COP31 Launches Antalya AI Pledge as Infrastructure Realities Expose the Environmental Limits of Compute
The COP31 Antalya AI Pledge sets green targets, but Brazil infrastructure conflicts and missing computing superpowers expose real resource limits.

During the COP31 climate summit in Antalya, Turkey, the conference presidency and the International Telecommunication Union officially launched the Antalya Artificial Intelligence Pledge, establishing the first multilateral declaration structured around the environmental footprint of machine learning. The initiative frames computing infrastructure as both a necessary catalyst for decarbonization and an accelerating source of physical strain. While international delegates promoted algorithmic efficiency as a tool for ecological modeling, the pledge itself cites International Energy Agency findings noting that electricity consumption from AI-focused data centres significantly rising by 50% in 2025.
This tension between climate ambition and computational appetite outlines the central dilemma confronting global digital policy. While artificial intelligence is marketed by corporate developers as an essential instrument for optimizing power grids and climate adaptation, training and running modern foundational models requires expanding volumes of baseload electricity, cooling water, and physical grid capacity. The Antalya agreement attempts to establish shared transparency and efficiency principles, yet its arrival highlights an intensifying confrontation between software scale and ecological boundaries.
The Antalya AI Pledge and the Missing Compute Superpowers
The Antalya Artificial Intelligence Pledge organizes international efforts across four foundational pillars: powering computing hardware with low-emission energy, applying circular principles across equipment life cycles, standardizing environmental impact measurement through frameworks like ITU-T L.1801, and expanding technical support for developing economies. According to the official text released by the COP31 presidency, signatories agree to share operational metrics and participate in structured progress assessments conducted through the ITU Green Digital Action initiative.
The diplomatic scope of the document, however, reveals structural limitations. The declaration affirms explicitly that its targets are voluntary and non-prescriptive, create no new obligations for participating entities. Reports published during the summit by Agence France-Presse and The Japan Times indicated that while approximately 80 international organizations and industry trade groups endorsed the pledge, only nine countries had formally joined the agreement at launch.
The list of initial signatories omitted the three jurisdictions that dominate global artificial intelligence development: the United States, China, and the European Union. Furthermore, foundational frontier labs and chipmakers including OpenAI, Anthropic, and Nvidia did not sign the agreement, though an industry association representing companies such as Amazon, Google, and Meta appeared among supporting entities. The environmental coalition Beyond Fossil Fuels criticized the voluntary architecture, warning that non-binding declarations risk providing corporate public relations cover without requiring enforceable operational cuts in fossil fuel reliance.
The Efficiency Paradox and Thermal Constraints
The core assumption behind voluntary technology pledges is that rapid advances in algorithmic efficiency and hardware architecture will naturally temper resource demand. Within computer engineering, however, this expectation frequently encounters the Jevons paradox: reducing the computational and economic cost of executing an inference query typically multiplies aggregate query volume, autonomous agent deployments, and continuous background processing tasks.
Physical constraints further complicate facility engineering. In data center cooling architectures, operators often face a direct technical exchange between water consumption and electricity draw. Closed-loop dry cooling and air-assisted fan arrays reduce evaporation losses from local aquifers, but running large mechanical cooling systems 24 hours a day drives up electrical usage, as researchers from Brazil’s Institute for Research on Law and Technology (IP.rec) noted during legislative hearings. Conversely, evaporative water cooling reduces facility electricity strain while consuming millions of liters from regional water reserves.
These physical balances carry direct geographic consequences. Global data center electricity demand is projected by the International Energy Agency to generate 350 million metric tons of carbon dioxide equivalent by 2035 in central forecast scenarios, with capital expenditures by major cloud operators expected to exceed $700 billion in 2026. A Harvard University study examining 403 large American facilities between May 2024 and April 2025 measured 82 terawatt-hours of power consumption and nearly 26 million tons of emissions, with over half the energy originating from fossil generation. Crucially, the IEA estimates that nearly 40% of identified AI data center capacity operates in regions already experiencing elevated or extreme water stress.
Brazilian Data Centers Face Grid Strain and Watershed Conflicts
This global computational demand is increasingly moving into developing markets that offer land, cooling water, and clean energy matrices, and Brazil has emerged as a primary target. In September 2026, an investigative report published jointly by Mongabay and The Intercept Brasil revealed that Atlas Renewable Energy is planning a 1-gigawatt data center in the Paracatu municipality, located in the state of Minas Gerais. The proposed complex, budgeted at 35 billion reais ($6.8 billion) across 13 computational facilities planned through July 2030, would exceed the approximately 800 megawatts of combined capacity installed across all existing data centers in Brazil.
Paracatu illustrates the acute environmental friction generated by gigawatt-scale computing. According to municipal water basin data, around 80% of the territory faces water conflicts, and local demand already exceeds available groundwater replenishment rates. Under efficiency thresholds established by proposed federal legislation, the facility could consume up to 1.2 million liters of water per day, representing 438 million liters annually from the Cerrado headwaters. The investigation documented at least 68 data centers are planned for the country in 37 different municipalities seeking grid interconnection points.
Similar structural friction reached the judicial system in Ceará. On September 10, 2026, the Federal Public Prosecutor’s Office (MPF) and the Public Defender’s Office (DPU) filed a civil action (Case 0080626-66.2026.4.05.8100) to halt the operation of the Data Center Pecém in Caucaia. The 300-megawatt project, developed by Omnia, occupies roughly 70 hectares and incorporates 120 diesel backup generators and underground aquifer extraction less than two kilometers from the Lagamar do Cauípe preservation area and the Anacé indigenous territory. The prosecutors challenged the state environmental agency’s decision to classify the complex as a medium-impact development without requiring a full Environmental Impact Assessment (EIA/RIMA) or conducting free, prior, and informed consultation with the Anacé people under International Labour Organization Convention 169.
Legislative Scrutiny and the Annual Accountability Timeline
The expansion of massive artificial intelligence facilities has triggered legislative debate in Brasília. During a public hearing conducted by the Senate Science and Technology Commission on August 4, 2026, consumer advocates from the Brazilian Institute for Consumer Protection (Idec) warned that data centers operate with continuous baseload profiles that lack peak-hour flexibility. Because utility transmission upgrades and substations are integrated into broader system costs, unmanaged infrastructure additions risk transferring grid expansion expenses onto household electricity tariffs without guaranteeing domestic scientific or economic retention.
Federal lawmakers are reviewing Bill PL 3.018/2024 to establish technical standards and water protection criteria for hyperscale computational facilities across Brazil. On the international stage, the International Telecommunication Union announced that collective progress under the Antalya AI Pledge will be tracked at each annual United Nations climate summit through 2030, starting with COP32. These complementary legislative and multilateral frameworks will determine whether regional hosts of artificial intelligence infrastructure can enforce measurable resource protections, or whether the environmental cost of digital scaling will remain externalized to local communities.
Written by Adriano Nascentes
Founder & Lead AI Systems Analyst · Verified Staff
Founder and editor-in-chief of Unknown Observer. Technology researcher and software engineer investigating foundation models, autonomous agent runtimes, cognitive architectures, and frontier AI safety.
Was this article helpful?
Your feedback helps our editorial team refine reporting depth and accuracy.



