Uploaded March 2026 | Updated September 2026, 2 hours ago
Every AI interaction—from training to inference—draws on electricity, water and infrastructure. But how much, exactly? Until recently, a lack of transparency has kept researchers, businesspeople, policymakers and even users in the dark. Now, new data from Google on Gemini’s per-prompt consumption shows that AI-specific chips account for more than half of total demand, with cooling, CPUs and idle backup systems making up the rest. How can efficiency gains in chips, data-centre design and model architecture keep pace with exponential growth in usage?
Moderator:
Vijay Vaitheeswaran
Global energy and climate innovation editor, The Economist
Speakers
Carol Yan, General manager, energy and utilities UKI, Amazon Web Services
Bruno Berti, Senior vice-president Global Product Management, NTT Global Data Centres
Maud Texier, Global vice president, data centres and industrials, Octopus Energy
Supported by NTT Global Data Centers
Every AI interaction—from training to inference—draws on electricity, water and infrastructure. But how much, exactly? Until recently, a lack of transparency has kept researchers, businesspeople, policymakers and even users in the dark. Now, new data from Google on Gemini’s per-prompt consumption shows that AI-specific chips account for more than half of total demand, with cooling, CPUs and idle backup systems making up the rest. How can efficiency gains in chips, data-centre design and model architecture keep pace with exponential growth in usage?
Moderator:
Vijay Vaitheeswaran
Global energy and climate innovation editor, The Economist
Speakers
Carol Yan, General manager, energy and utilities UKI, Amazon Web Services
Bruno Berti, Senior vice-president Global Product Management, NTT Global Data Centres
Maud Texier, Global vice president, data centres and industrials, Octopus Energy
Supported by NTT Global Data Centers
