios no se consultaron uno por uno.

other · Environmental management · the paper, 15 Sep 2026 · free

AI Data Centres Are Drawn to Dirty Power. The Study That Found It Can't Tell You About the One Near You.

In 247 announced projects, higher-carbon grids attracted more capacity, and modelled emissions and water use exceeded what home grids would have cost. The figures rest on announced designs, not metered use.

Short version · the longer version follows, about 6 min

Ask the weeklyAI chat

Ask me about AI in your life: I point you to the stories weeklyAI has published on it, this week and before.

Conversations are saved for as long as weeklyAI exists, to improve the publication. Answers come in the language you write in.

The study at a glance
Who
Announced large AI data-centre projects
How many
247
Where
34 host jurisdictions, including US states, Latin America, the EU and Asia
When
2018 to early 2025
Kind of study
analysis of what people did
Who did it
Punjab Bar Council and Mohi-ud-Din Islamic University
The limit that matters
Figures rest on announced designs and assumed 85 percent use, not metered draw

What the study compared

A host grid 100 grams of carbon dioxide per kilowatt-hour dirtieragainstA cleaner host grid

The dirtier grid went with an 8.1 percentage point higher chance of being chosen

Projects weighted by disclosed spendingagainstProjects counted equally

The pull toward dirty grids rose to 11.4 percentage points

High water-stress placeagainstLow water-stress place

High stress lowered the chance of being chosen by 5.7 percentage points

Same computing run on the customer's home gridagainstSame computing run on the host grid

The host grid came out 348 kilograms of carbon dioxide and 0.49 litres of water higher per kilowatt-hour

So read the utility filing and the water permit, not the press release.
weeklyAI's reading

A large data centre is planned near you. The company promises efficiency and renewable power. What the study behind this story found is a pattern across 247 large AI projects announced between 2018 and early 2025, in 34 jurisdictions that include US states, Latin America, the EU and Asia.

Two legal academics, Hazrat Usman of the Punjab Bar Council and Sidra Zakir of Mohi-ud-Din Islamic University, gathered the announcements and compared them with grid-carbon and water-stress data.

On the power side, the pattern was blunt. Every 100 grams of carbon dioxide per kilowatt-hour more on a host grid went with an 8.1 percentage point higher chance that a project chose that jurisdiction. The study reports an association, not proof that dirty grids caused the choice.

On water, the finding was stranger. Moving from low to high water stress lowered the chance of a project by 5.7 percentage points. But where a low-carbon grid was also available, high water stress did not deter the projects at all.

The authors then modelled what the same computing would have cost on the customer's home grid. The typical project's offshore emissions came out at 348 kilograms of carbon dioxide per kilowatt-hour of delivered computing, and its water use 0.49 litres higher. The paper prints the unit as kilograms per kilowatt-hour, about a thousand times what any power grid emits per kilowatt-hour, so the figure is probably in grams. Added up for the whole sample in its first year, that is 21.4 million tonnes of carbon dioxide. The 1.78 billion litres of water is the yearly figure for one typical 150-megawatt site.

These are estimates, not readings. They rest on announced design loads and an assumed 85 percent utilisation, and on regional average cooling figures rather than the equipment at any one site. Treat them as a scale, not a forecast for your local project.

The study adds two findings about who carries the cost. It mapped 42 percent of the projected electricity load to US counties above the national median for both poverty and non-white population share. It also found that only 18 percent of the companies covered by the EU's due-diligence directive disclose project-level emissions, and 7 percent report any water figure.

The authors argue that hourly reporting rules and border adjustments tied to computing could roughly halve the modelled leakage. Their own scenarios are models. No court has tested whether the duty reaches a foreign data centre.

The study may miss smaller computing projects below 20 megawatts, because they do not appear in public announcements.

When a centre is proposed, the water permit and the utility interconnection filing state the actual draw. Ask for those two documents before the efficiency claims.

What this means for you

When a centre is proposed near you, the water permit and the utility filing say what it will draw, and those are worth asking for before any efficiency claim. The study's figures are modelled from announced designs, so they set a scale for the whole sample and cannot tell you about your local project.

Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1

Who paid: The authors state they received no funding, grants or support from any public, commercial or not-for-profit entity, and that no outside individual or institution contributed to the study.

The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1239 words · about 6 minRead it →Close

Higher-carbon power grids were more likely to attract new AI data centres

An analysis of 247 announced projects finds that the power and water they use offshore are large enough to matter for the law — and that the companies involved mostly do not report them.

A water bill, a power bill, a drought notice: three pieces of paper that tell a household what it can afford. Now put them next to an announcement that a computing campus is coming to town.

The study examined 247 large data-centre projects built for artificial intelligence, announced between 2018 and early 2025, across 34 host jurisdictions. The authors received no funding and declared no competing interests. They did two things: they measured where these projects went, and they asked what international environmental law can do about the costs that cross borders.

The central finding is a pull, not a proof. Each 100 grams of carbon dioxide per kilowatt-hour that a host grid emits above the customer's home grid was associated with an 8.1 percentage-point rise in the chance that a project would be sited there. Weighted by disclosed spending, the pull reached 11.4 points12. Water stress worked the other way, but weakly: moving from low to high stress lowered the chance of selection by 5.7 points, and the deterrent disappeared when a low-carbon grid was on offer3.

The modelled leakage is large. The median offshore surplus was 348 kilograms of carbon dioxide per delivered kilowatt-hour and 0.49 litres of fresh water, totalling 21.4 million tonnes of carbon dioxide and 1.78 billion litres of water in the sample's first year. The paper prints the unit as kilograms per kilowatt-hour, about a thousand times what any power grid emits per kilowatt-hour, so the figure is probably in grams. That is roughly Croatia's 2022 national emissions, and enough water for a city's worth of households. Central Arizona, northern Virginia and Madrid alone carried 58 percent of the cross-border water burden4. Only 18 percent of the firms covered by European due-diligence rules disclose project-level emissions, and 7 percent report any water figure.

Consider what the building actually is. The article describes a study in which a single 3.6-megawatt cluster of the chips that train AI models puts about 70 percent of its lifetime greenhouse gases into manufacturing the servers and a further 15 percent into cooling them5. The article describes a study in which the global stock of server capacity is forecast to pass 205 gigawatts by 2030, 90 percent of it driven by AI6. The article describes a study in which data-centre water use in the United States sits near 292 million gallons a day and is trending toward 450 million by 20307.

The mechanism is simple enough to see from a kitchen table. Electricity from a coal-heavy grid carries more carbon per unit than power from a nuclear or hydro grid, so the same AI workload costs the atmosphere more in one place than another. The article describes a study in which providers site campuses where electricity is cheap, water rights are permissive and carbon rules are looser — what the German Environment Agency calls AI-driven carbon-leakage risk8. The customer abroad gets the answer; the host community gets the draw on its grid and its aquifer. The article describes a study in which a planned 1.2-gigawatt campus in Goodyear, Arizona, would need more than 4.5 billion gallons of water a year under standard cooling, about 20 percent of the city's current municipal withdrawals9.

States compete for these buildings. The article describes a study in which nine US states exempt data-centre electricity from sales tax, and seven grant property-tax holidays longer than 20 years10. The trade-off is plain: tax revenue and construction jobs against power demand, water draw and strain on local grids. The article describes a study in which case studies from Iceland, Norway and Greenland describe a "digital resource curse" — housing inflation, boom-bust labour and fragile grids built for aluminium smelting rather than 24-hour server loads11.

Other research has looked at the same ground. According to the summary of a study of the water footprint of large AI models, more than half of the world's top 100 data centres are projected to sit in high water-stress regions by 2030; we could read only the summary12. The summary of a separate analysis of 9,347 identified data centres found that a quarter would already have faced at least one month a year of water demand exceeding supply under 2001–2010 conditions; we could read only the summary13. That study also found the United States hosts 39 percent of data centres, with 21 percent in water-scarce regions14. Its authors say it remains an open question whether that exposure translates into measurable consequences for other water users15.

