Will AI break the grid?

Will AI break the grid?

Chris Pennington

As we start the year, the data center industry is facing a cluster of energy and sustainability challenges that at first sight seem hard to solve simultaneously.

We have seen phenomenal load growth and rocketing demand for new capacity, to the point that nearly all of the readily accessible grid power has now been reserved

At the same time AI is shifting from a predominantly training model to a far more geographically widespread architecture, as AI inference takes off and nations have prioritized building domestic AI capabilities. This will create many local demand pinch points.

While this headlong rush takes place, the need to battle climate change remains intact. The reality of global warming far in excess of the 1.5 degree level targeted by climate scientists is looming closer, accelerating the need to decarbonize faster.

Maybe we can tackle one or two of these challenges, but can we address all three? Over time, the answer must be yes. Because the spirit of innovation and collaboration in our industry is strong, we are developing new practical approaches based on new partnerships, renewable energy, and improved reporting.

Data center demand is a local challenge

The best way to tackle the power demand issue is by putting it in perspective. Globally, data centers still account for just 2-4% of world power usage. According to the latest World Energy Outlook from the IEA, AI will account for nearly 10% of the growth in electricity demand through 2030. While 10% of global demand growth is significant, it is worth noting that this is around the same increase as for space and water heating, and less than the forecast growth in demand for space cooling (12%), vehicle electrification (14%), and appliances and fridges (17%).

Demand is a far more critical issue in specific locations, and there are many communities where data center loads are pushing beyond 30% of total electricity use. This will be further extended by the next phase of AI infrastructure development. An increasing proportion of data center demand will be driven by the sharp growth in inference, where low-latency access to large numbers of users will be a critical factor. According to Structure Research, AI inferencing infrastructure will overtake training infrastructure in the coming year, growing at 79% through 2030, by which time it will account for over 80% of all AI infrastructure.

This means we are looking at a mixed landscape of gigawatt-plus campuses for training in selected markets that can accommodate this scale, plus many smaller AI-focused extensions to existing availability zones in mature, densely populated markets. Many of these markets are already power-constrained. Finding practical, local solutions to the stress this puts on the grid will be a key challenge. As data center loads emerge as the largest load in more local grids, the need for data center load flexibility in support of reliable and affordable power is emerging as a key opportunity.

Flexible power can help grids adapt

How might this flexibility work? Earlier solutions for balancing generation and demand will need to evolve. Large loads have the opportunity to move from being simply utility customers to actively providing solutions to grid challenges. In other words we can be utility partners, not just customers:

• Shifting the load

  Battery energy storage systems (BESS) have emerged as an accessible and impactful behind the meter (BTM) solution. Coupled with a large data center load and the flexibility it achieves can unlock gigawatts of stranded capacity by reducing power on the grid during peak demand. Iron Mountain Data Centers are currently pioneering a 12 MW BESS in New Jersey. On track for construction this year, it will be capable of supporting the facility load for over an hour, essentially removing load from the grid during peak grid demand.

  Data center operators are also designing fully islanded microgrids that meet the need for new demand without connecting to local grids. Where connection queues are very long, full BTM builds can be completed long before a grid connection becomes available, and potentially become grid-serving assets in the future when connected to the grid. While these microgrids use fossil gas generation, in some markets this can contribute to reducing grid emissions at the same time as providing critical grid reliability.

• Varying the load

  Load management can also help in some cases. Some sites may support the local grid by ramping workloads up and down instead of shifting the time of load with batteries. This is not something that will develop at industry-wide scale, but where possible it can contribute a low carbon solution for grid stability. It is already being demonstrated by operators with more flexible loads like the Nvidia Omniverse DSK Blueprint.

• Rationalizing the risk

  With the considerable infrastructure investment needed to support future demand it is critical that we have a fair and accurate view on what that future load will be. Speculative load requests have driven up projected grid costs, and some utilities have responded by requiring substantial financial assurances in “take or pay” deals. This is an understandable effort to shield all rate payers from the cost of stranded assets, but it can tie up large amounts of capital and credit, slowing investment.

  Clarity, accurate planning and the avoidance of duplicated risk make for fairer contracts - for example, the Data Center Coalition, together with E3 have made recommendations for an approach that is fair, non-discriminatory, and justified, based on quantifiable utility exposure.

Renewables have passed a major milestone

The future for renewables is bright. In the first half of 2025, renewable energy overtook coal as the world's leading source of electricity production (source: Ember). In fact, the growth in solar and wind was so strong that it effectively met all of the new electricity demand, helping drive a slight decline in coal and gas use. Over 90% of new grid assets are carbon free, clearly demonstrating that wind, solar and storage have moved from the margins to the mainstream.

This is because renewables are no longer simply the right choice to address the climate crisis, they now deliver overwhelming advantages in terms of both energy security and lower cost. These advantages will continue to grow as renewables innovate and scale further, bringing unit costs down, while fossil fuels fail to find comparable advantages. You simply need to look at the hourly carbon heatmap of power on one hand, and track it against the market electricity price on the other. The lower the carbon the lower the cost, as seen in European national grids over the last five years in this excellent illustration from Energy Tag.

Here again, we need to remain mindful of local communities. While grids get cleaner globally, markets with large load growth will see part of that need met by installing new fossil gas fueled generation plants. The journey to wider reaching grid decarbonization requires investment in transmission that can get clean power to the new loads. This will take time.

Corporates climate targets

We have moved on from the era of promises and pledges that followed the Paris Agreement, to an era of performance and progress. While goals remain an important part of driving change, it is also important to achieve this change in practice.

Examples of this new era include recent regulations for increased reporting on emissions and data center efficiency, as well as the work currently underway to increase the granularity of the greenhouse gas reporting protocols. We can expect recalibration of corporate emissions targets in response to this which will lead to clearer, more objective reporting in the long run.

This new landscape of increased transparency and granularity of data will enable us to focus on genuine, incremental progress for lower carbon solutions. Companies that continue to innovate in this area will have a great new platform on which to share the details of their performance.

Working together for a successful transition

AI infrastructure will not break the grid, but there will be many practical challenges to solve. The medium-term picture is mixed but the longer-term outlook is positive as we enter an era of unprecedented access to new energy sources, new ideas and increasingly detailed data.

Emissions in developed markets may rise, particularly in the USA, as gas-fueled generation is used to meet part of the immediate need for load growth. Unlocking existing capacity to optimize grid utilization is critical to meet new load demand, with a focus on affordable power.

Progress is also being made in providing support for grids as they decarbonize through bringing more clean energy resources online to serve growing demand. As new solutions require innovation and negotiation, collaboration will be the key to unlocking this growing set of solutions.