AI and Power Systems Are Rewriting Grid Infrastructure Rules as Energy Demand Surges

How AI and power systems are rewriting each other's rules

The task now is to build for a world in which intelligence is infrastructure - and in which decisions about AI procurement, grid expansion and energy storage are no longer made in separate rooms. What happens when two critical infrastructures - AI on the one hand, and our electricity infrastructure on the other - start to rewrite one another's rules? That is the question grid operators, hyperscalers and energy policymakers are now asking. AI is changing how electricity systems are planned, built and run. Electricity systems, in turn, are setting the boundaries of what AI can become. The grid and the algorithm are no longer separate stories. This convergence matters because their clocks are out of sync. The International Energy Agency (IEA) estimates that electricity consumption from data centres could roughly double by 2030. But infrastructure buildouts do not move at the speed of software. New transmission lines can take years to permit, finance and build; new AI workloads can arrive in months. That mismatch between infrastructure and computational timescales is becoming one of the defining features of the decade ahead. The World Economic Forum's Innovation Playbook for Future Power Systems captures this shift with clarity. What emerges from its case studies is a portrait of three core themes across this mutual transformation: 1. AI is entering the operating core of power systems; 2. Power systems are becoming a binding condition for AI deployment; 3. AI is emerging as a mechanism through which grids absorb shocks. The future of power will be determined by how intelligently this emerging system is managed. Here's what you need to know about each theme. The clearest sign that AI has moved from being a tool for the grid to becoming part of the grid itself is the fading distinction between energy producer and energy user. Industrial facilities, commercial buildings and distributed storage assets are beginning to run continuous optimization loops, deciding when to generate, store, consume or sell electricity. Take, for example, Envision, a Chinese company that provides wind turbines, energy storage systems and energy management software. Its AI Power System for industrial parks shows what this looks like at the edge of the grid. "Industrial facilities that were once passive users of electricity are becoming autonomous energy managers, continuously balancing on-site solar, battery storage and grid interaction without compromising production," Lei Zhang, Envision's Founder and CEO, said. The implications extend beyond any single facility. When industrial users become active grid participants, the demand side of the electricity system acquires the intelligence that has historically sat mainly on the supply side. The grid gains not just more assets, but more decision-makers: millions of buildings, batteries and industrial sites are able to absorb variability and respond to prices. AI's growth, however, is creating a category of electricity demand that existing power systems were not designed to serve. This is thanks to both the volume required, and the character of that volume. AI data centres require power that is highly reliable, increasingly clean and available on timelines that do not match conventional grid-expansion cycles. The problem is not only that this demand is large; it is that it is arriving faster than the grid can respond. China-based battery and storage solution provider Hithium is working to address that gap. Their approach pairs long-duration storage at the energy source with a lithium-sodium system at the load side, giving operators grid-scale reliability and millisecond response on deployment timelines of one to two years rather than five to ten. For hyperscalers, that storage - and the ability to scale it up quickly - is becoming a prerequisite for AI deployment at scale. "AI is scaling faster than power infrastructure can be built. Fast-response storage combined with long-duration flexibility provides the reliability that AI data centre operators require, while enabling co-location of the large renewable capacity that hyperscalers increasingly demand," Dr. Nazar Yi, Hithium Board Member and Vice President, said. China's new energy plans for the 15th Five-Year Plan period (2026-2030) reflect this convergence, calling for closer coordination between computing, power and renewable-energy infrastructure, linking "power for AI" with "AI for power". The third current is the least visible and may prove the most consequential: AI as the means by which power systems see trouble, absorb it and recover. As grids carry more variable renewables, and as extreme weather and cyber risks multiply, stability is becoming less a matter of reserve capacity alone and more a question of speed. A system that cannot react fast enough cannot remain stable. "In the AI era, resilience is no longer defined by backup power alone. It is the ability of energy storage systems to sense, respond and adapt in real time, keeping critical infrastructure operating even as power systems become more dynamic and uncertain," Dr. Nazar Yi, Hithium Board Member and Vice President, said. The next step is infrastructure with intelligence stitched through it. Human operators work in minutes; AI-enabled systems can detect anomalies, reroute flows and coordinate storage in milliseconds. Such capabilities change the economics of resilience: intelligent protection is becoming more efficient just as outages for supply chains, data centres and financial systems are becoming more expensive. These three currents are already interacting and converging. AI systems that manage industrial energy autonomously are also the assets that virtual power plants aggregate. Storage systems designed for data-centre reliability are also buffers for renewable-heavy grids. Resilience intelligence that protects a facility also contributes to the grid-level stability that keeps digital infrastructure online. That means decisions once made in separate rooms - power purchase agreements, grid expansion plans, cybersecurity protocols and AI procurement - are now interdependent. Technology companies that separate AI infrastructure decisions from energy security create fragility. Grid operators that plan transmission without modelling AI-driven demand behaviour plan for a load that no longer exists. Policymakers who regulate energy storage for grid stability but not for compute competitiveness miss the system they are actually governing. The deeper question is no longer whether grids and algorithms will become interdependent. They already are. The task now is to adapt and build for a world in which intelligence is infrastructure.