The legal half is an argument, not a ruling. Because grid-carbon and water-stress figures are public, the authors argue operators cannot claim ignorance, and that the volumes pass the "appreciable harm" threshold that anchored the duty to prevent transboundary damage in the Trail Smelter case. The European Union's Corporate Sustainability Due Diligence Directive covers firms above 1,000 employees and €450 million in turnover; 71 percent of the modelled carbon leakage and 64 percent of the water leakage trace to firms within that scope. No court has yet ruled on whether foreign data-centre emissions fall inside a company's "chain of activities."

Here is how we read it. The pattern is distance: the people who benefit from a computation and the people who supply its power and water are often in different countries, and the second group seldom hears about the first. We expect that when a centre is proposed near you, the announcement will lead with investment and jobs, and the numbers that matter — contracted draw, water permit, tax abatement, backup diesel — will surface later, in a rate case, a permit hearing or a drought order. You would know we are wrong if audited project-level figures were published and debated before any dispute. So read the utility filing and the water permit, not the press release.

We also read a risk in how these arguments get settled. Residents are often told that a water-stress index or an emissions factor has answered the question of what their town can bear. Models can say what a basin averages; they cannot say what your neighbours will accept. Bring your own evidence — meter readings, well levels, dated photographs — and insist the limit is set with you, not only for you.

The authors' own scenario is the hopeful part, and it is a model, not a result. Combining round-the-clock carbon-free power purchases, hybrid cooling and a border adjustment tied to computing could halve both the carbon and the water leakage16. The limits are real: the figures rest on announced design loads and an assumed 85 percent utilisation, so treat them as a scale, not a forecast for your local project. Water numbers use regional average cooling, so ask for the permit. The dataset may undercount smaller edge-computing projects below 20 megawatts that escape public announcements.

The last thing to carry away is a question you can ask at the next council meeting: who are the customers of this centre, and where do they live? If the answer is a company and a city on another continent, you now know what that means for your grid and your well — and that the bill for the answer is being settled where you live.

