AI’s voracious appetite is straining energy grids, prompting new power strategies and longer planning horizons
In sum — what to know:
The gating factor: Power is now the biggest gating factor for AI infrastructure, with water to follow next.
Bring your own power: To avoid grid struggles, businesses must consider adopting the bring your own power (BYOP) strategy, which involves owning independent micro-grids and tapping into multiple energy sources.
Nuclear power, a distant solution: Nuclear power as an alternative energy source for AI remains years away.
AI infrastructure is at a crossroads. While on one hand demand for infrastructure is growing at a manic speed, power constraint is choking structural expansion plans.
That was the message from Marc Ganzi, CEO of DigitalBridge — an asset management firm in the digital infrastructure space soon to be acquired by SoftBank — at Metro Connect USA in Fort Lauderdale.
Speaking at the keynote, Ganzi said, “To enable AI, we all know the cliche is, power availability becomes the currency. That’s really the currency that’s driving critical workloads, that’s driving new data centers, and ultimately drives an ecosystem that feeds off of those data centers.”
The power crunch is here to stay
While there is abundant free land, land without power is of no value to data centers, nor is a “will-serve” letter from an utility provider promising capacity and intent to supply.
“A will-cert letter does not mean you have a connection date,” he reminded, noting that developers today are looking at connection dates starting 2030 to 2032.
He said that gaining access to energy is now a power play which really comes down to how politically connected one is in that state, arguing that companies in digital infrastructure business are essentially in power enablement business.
In order to sidestep state-level politics and long grid queues, developers must own grid-independent power, micro-grids, and multi-source energy strategies as a core part of their business model. Also known as bring your own power (BYOP), the strategy promotes onsite power generation over reliance on public grids. Large-scale data centers are increasingly adopting BYOP solutions like temporary generators and micro-grids to get around the capacity crunch and obtain supply in a timeline manner.
This is all the more essential as the power problem is here to stay, Ganzi said. The industry currently leases 12 gigawatts of power every year, while adding four gigawatts of incremental capacity.
“You guys can do the math. We have a deficit of power,” he said, adding that the gap will only widen in the coming years.
And as the energy discourse is turning towards nuclear power as an alternative source for meeting AI’s appetite for baseload power, Ganzi’s message was blunt and clear: “We don’t believe that magically 40 to 60 gigawatts of nuclear power is going to show up in eight years. It’s not.” He predicted that it will take well over a decade for nuclear to become adequately available.
“If we’re all hoping for some magic bullet to save the power problem, I’m here to tell you it’s not going to happen,” he repeated.
More water for cooling
With the data center industry already facing public backlash from local communities for building projects in their backyards, Ganzi predicted that the next challenge will be to face the water controversy.
Large data centers consume up to 5 million gallons of water daily. That’s equivalent to water consumed by a big town of 10,000 to 50,000 people. That amounts to an alarming amount of water wasted yearly. As an example, during a legal battle, it was found that Google data centers in the Oregon region consumed over 355 million gal. in 2021.
With many watersheds already moderately to highly stressed, companies are now looking to tapp reclaimed or recycled water for data center cooling.
“We’ll have to defend our water usage,” Ganzi said. He noted that it is pressing on the industry to design sustainable water cooling systems that call for zero waste.
Already work is underway. Many companies are designing cooling systems that can check water consumption through prevention of evaporation. In 2024, Microsoft launched a green data center design that it says “consumes zero water for cooling”. The design leverages chi-level cooling to achieve the right temperature without water evaporation — an effort to preserve local watersheds.
As a potential solution to the ongoing energy struggle, Ganzi urged the industry to think in longer timelines and planning horizons. He recommended a 36 to 48 month build cycle as opposed to a shorter five-year model, and advised companies to start planning for tomorrow’s supply issues today, with several years of time on hand.
“That’s the mentality you have to have for success,” he said. “The opportunity is huge. The challenges are different. Every year the chessboard changes, and it’s a lot of fun.”
