The race to build more powerful artificial intelligence is creating an equally urgent race for electricity.
AI data centers require enormous amounts of power, yet new grid connections and major utility upgrades can take years. As demand grows faster than infrastructure can expand, technology companies are increasingly turning to expensive on-site generation or building facilities in locations chosen primarily for power availability.
Broc TenHouten, co-founder and CEO of Intrinsic Power, believes the industry is overlooking a faster solution.
The immediate problem is not simply that the grid lacks power. It is that data centers cannot use their existing connections efficiently.
Data Centers Are Built Around Their Worst Moments
Modern AI workloads do not consume electricity at a steady rate.
Thousands of graphics processing units can begin working simultaneously, causing power demand to surge within milliseconds. Those spikes create serious risks. If a facility exceeds the limit of its grid connection, breakers can trip, equipment can shut down, and instability can spread beyond the data center.
To prevent that outcome, operators often run facilities well below their maximum electrical capacity.
TenHouten says many data centers operate at roughly half of what their grid connections could theoretically support. That margin provides room for sudden demand spikes, but it also leaves valuable capacity unused.
In effect, facilities are designed around the possibility of an extreme moment rather than their typical operating needs.
Predicting Demand Changes the Equation
Traditional backup power and battery systems tend to respond after a change in demand has already occurred.
Intrinsic Power is developing a more proactive approach.
The company combines high-voltage power electronics, energy storage, distributed sensors, and artificial intelligence to monitor electricity moving throughout a data center. Its system analyzes activity across computing equipment, cooling systems, renewable sources, batteries, and the grid connection itself.
Rather than reacting to historical averages, the technology attempts to anticipate near-term changes and determine when stored electricity should be supplied or replenished.
By controlling spikes more precisely, data centers may be able to operate much closer to the full capacity of their existing grid connections.
For operators, that could mean adding more powerful equipment without waiting for a major utility expansion.
The Fastest Infrastructure Upgrade May Be a Retrofit
The conventional response to rising AI demand is to build more.
That might mean new power plants, larger grid connections, additional data centers, or on-site generators. Each approach carries significant costs, delays, and environmental consequences.
TenHouten argues that improving existing facilities should come first.
Intrinsic Power has designed its systems to operate alongside current electrical infrastructure. Early deployments can be installed in parallel, allowing operators to test the technology without immediately replacing their entire power architecture.
This retrofit model addresses one of the most difficult challenges in infrastructure innovation. Data center operators cannot simply shut down working facilities and rebuild them around an unproven system.
New technology must demonstrate value while the existing operation continues.
More Efficient Power Use Reduces Other Constraints
Electricity shortages create problems beyond computing capacity.
Data centers require substantial cooling, and cooling itself consumes power. When electricity is limited, operators may rely on practices that use more water because recirculating and rechilling that water would require additional energy.
Some facilities are also installing local natural gas or diesel generation to bypass delays in grid expansion. Those systems may provide power quickly, but they can be more expensive and less efficient than centralized generation.
Better load management can reduce pressure across the entire operation.
A facility that uses its grid connection more effectively may require less on-site generation, support more advanced cooling systems, and integrate battery storage or renewable energy more successfully.
The value is not merely a lower electricity bill. It is the ability to produce more computing output from the infrastructure already in place.
AI Infrastructure Needs Intelligence of Its Own
The development of artificial intelligence is often framed as a software challenge.
TenHouten’s work reveals how dependent that progress is on physical systems. Faster chips and larger models cannot create value if the facilities supporting them cannot access reliable power.
The next phase of AI infrastructure may therefore depend on making electricity distribution as intelligent as the computing workloads it supports.
Instead of waiting for the grid to double in size, data centers can become more responsive participants in the energy system. Batteries, sensors, and predictive controls can absorb volatility locally rather than sending sudden demand back through utility networks.
AI’s power challenge will not disappear through one technology alone. New generation and grid construction will remain necessary.
But during the years those projects take to complete, better use of existing capacity could determine which companies are able to keep scaling and which are forced to wait.
Want more Plugged In? Read more or listen on Apple Podcasts.