Where each piece of context comes from, and how much of it we read

  1. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - the article this story is about — the whole article — the passage: "Every 100 g CO₂ kWh⁻¹ increase in host-grid intensity raises the probability of attracting new capacity by 8.1 percentage points, while moderate water stress deters investment only when carbon advantages are absent."
  2. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - the article this story is about — the whole article — the passage: "When the same projects were re-weighted by publicly disclosed capital expenditure, the elasticity rose to 11.4 percentage points, confirming that the largest investments gravitate most strongly toward high-carbon grids"
  3. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - the article this story is about — the whole article — the passage: "Controlling for carbon intensity, corporate-tax rates, and data-sovereignty restrictions, moving from a “low” to a “high” water-stress classification lowered siting probability by 5.7 percentage points (p = .03)."
  4. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - the article this story is about — the whole article — the passage: "Projects located in central Arizona, northern Virginia, and Madrid together accounted for 58 percent of the total cross-border withdrawal burden"
  5. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - the article this story is about — the whole article — the passage: "life-cycle assessments for a single 3.6-megawatt AI cluster anchored on NVIDIA H100 accelerators attribute approximately 70 percent of cradle-to-grave greenhouse gas (GHG) emissions to embodied server manufacturing and a further 15 percent to cooling operations (d’Orgeval et al., 2024)."
  6. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - the article this story is about — the whole article — the passage: "Industry forecasters predict that the global stock of server capacity will exceed 205 gigawatts by 2030, 90 percent of which is driven by AI workloads (Infrastructure Masons, 2025)."
  7. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - the article this story is about — the whole article — the passage: "Water withdrawals are similarly non-trivial; current estimates place datacenter consumption near 292 million gallons per day and trending toward 450 million gallons per day by 2030 (Friedmann, 2024)."
  8. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - the article this story is about — the whole article — the passage: "Cloud providers routinely site new campuses in jurisdictions with inexpensive electricity or permissive water rights regimes and increasingly with laxer carbon-intensity profiles, creating what the German Environment Agency terms “AIdriven carbon-leakage risk’ (INFRAS, 2025, p. 12)."
  9. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - the article this story is about — the whole article — the passage: "Microsoft’s 2024 announcement of a 1.2-gigawatt campus in Goodyear, Arizona, for generative-AI inference would demand more than 4.5 billion gallons of water annually under standard Water Usage Effectiveness values, equal to roughly 20 percent of the city’s current municipal withdrawals (Food & Water Watch, 2025)."
  10. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - the article this story is about — the whole article — the passage: "In the United States, nine states exempt data-center electricity from sales tax and seven grant “mega-project” property tax holidays exceeding 20 years (Urban Land Institute, 2024)."
  11. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - the article this story is about — the whole article — the passage: "Case-study evidence from Iceland, Norway, and Greenland shows that while renewable siting offers low operational carbon intensity, it can trigger a “digital resource curse”: housing inflation, boom-bust labor cycles, and strain on fragile grids built for aluminum smelting rather than 24-hour server loads (Sovacool et al., 2022)."
  12. Tahmid T, Upaul S. (2026). Water footprint of large AI models: risks to global water security and sustainable optimization strategies. Water Policy. https://doi.org/10.2166/wp.2026.233 — only the abstract - the full text could not be fetched — the passage: "Critically, over 52% of the world's top 100 data centers are projected to be located in high water stress regions by 2030, where AI infrastructure increasingly competes with agricultural, municipal, and ecological needs."
  13. Rosa L, Sangiorgio M, Scanlon BR, Mazzotti M, Berry J, Michalak AM. (2026). One Quarter of Global Data Centers are Located in Water-Stressed Areas. Environmental Science & Technology. 10.1021/acs.est.6c08088 — only the abstract - the full text could not be fetched — the passage: "We find that 25% of data centers would already have experienced at least 1 month of water gap annually under 2001-2010 hydrologic conditions."
  14. Rosa L, Sangiorgio M, Scanlon BR, Mazzotti M, Berry J, Michalak AM. (2026). One Quarter of Global Data Centers are Located in Water-Stressed Areas. Environmental Science & Technology. 10.1021/acs.est.6c08088 — only the abstract - the full text could not be fetched — the passage: "The United States hosts 39% of data centers, with 21% located in water-scarce regions, where overlap with municipal and agricultural demands is pronounced."
  15. Rosa L, Sangiorgio M, Scanlon BR, Mazzotti M, Berry J, Michalak AM. (2026). One Quarter of Global Data Centers are Located in Water-Stressed Areas. Environmental Science & Technology. 10.1021/acs.est.6c08088 — only the abstract - the full text could not be fetched — the passage: "The extent to which water-stress exposure translates into operational impacts for data centers themselves or into measurable consequences for other water users remains an open question to address."
  16. Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1 - the article this story is about — the whole article — the passage: "Scenario analysis showed that 24/7 carbon-free procurement, hybrid cooling, and compute-adjusted border adjustments could halve both carbon and water leakage."

Usman, H., Zakir, S. (2026). Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. Environmental Management. https://doi.org/10.1007/s00267-026-02610-1

Who paid: The authors state they received no funding, grants or support from any public, commercial or not-for-profit entity, and that no outside individual or institution contributed to the study.

The findings of other studies mentioned here come from this document; those studies were not opened one by one.

The free weekly drip by email

One email a week: the free story of every section, with its picture, and the headlines of the rest.

No payment, nothing to cancel. Unsubscribe from any issue.

El goteo semanal gratis por correo

Un correo por semana: la nota gratuita de cada sección, con su imagen, y los titulares de las demás.

Sin pago y sin nada que cancelar. Puede darse de baja desde cualquier envío.

From the house

About

Every week the new research on AI is read here, and this page tells you what was found, how far it can be trusted, and what you can do with it. This page is free, and every line on it says what was found.

The studies are read by a machine built and ruled by a professor of communication at Sacramento State and Universidad Nacional Autónoma de México (UNAM). Who makes this →

Diego Bonilla, Ph.D.