Can AI solve the energy problem it created?

We have all the electricity we need to power AI's data center demand -- and AI can unlock it for next to nothing. It costs about $5 and takes roughly 15 minutes to train an AI model that can help determine where America's next data centers should be built. That's not a rounding error. The world is debating hundred-billion-dollar AI investments and warning that we need fleets of new power plants, yet one of the most important pieces of the puzzle costs about as much as a pack of gum. Before we build another plant, we should figure out how much of the grid we're already leaving on the table. Much has been made of AI's power requirements, and for good reason. Bloomberg projects that data centers will consume up to one-fifth of all power in the U.S. by 2035 -- up to 200 gigawatts -- with much of that demand landing in regions where power is already constrained. Power generation wait times can stretch past five years on a grid where much of the infrastructure is 50 to 70 years old. The pressure is real, and it's arriving faster than utilities can plan for it with current systems and processes. What most people don't realize is that our existing grid massively overproduces the energy we need, and much of it sits idle. Power systems are sized for the handful of hours a year when demand peaks, the hottest afternoons and the coldest mornings. The rest of the time, that capacity sits there, already built and already paid for. Across most hours, we use only about half of it. It doesn't take a math degree to appreciate that we already have the power we need; we just need to ensure the power that would go wasted is put to good use. This is a planning problem, and planning problems are exactly what AI is good at. New models can evaluate where and when the grid has unused power, along with the costs and timelines required to build the infrastructure that reroutes it, which I call capacity mining. It lets the AI revolution advance at pace while optimizing longer-term backbone investments and minimizing backlash. This doesn't require any new hardware or a scientific breakthrough. It requires asking the right questions with the right, highly efficient, and inexpensive models. Here's the sequence that will make this work for everyone involved. First, apply AI models to match data center demand against the headroom that already exists, citing locations and operating profiles where there's available power and operational flexibility rather than forcing utilities to build new generation and transmission capacity around a site chosen for other reasons. Second, where matching alone isn't enough, apply AI to optimize the expansion of transmission lines and battery systems to route power to where it's needed or store it until it is. New power plants, which are costlier, slower to build and often polluting, are the last resort. In most cases they're avoidable. The result is a buildout that moves at the speed of software, not steel and permitting. Getting that order right also helps utilities deliver on their recent White House pledge to shield consumers from rising electric bills tied to data center growth. Capacity that already exists doesn't need to be paid for twice, and finding it costs relatively nothing. But the larger point is that data centers don't have to raise bills at all, and can actually lower them. At one large investor-owned utility, we've calculated that large loads, including data centers, can generate roughly $1 million per megawatt per year in new revenue, about $1 billion per gigawatt. Structured correctly, that revenue offsets fixed grid costs every other customer currently shoulders alone. Reliability works the same way. A buildout done right pulls investment into stronger transmission, better monitoring, and improved protection and control systems, and those upgrades serve every household on the line, not just the hyperscaler at the end of it. I also believe hyperscalers will gladly cover more of the cost in exchange for faster access to power. Think of a data center as a mini-utility. Its buildout should put money back into the community it lands in, both in quality local jobs and in more modern, reliable infrastructure for everyone on that grid. There's no good reason to make regular people pay for electricity that data centers use, and deflating that argument will do more to cool the backlash than any ad campaign. This is a crossroads for utilities, which are not historically growth businesses and have been flat for decades. If they do data centers right, they can drive the most revenue growth in a generation while optimizing their grids for every other customer. Get it wrong, and they'll create stranded assets and raise rates, push hyperscalers to build behind-the-meter generation themselves, lose that revenue entirely, and miss the biggest modernization opportunity they'll ever get. We don't have to choose between winning the AI race and protecting communities from unnecessary costs. The better question is whether AI can solve the problems AI is creating. On this one, it can, and with tools that are ready today. We have all the power we need; we just need to unlock it.