Researchers at Binghamton University have developed a new power conversion technology that could significantly improve energy efficiency in AI data centers, potentially reducing wasted energy by 10-12% across all operating conditions. The innovation addresses a fundamental challenge facing the industry as artificial intelligence demands continue to surge and data centers consume ever more electricity.

Why Are Data Centers Struggling With Power Efficiency?

The challenge stems from a physics problem that chip designers have been wrestling with for years. As transistors have become smaller, they've hit a wall. Moore's Law, the decades-old prediction that computing power would double every two years, has essentially reached its limit. Modern chips can't get much smaller than 4 to 7 nanometers without running into physical barriers.

To keep improving performance without shrinking transistors further, chip makers are pushing toward higher computing density and lower voltages. Traditionally, chips operated at 5 volts or 3.3 volts, but the industry is now moving toward voltages below 1 volt. Here's the catch: when voltage drops, the laws of physics require more electrical current to deliver the same computing power. That extra current creates heat, and managing that heat in densely packed data centers has become a major problem.

"Because we cannot make the transistor any smaller, we need to reduce the voltage that these chips handle or operate from. Traditionally, it has been 5 volts or 3.3 volts, and now chipmakers are looking below 1 volt. As the voltage is reduced, it allows electronics packaging in a more dense way," said Pritam Das.

Pritam Das, Associate Professor of Electrical and Computer Engineering at Binghamton University

What's the Bottleneck in Today's Data Centers?

The real problem lies in devices called point-of-load converters, or POLcs. These are small power converters that sit right next to graphics processing units (GPUs) and other AI chips, stepping down voltage from the main power distribution system to the ultra-low voltages the chips need. Traditional POLcs use multiple stages of conversion, and this multi-stage approach is inefficient. They typically achieve only about 80% efficiency, meaning that for every 100 watts of power delivered to a GPU, roughly 20 watts is lost as heat.

In a data center packed with thousands of GPUs all generating heat simultaneously, adding extra heat from inefficient power conversion makes cooling exponentially more difficult and expensive. The industry is planning to upgrade main power distribution from 12 volts to 48 volts, which would reduce transmission losses by 16 times. But that upgrade only works if the point-of-load converters can efficiently step that 48-volt power down to 1 volt or lower.

How Does the New Technology Work?

Das and his team, including PhD student Tuhin Sasmal, developed a single-stage point-of-load converter that eliminates the inefficiency of multiple conversion stages. A laboratory prototype achieved 10-12% higher efficiency across all load conditions and doubled the slew rate, which measures how quickly the converter can deliver power when demand spikes.

The speed improvement matters because AI workloads are unpredictable. When a GPU suddenly needs more power for a computational burst, the converter needs to deliver that current almost instantly. Das compared it to how the human brain needs fuel quickly when working hard. The new converter delivers power faster, reducing the lag between demand and supply.

The team has already secured one patent for single-stage conversion from 48 volts down to 1 volt, with another patent pending that would allow manufacturers to package the converter just 5 millimeters away from the microchip, roughly the thickness of a chocolate bar.

Steps to Commercialize Energy-Efficient Data Center Technology

• Secure Funding: Das received $100,000 from Binghamton University's Excellence in Entrepreneurship and Discovery program, supported by a National Science Foundation Accelerating Research Translation grant, to fund prototyping and data collection.
• Build and Test Prototypes: The team is using grant funds to construct working prototypes and gather real-world performance data to validate laboratory results before approaching industry partners.
• Attract Commercial Partners: The goal is to encourage startup companies and established manufacturers to license the technology and build commercial versions for hyperscalers like Amazon Web Services, Microsoft Azure, and Google Cloud.

What Does This Mean for the AI Industry?

The timing of this innovation is critical. Data centers are multiplying rapidly to support AI workloads, and energy efficiency has become a competitive advantage. A 10-12% improvement in power conversion efficiency might sound modest, but across thousands of GPUs running continuously, it translates to significant reductions in both energy consumption and cooling costs. For hyperscalers operating on razor-thin margins, that efficiency gain directly impacts profitability.

The innovation also addresses a broader global challenge. South Korea, a major semiconductor manufacturing hub, is facing pressure to update its national power planning. Energy experts warned that planned AI data centers and semiconductor fabrication plants could require 216.4 terawatt-hours of electricity annually, exceeding the 178 terawatt-hour increase forecast in the country's long-term power supply plan. Technologies like Das's converter could help reduce that demand pressure.

Meanwhile, in the United States, New Jersey has taken a different approach to the data center energy challenge. Governor Mikie Sherrill signed legislation requiring large data centers to pay for their own grid infrastructure and energy use rather than shifting those costs to residential ratepayers. The "Data Center Fair Share" law also incentivizes data centers to bring in their own clean power and requires them to reduce consumption during grid strain events before residential customers are affected.

These developments reflect a broader recognition that AI infrastructure's energy demands are no longer a niche concern. They're reshaping how governments plan power grids, how utilities structure rates, and how engineers design the fundamental building blocks of data center hardware. Das's work on power conversion efficiency represents one piece of a much larger puzzle: making AI infrastructure sustainable and affordable as the technology continues to scale globally.