Full Professor of Communication Studies, Sacramento State, California, United States
Graduate instructor, Universidad Nacional Autónoma de México (UNAM), Mexico

hypergraphia.com

I am a professor of communication at Sacramento State, where I created the course Information Management and Privacy and have taught it many times, and at UNAM, in Mexico City, I teach the graduate course Electronic and Generative Literature with AI. In both I teach AI. I have taught since 1989, at first mathematics and physics to children who found them hard, and since 1997 one subject that has changed its name many times: media literacy, computer literacy, digital literacy, information literacy, ICT literacy, and now AI literacy.

This page exists because most people cannot go back to a university, and these systems are changing their work, their children's schooling, their parents' care and their vote all the same. What a university teaches about them should reach the people who need it, every week, in plain language, without enrolling anywhere. That is what I do here, and it is what I do on Thursday mornings in a seminar for the Renaissance Society at Sacramento State, built on this publication, where anyone can send a question and hear it answered.

In 2003 my doctorate at Syracuse measured what a machine can infer about a person from the traces they leave, and what I found became the privacy course I still teach. In 2007 and 2009 I sat on the committee at ETS in Princeton that set the national standard for digital literacy, and from 2014 to 2016 I was the California State University's ICT Literacy Ambassador. Six names, one job: helping people not be at the mercy of the newest machine.

I have also worked on the other side of the counter. In 2000 I began making films that a machine assembles differently for every viewer; the first of them took first prize at the Moscow International Film Festival in 2004, two decades before the industry shipped a generative model. A generative video poem I made with a poet at UNAM is in the field's definitive anthology, and in 2025 my AI short film The Human Futures Market took Best AI Short Film at the Red Rocks AI Film Festival. My recent published research is on AI and elections, and on how these models represent the cultures they were not trained on. I do not read the papers myself, and this page will never tell you I do. I built the machine that reads them and wrote the rules it reads by, and I analyze what it gives me the way I have taught my students to analyze: asking how it knows, and who benefits from what it says.

The vendors began teaching AI literacy two years ago, because before that there was nothing to sell. I am not against them. I am simply not one of them, and no one who sells these systems pays for this page. Your subscription does, and that is what keeps it honest.

The record, dated
  • 1989 Teaching begins, Mexico City
  • 1995 Hypergraphia, an early Spanish-language web and hypermedia publisher
  • 1996 Mercadotecnia e imagen en Internet, with Jesús del Toro, a book on marketing on the Internet, among the first in Spanish, Editorial Iberoamérica
  • 1999 Autorretrato, first generative poem, published by UNAM's Instituto de Investigaciones Filológicas
  • 2003 Ph.D., Syracuse University, Doctoral Prize, for the dissertation The Medium Is the Measure of Itself
  • 2004 A Space of Time: First Prize, Internet/Multimedia, XXVI Moscow International Film Festival
  • 2007, 2009 ETS iSkills committee, Princeton
  • 2008 Making Sense of Tracking Data, a book on how the traces a person leaves online allow inferences about them, including how they think, VDM Verlag
  • 2011 Keynote on generative work, UNAM
  • 2014–2016 ICT Literacy Ambassador, California State University · California Open Educational Resources Council
  • 2019 Big Data, with Rodolfo Mata: Electronic Literature Collection, Volume 4
  • 2020 Tablada Hipertextual, with Rodolfo Mata, UNAM: the computer reads a poet's words and builds a different reading from them, the ground the language models now work on
  • 2021 Uku Pacha, a generative VR film: world premiere, Vancouver International Film Festival
  • 2023 Keynote on the representation of Latin American cultures in AI, Ibero-American Meeting of Audiovisual Narratives, Trujillo, Peru · Graduate teaching at UNAM begins
  • 2023–2026 Directed the faculty learning community Advancing AI Literacy, Sacramento State · Principles for Cognitive and Pedagogical Sovereignty
  • 2024 A weekly AI publication run on real pipelines, every week since
  • 2025 Best AI Short Film, Red Rocks AI Film Festival
  • 2026 Keynote, with Nancy Huante-Tzintzun, Latin American and Caribbean Digital Humanities symposium, UT San Antonio · AI Watch, a weekly seminar for the Renaissance Society at Sacramento State

The research record →

TertulIA: weeklyAI.watch/tertulia/en/Pictures are illustrations generated by weeklyAI.watch, never photographs