NextFin News - Artificial intelligence is no longer just a software story. It is becoming a power story, and the bottleneck is shifting from chips and models to wires, transformers, substations, storage and the grid queues that decide when new load can actually connect. Morgan Stanley said on June 26 that the race to supply reliable electricity for AI is already reshaping investment priorities, while Goldman Sachs and the International Energy Agency both point to a rapid rise in data-center power demand that is forcing utilities, developers and policymakers to rethink how the next wave of infrastructure gets built.
The reason the market is paying attention is simple: power is turning into a gatekeeper for AI growth. Morgan Stanley said the world is entering the early stages of a structural shift in power consumption, driven by the rapid expansion of artificial intelligence, data centers, electrification and industrial reshoring. Goldman Sachs says U.S. data center power demand is expected to climb from 31 gigawatts in 2025 to 41 gigawatts in 2026 and 66 gigawatts in 2027. The IEA says electricity consumption from AI-focused data centers surged 50% in 2025 and could triple by 2030, while total data-center electricity use is projected to roughly double to 950 terawatt-hours and approach 3% of global electricity demand.
That combination matters because the next phase of AI buildout is no longer limited to servers and semiconductors. Every additional megawatt of load requires real-world assets: land, transmission, interconnection, backup generation, cooling and often storage. The result is a capital cycle that looks less like a pure technology trade and more like an infrastructure race. Companies that can secure electricity faster may win more projects. Regions that can expand power systems faster may attract more investment. And utilities, grid-equipment suppliers, gas infrastructure firms and storage developers may all become more central to the AI story than they were when the narrative was only about compute.
The shift is also changing how projects are financed and sited. As the amount of power required rises, the constraint is moving from theoretical capacity to actual delivery. That is why time-to-power is becoming a critical metric for data-center developers, and why the market is seeing more interest in off-grid and behind-the-meter solutions. The AI boom is starting to look like a demand shock that reaches far beyond the chip supply chain.
Power Availability Is Becoming The New Site-Selection Filter
The clearest takeaway from the Morgan Stanley discussion is that the grid itself is now part of the competitive landscape. The bank said that in the United States through 2028, roughly 80 gigawatts of data centers would need to be powered up, but only about 25 gigawatts are expected to have grid access within that period. That leaves about 55 gigawatts that would not have the standard ability to connect to the grid. In practical terms, that is a severe bottleneck. It means the largest constraint on AI infrastructure is no longer just capital or chips, but the ability to secure dependable electricity on a usable timeline.
That is why Goldman Sachs’ location analysis is so important. The firm said power availability and time-to-client are the primary factors driving where data centers choose to locate. That marks a shift in investment logic. Land, tax incentives and proximity to fiber still matter, but they are no longer enough if a project cannot get through the interconnection queue. The market is effectively repricing regions with faster access to grid capacity and punishing those where power delivery is slow or uncertain.
The same pressure is showing up in the broader energy economy. When data centers require more power, utilities need more generation, more transmission and more distribution investment. That creates a second-order demand wave for transformers, switchgear, substations and engineering services. It also creates a timing problem: the AI industry can move in months, but the power system often moves in years. That mismatch is the heart of the current story.
The result is a reordering of who has leverage in the AI buildout. Hyperscalers still control demand, but utilities control the connection, and project developers that can solve power first are likely to move ahead of competitors. In that sense, electricity is becoming a form of strategic capacity, much like memory chips or advanced packaging were in earlier phases of the AI cycle.
“Power availability and time-to-client are the primary factors driving where data centers choose to locate.”
That sentence from Goldman Sachs captures the market shift in one line. The relevant edge is not just the cheapest kilowatt-hour; it is the kilowatt-hour that arrives on time.
Off-Grid And Behind-The-Meter Solutions Are Moving From Backup Plan To Core Strategy
As the grid constraint tightens, hyperscalers are increasingly looking for ways to bypass it. Morgan Stanley said many are pursuing “time-to-power” strategies that prioritize faster deployment through alternatives to the traditional electric grid. The bank’s roundtable described a market in which turbine islands, fuel cells and other onsite power systems are becoming increasingly relevant because they can be deployed quickly and provide reliable electricity while grid access remains limited.
This matters because AI economics punish delay. A data center that cannot get power on time cannot train models, serve inference or monetize capacity. That delay can be more expensive than paying a premium for a private power solution. As a result, what used to be a contingency plan is turning into a deliberate development strategy. The shift is visible in the growing interest in batteries, microgrids, onsite generation and hybrid power architectures that combine grid supply with private backup or primary generation.
