Global AI data centers are engaged in an intense competition for power and computing resources, giving rise to a series of creative "acceleration strategies"—ranging from constructing AI "tent cities" and recommissioning jet engines to repurposing abandoned industrial sites. The most surprising aspect is the diverse array of improvised and adaptive measures major AI data centers are employing to meet the enormous electricity demands of artificial intelligence.
Here are some highly ingenious strategies companies are using to rapidly scale up their computing capacity.
Strategy One: Leveraging Existing Real Estate to Maximize Power Usage In early 2023, as the AI data center boom began, Alphabet and Meta Platforms, Inc. held valuable power usage rights in New Albany, Ohio (near Columbus). As previously discussed, New Albany, with its high-voltage transmission network—a major power "artery" with substantial capacity—has become a new hub for the AI industry. The main campuses of Alphabet and Meta Platforms, Inc. are situated just across a boulevard from each other.
To fully utilize the local power supply, Meta Platforms, Inc. expanded its AI data center capacity. It deployed a dedicated team of engineers to five of its traditional data centers, replacing rows of CPU chip racks with AI-specific GPU chips. This created a massive AI computing cluster within just a few months, as stated in a Meta Platforms, Inc. blog post last September.
Meta Platforms, Inc. also erected an AI "tent city" on its own land in New Albany, using temporary structures commonly seen at sporting events and exhibitions to shelter valuable computing and cooling equipment. "A year ago, if someone said, 'We're going to put a bunch of expensive GPUs in a tent,' it would have been unimaginable," said former Meta Platforms, Inc. executive Chris Malone last fall during a presentation. By that time, he had moved to OpenAI to oversee its data center construction.
Alphabet, meanwhile, interconnected multiple data centers in Ohio, Nebraska, and Iowa via massive fiber optic cables to extract more computational power from its chips. This effectively consolidated hundreds of megawatts (potentially totaling several gigawatts) of precious electricity into a "super-brain," ensuring that training processes can continue running even if individual chips fail.
Strategy Two: Securing Others' Power Allocations for Rapid Deployment As early as 2022, Microsoft was acutely aware of AI's "insatiable" appetite for power—it was then providing the computing power for training OpenAI's ChatGPT. Consequently, when Microsoft discovered that one of the nation's largest electrical substations was newly built but sitting idle, it moved quickly. The substation had been constructed by a utility company for Taiwanese manufacturer Foxconn in Mount Pleasant, Wisconsin, but Foxconn later abandoned its plans to build a large LCD factory there.
"The secret to getting power fast right now is to find resources that are already interconnected to the grid but where the project hasn't fully materialized," said Landon Lill, an energy regulatory law partner at Baker Botts, the law firm coordinating power supply for AI data centers.
According to SemiAnalysis, Microsoft's newly built Fairwater data center currently supports over 350 megawatts of AI workload, with potential for expansion to 2 gigawatts and beyond (1-2 GW is roughly equivalent to the power demand of a medium-sized city). Microsoft is also building an identical data center near Atlanta and connecting the two directly with new high-speed data links.
Strategy Three: Repurposing "Stranded" Green Energy Projects from the Trump Era Data centers have been utilizing power resources developed for cryptocurrency miners for several years, but another category of power assets had remained largely untapped. Green hydrogen was once touted as a clean fuel for transportation and manufacturing, attracting significant attention. However, its adoption has been hampered because its production consumes vast amounts of renewable energy, costs are high, and transportation is extremely difficult. The Trump administration further reduced tax credits for green hydrogen projects.
Now, data center developers are targeting solar and wind farms that were built or permitted specifically for green hydrogen projects. This strategy was previously mentioned in the context of Alphabet's acquisition of renewable energy company Intersect. These sites are not currently running AI workloads but are expected to be operational soon.
The pivot away from green hydrogen also facilitated a new agreement between Alphabet and power producer AES (which investors like BlackRock are acquiring). AES will build renewable energy facilities for Alphabet's supercomputer in Willacy County, Texas. The site was originally planned for green hydrogen production in collaboration with industrial gas supplier Air Products.
After Air Products exited the project, AES had the land prepared for power generation and the grid interconnection ready. This enabled it to take advantage of a new Texas regulation that grants grid interconnection priority to data centers that bring some of their own power supply.
Strategy Four: Recommissioning Reliable Machinery from a Bygone Era Elon Musk's xAI (now merged into SpaceX) is using refurbished General Electric aircraft engines—which once powered Boeing 747s and 767s—to generate electricity for its second Colossus data center project outside Memphis, Mississippi. Despite air pollution concerns, the project received a permanent operating permit this month.
Proenergy, a company specializing in refurbishing decades-old aeroderivative engines for land-based use by utilities or industrial sites, stated that about 1,000 such engines are scheduled for retirement over the next decade. With some GE patents expired, Proenergy can manufacture its own replacement parts.
Refurbishing old engines is faster than the high-profile approach of AI data center developer Crusoe, which purchases new jet engines. Crusoe became the launch customer for supersonic jet manufacturer Boom Supersonic—which pivoted to the AI sector months ago needing funds for engine development. Boom will modify engines for land-based use at its new Superpower superfactory in Denver. Crusoe has ordered 29 turbines with a total capacity of 1.21 gigawatts. The concept is that if these engines can be ready within 1-2 years, they could generate more power in hot weather than older turbines designed for cold, high-altitude conditions.
Strategy Five: Building Independent Power Generation for Self-Sufficiency Applied Digital, which is building large data centers in North Dakota for companies like CoreWeave, recently established a new independent power producer called Base Electron. Days ago, Base Electron placed an order with veteran coal-fired engineering firm Babcock & Wilcox to build a natural gas power plant using boiler and steam turbine technology derived from the coal era. Applied Digital and Babcock & Wilcox claim this approach can be deployed faster than the currently scarce simple-cycle and combined-cycle gas turbines.
Industry News
Emerald AI announced demonstration results from its collaboration with Nvidia today: its load-shifting software helps data centers flexibly adjust power consumption in real-time. During peak grid stress periods, it can ensure heavy AI workloads continue running while reducing energy consumption by over 20%. Testing simulated scenarios included sudden nighttime temperature drops and prolonged heatwaves.
KKR is considering a multi-billion dollar sale of data center cooling company Cool IT Systems.
The U.S. Nuclear Regulatory Commission issued the first construction permit for a next-generation nuclear reactor in Wyoming to TerraPower, a startup backed by Bill Gates.
BlackRock has established a $100 million philanthropic fund dedicated to workforce development for technical trades needed in AI and other infrastructure sectors.
Comments