Energy barriers surpassed
1,000 TWh
In 2026, global data centers already break the 1,000 TWh barrier, and AI accounts for roughly 30% of that load.
Sustainability
Growth scenarios, energy tension, and sustainability as demand scales.
AI expansion is no longer just about benchmarks. This report projects how three forces collide: consumption growth, the geopolitical fragility of infrastructure, and the economic pressure of deployment.
Energy barriers surpassed
1,000 TWh
In 2026, global data centers already break the 1,000 TWh barrier, and AI accounts for roughly 30% of that load.
Efficiency improvement per token
> 280 x
Efficiency improves at a historic rate, but usage growth ends up absorbing those savings.
Potential pressure on electricity prices
+ 80 %
AI clusters are already straining water, grid, and generation capacity in several markets.
AISHA's conclusion is not that AI will slow down on its own, but that the combination of demand, infrastructure, and cost is pushing the system toward physical limits that are increasingly less abstract.
In 2024, AI was still a minor fraction of the sector's demand. By 2028, in a mass adoption scenario, it already competes with entire countries in cumulative consumption.
The bar stacks AI-attributable consumption against the rest of the data center fleet.
Efficiency per query improves radically, but the cheapening of usage accelerates demand much faster than per-inference savings.
The two curves no longer move together: technical improvement lowers the cost per query, while usage volume takes off.
Not all AI workloads are equal. Text, reasoning, image, agents, and video push infrastructure in very different ways.
Volume doesn't tell the whole story: reasoning, agents, and video concentrate far more energy per task.
Still concentrates volume, but with a small marginal cost. It is the most optimized modality.
Rapidly raises consumption per query by generating internal thinking and long inference chains.
Each generation consumes far more than a text query and scales poorly if repeated until getting it right.
The cost is no longer a single response, but an entire session with context, searches, and tools.
It is the most energy-aggressive modality and the one most likely to trigger commercial demand spikes.
The real sustainability of AI doesn't depend only on the model. It also depends on how it's cooled, what electricity mix powers it, and where it's deployed.
Includes both direct cooling and water associated with electricity generation.
The grid's carbon intensity continues to determine an enormous difference between deployments.
Efficiency improvements don't eliminate the physical problem: if growth relies on scarce water, local diesel or gas, and saturated grids, the total environmental cost keeps rising.
The demand for continuous 24/7 load forces the search for cheap power, cluster relocation, and acceptance of rising capacity premiums.
+ 833 %
The capacity auction spikes dramatically and the residential impact is passed through to the final bill.
+ 79 %
The combined pressure from AI clusters and transmission bottlenecks already projects strong wholesale price increases.
Cheap power
Companies seek hydroelectric, massive solar, and less saturated regions to sustain growth.
The pattern repeats: AI doesn't just consume energy — it also reshapes where it makes sense to install capacity, who pays for it, and which regions absorb the logistical friction of growth.
Same category
Wed Apr 01 2026 00:00:00 GMT+0200 (Central European Summer Time)
Dashboard to explore consumption, models, modalities, multipliers and real-world equivalences.
Wed Apr 01 2026 00:00:00 GMT+0200 (Central European Summer Time)
Síntesis prospectiva sobre emisiones, infraestructura y límites operativos de la expansión de IA.