How AI Data Centers Are Challenging the Electricity Grid

The system operator of Mexico's electricity grid, CENACE, must fully understand the behavior of demand to properly plan for both the day-to-day generation and the grid's expansion. To do this, it projects demand yearly, monthly, weekly, and hourly, based on user behavior — during a World Cup soccer match, for example — and the latest technological trends in industry. 

CENACE also needs to have a close relationship with the users that create the biggest disruptions in the system. Take the steel industry. This well-studied industry creates massive peaks of energy during its processes. When a large steel plant runs its electric arc furnaces to melt steel, it increases its demand by about 300MW for around 30 minutes (imagine the city of Merida appearing and disappearing from the grid half an hour later). They do this several times a day, and CENACE knows exactly when these furnaces fire up. However complex, the system operates within known variables.

AI is the new industrial load reshaping the grid. Steel took decades to become legible to CENACE. AI is a new industry and yet, it's already consuming more electricity than many industries. Nobody — not the grid operators, not the data centers, not the people training the models — can tell you with precision what that consumption will look like in three years. 

Today’s widespread AI is based on Large Language Models (LLMs). These are models whose only purpose is to predict the next word or character in a sentence. In order to make these predictions, the computer has to feed the text through a series of millions — and in most cases billions — of parameters that calculate the final result. The process is non-deterministic, meaning that the same input can deliver different results in every iteration. This alone makes the estimation of energy consumption very challenging.

On top of that, the AI industry is using these text prediction capabilities to implement autonomous agents. This is software that runs in loops, using one or several LLMs in every turn to decide how and when to take some action. This takes the energy consumption to a different order of magnitude compared to just the text prediction of an LLM. The promise of higher output and reduced costs is rapidly making enterprises transition to this new paradigm of productivity.

On the other side, we have the Mexican National Electricity System (SEN), a system that has proper generation conditions and reserves during the cold months of the year (September-April), but collapses as soon as the heat starts and demand for air conditioners peaks (May-August). 

Avoiding a data center exodus to countries with better energy conditions starts with a hard truth: no country has the kind of energy surplus data centers require at this scale. This is especially true when we consider the variables that have forced electricity systems to be in constant stress and evolution: decarbonization targets, the electrification of mobility, variable renewables, and a constant increase of extreme weather events, including heat waves.

The difference between data centers in Mexico or in other countries raises the question of whether or not we will be able to adapt our system, our regulation, and our technologies to adapt to this new massive demand. Both the grid and the data centers will need to contribute to achieve the energization of these new projects. 

Considering this complex scenario, some possible solutions are: 

(1) Integrating data centers under a series of regulations to maintain a close interaction with CENACE, just as the steel industry did when it was ramping up. This includes reducing the variability of the demand through mandatory on-site flexibility, collaboration in generating demand profiles, and an ongoing dialogue mechanism to identify where in the grid large data centers make the most sense to integrate, as well as identifying the ideal "size" that keeps the industry competitive without concentrating too much load on a single point in the Mexican grid. 

(2) Request a minimum of on-site generation for every data center in the country. Currently, the legal figure “Autoconsumo” (self-consumption) is growing and becoming more accessible, turning this condition into a profitable project for the private and public sectors.

(3) Support state governments to develop decentralization programs for the promotion of Distributed Generation and Autoconsumo among all end users. This will reduce the pressure on the national grid as new actors disrupt the national economy.

Whether the national grid is capable of integrating data centers or not depends on its adaptability to the new environment. AI will become a necessity for most companies in the near future, and being able to run models in Mexico will help us take advantage of nearshoring in the short term, and even to be in a good position to address medium-term issues such as data sovereignty. 

AI is disrupting national grids today. In 20 years, a different technology will force the same conversation. The question is whether we build the regulatory muscle now, or relearn it from scratch every time.