The IEA’s analysis helps explain why that shift is accelerating. It said electricity use from AI-focused data centers surged 50% in 2025, while total data-center electricity demand is on track to reach 950 terawatt-hours by 2030. That scale of demand does not just require more electrons; it requires more flexibility. AI workloads can create large and rapid power swings, which makes storage and dispatchable backup more valuable. The IEA said around 20-25 gigawatts of battery storage could be installed in data centers globally by 2030, potentially turning them into grid assets if incentives are aligned.
That opens a new valuation framework for the market. The companies most exposed to the AI power buildout are not only utilities. They also include firms that can sell flexibility: battery developers, power-electronics providers, gas-turbine suppliers, fuel-cell manufacturers and companies that can integrate multiple power sources into one campus. The shift is also favorable for infrastructure financiers because the projects are capital-intensive, long-duration and contract-driven.
In other words, AI is not just pulling electricity into the economy. It is pulling a much broader set of industrial capabilities into the center of the technology trade.
The Morgan Stanley panel said “time-to-power solutions are coming into the picture.”
That is the key phrase. If time-to-power becomes the real bottleneck, then the winners are the developers that can solve reliability first and scale second.
Why This Energy Cycle Looks Different From Earlier Technology Booms
The current cycle is different because the load is large, fast and localized. Earlier technology booms did not require this kind of immediate physical buildout across the power system. AI does. Morgan Stanley said U.S. electricity demand is likely to rise at an annual average rate of 2.6% over the next decade, surpassing previous peaks, and that data centers are projected to account for a significant share of that growth. Goldman Sachs says U.S. data center power demand alone could rise by 35 gigawatts between 2025 and 2027. The IEA says AI-focused data centers tripled their electricity growth trajectory relative to the broader data-center market in 2025.
The implication is that this is not a standard demand cycle. It is a structural shift in the way power systems are used. Data centers cluster in specific geographies, which makes local demand spikes more intense than broad national averages imply. A region may have enough aggregate generation on paper, yet still face shortages because transmission is congested or nearby substations are full. That means the grid’s weakest links matter more than its headline capacity.
It also means utility planning is getting more complicated. A traditional utility buildout is slow and regulated. AI demand is fast and concentrated. The result is a mismatch between the speed of private investment and the speed of public infrastructure. That mismatch can produce rising capex, more rate-case pressure, longer equipment lead times and a broader scramble for dispatchable power.
Crucially, the market is not reacting to one company or one region. It is reacting to a systemwide squeeze. The same constraint can support multiple parts of the energy stack: transmission owners, equipment manufacturers, gas developers, storage providers and firms that can offer fast, private power. The story is therefore less about a single utility winner and more about an ecosystem of beneficiaries tied together by scarcity.
“The electricity market is facing a new reality of capacity constraints, particularly around grid infrastructure and power availability.”
That Morgan Stanley line suggests the AI buildout is already changing how the power market thinks about scarcity. Instead of asking only how much electricity exists, the question is whether it can be delivered where it is needed, when it is needed and at a price that preserves project economics.
What Happens Next Depends On How Fast The Power Stack Can Scale
The next stage of the story will be decided by execution. The obvious catalysts are utility capital plans, interconnection timelines, transmission approvals, equipment lead times and the pace at which hyperscalers sign contracts for onsite or off-grid power. If those processes accelerate, more data-center projects can move forward. If they stall, the AI infrastructure cycle could become more uneven and more expensive than the market currently expects.
Investors will also watch geography closely. Goldman Sachs said power availability and time-to-client are driving site selection. That means the most attractive regions are likely to be the ones that can add generation, transmission and grid access quickly enough to support AI demand. Areas with strong digital infrastructure but weak power delivery may find themselves at a disadvantage. Areas with power access and permitting flexibility may become the next magnets for data-center investment.
For the broader market, the message is that AI has crossed from a software-led boom into an infrastructure-led one. That changes the list of beneficiaries and the pace of returns. Semiconductor demand can move quickly. Power infrastructure cannot. The payoff from the AI cycle may therefore be spread across a wider set of assets, but it may also arrive more slowly, through permitting cycles, capital expenditure programs and long-term contracts rather than headline growth alone.
The central conclusion is straightforward. AI still begins with compute, but it increasingly ends with electricity. And in a constrained power market, the most valuable asset may be the one that can deliver reliable megawatts on schedule.
That is the new economics of AI infrastructure: not just more demand, but a race to the grid.
Explore more exclusive insights at nextfin.ai.
