Delivery of a transformer
The announcement from DeepSeek that its carbon footprint is a fraction of rivals such as OpenAI is shifting the competitive landscape of artificial intelligence. But there is no question that the growth of power-hungry AI applications is putting enormous demand on data center developers and operators who are keen to have their facilities up and running as quickly as possible.
For them, the fast, reliable supply of long-lead components such as transformers is critical to meeting demand, said Jake Ring, Power and Data Management managing partner and president.
“It can take years to deliver a data center, from achieving funding to getting the distribution grid to provide power to the location,” he said. “The utility might need to build their substation and will need transformers and breakers to distribute power as quickly as possible.”
Data centers’ demand for power is expected to increase 19-22% annually between 2023 and 2030 from about 60 gigawatts a year globally to 171-219 gigawatts. Estimates suggest about 70% of demand by 2030 will be for data centers equipped to host advanced AI workloads.
Developers and operators need to not only create facilities with enough power but also deliver a much higher level of power to support each IT rack. The more sophisticated the AI chip, the more power it needs.
“Not long ago, a typical data center would supply racks with power of 7-10 kilowatts,” Ring said. “With the advent of cloud services, these racks needed up to 20-25 kilowatts. The next iteration of Nvidia’s chip to power AI, Blackwell, will require up to 160k kW per rack — far more than what many older data centers have been constructed to provide. The next iteration, Reuben, could require 260 kW per rack on startup.”
Data storage is an additional challenge for data centers running AI applications, Ring said, as they use around 10-12 kW per rack. Balancing the load across multiple rows of racks requires the design be tailored to the expected application during fit-out.
Transformers are critical to provide the right level of power for each piece of data center equipment, especially when an operator can’t afford power to be disrupted. Operators need flexibility in terms of the power provided to each element — two megawatts delivered to one hall and 10 megawatts delivered to another since not only the server load but the heat rejection systems need critical power support, Ring said.
“Our goal is to provide reliable, customized power distribution and unparalleled speed of delivery,” he said. “No one wants their AI application to stop halfway through training due to harmonic interference or overheating. So we build our products to support continuous duty for extended overloads.”
Given the global political situation, including potential tariffs on products from overseas, a domestic supply of transformers is extremely important, Ring said. PDM manufactures transformers in the U.S. using domestic and imported components. The lead time for the company’s products is as low as nine weeks, far quicker than the industry high of up to 115 weeks and six months for high voltage breakers, where some providers are quoting four years, Ring said.
“The 2024 opening of our Salt Lake City facility was extremely well timed,” he said. “We have brought down the lead time of transformers massively, which makes a huge difference for operators and developers that are active in the race to build data centers for AI applications.”
PDM is working closely with developers and operators to innovate in the design of data centers. For example, it seeks to streamline the critical path for power through a series of transformers that incorporate a vacuum-fault interrupting circuit breaker and control relay to provide the right overcurrent protection for each piece of hardware.
“Each time you transform power, you lose a bit of efficiency,” Ring said. “We can support whatever the requirement might be to distribute the power through breakers and transformers in the most effective way, delivering the combination of voltage and current required in as few steps as possible. As a result, the infrastructure overall will require less copper, less space, less equipment expense and so on.”
Other innovations include replacing the use of mineral oils with seed oil, such as FR3 from Cargill. As this is less flammable, the transformer can run at a higher heat rating and provide as much as 25% more power, Ring said. PDM is also seeing a greater use of aluminum to replace increasingly expensive copper.
PDM recently has been delivering more transformers that need greater protection against harmonics because AI applications are driving more use of variable frequency drives for pumps and fan walls. Harmonics are also being generated by GPU processors.
“When an AI chip is in training mode, it can peak at 150% above its rated power utilization, drop down to 100%, then drop to 10% before starting the cycle over again, which creates harmonic patterns,” Ring said. “To prevent this from impacting the power supply or traveling outside the data center, we’re providing more dry-type transformers with K-factor ratings of 9 and 13. AI is forcing people to take a more holistic look at the facility.”
While the impact of using so much power to operate AI data centers is yet to be fully understood, the capabilities that come with AI applications are worth it, Ring said. As technology evolves and more benefits are realized, there’s a direct tie between the amount of power data centers are using and the value they create.
Ring cited the example of how someone with diabetes can now send blood sugar levels instantly to their endocrinologist using a smartphone app for immediate AI-driven analysis, which can help better regulate or even reduce their insulin use. Or how a smart car interacting with a smart traffic system can work out the best lane and speed to pass through lights efficiently as well as better clear the way for emergency responders.
“These gains can be life-changing and can help save the use of power elsewhere since greater electrification in transportation, healthcare and other sectors is also driving the need for more power generation,” he said. “Faster delivery of transformers is needed to support this growth.”
This article was produced in collaboration between Power and Data Management and Studio B. Bisnow news staff was not involved in the production of this content.
Studio B is Bisnow’s in-house content and design studio. To learn more about how Studio B can help your team, reach out to studio@bisnow.com.
