The race to power artificial intelligence is moving beyond Earth. Starcloud Inc., an artificial intelligence hardware startup, announced a $250 million funding round to construct data centers in orbit, where solar panels generate power continuously and expensive cooling systems become unnecessary. Meanwhile, universities like Duke are designing energy-efficient GPU facilities on the ground, signaling a broader shift in how institutions are approaching the massive power demands of AI computing.

Why Are Companies Looking to Space for AI Computing Power?

Starcloud's orbital approach addresses one of the most pressing challenges facing AI infrastructure: energy consumption. On Earth, data centers require elaborate cooling systems to prevent graphics processing units (GPUs), the powerful semiconductors that run AI workloads, from overheating. In space, satellites can be positioned to face the sun continuously, generating power 24/7 without the weather interruptions that plague terrestrial solar arrays. The company has already launched an Nvidia graphics card into orbit aboard a small satellite and completed the first-ever orbital AI training run.

The funding round, led by investment firm Manhattan West and including backers like Nvidia and Cisco Investments, reflects growing confidence in the orbital model. Starcloud plans to eventually launch 88,000 satellites with a combined computing capacity of 20 gigawatts, though the company's immediate focus is on expanding manufacturing capacity and booking launch space with SpaceX.

Starcloud's roadmap includes four generations of satellites. The Starcloud-2 system, scheduled to launch in 2027, will handle both AI training and inference workloads, processing sensory data in orbit rather than beaming raw datasets to Earth. The Starcloud-3 will consume 200 kilowatts of power, equivalent to the energy use of a few dozen homes, while the most advanced Starcloud-4 envisions a cylindrical structure attached to a 2.5-square-mile solar array with liquid-cooled server modules.

How Are Universities Balancing AI Computing Needs With Sustainability?

While private companies pursue orbital solutions, academic institutions are taking a different approach by designing highly efficient ground-based facilities. Duke University is constructing a GPU Center expected to open in 2027 that will serve as a model for sustainable computing infrastructure. Unlike hyperscale data centers that consume 100 megawatts or more of electricity and span hundreds of acres, Duke's facility will occupy less than one-third of an acre and use just 1.5 to 3 megawatts at peak load, representing only 2 to 3 percent of the university's total annual energy consumption.

The facility incorporates several innovations to minimize environmental impact. It will connect to Duke's closed-loop chilled water system, which recovers heat from GPU cooling and repurposes it to heat campus buildings, turning what is typically waste into a resource. The design also minimizes water consumption and noise from cooling equipment, common concerns about hyperscale data centers.

"Demand for computing power is accelerating rapidly in both applied and basic research. Providing the computing capacity our researchers need is a must-have for recruiting and retaining faculty talent," said Tracy Futhey, vice president and chief information officer at Duke University.

Tracy Futhey, Vice President and Chief Information Officer, Duke University

What Is Driving the Urgent Need for More AI Computing Power?

The explosion in AI demand is reshaping electricity grids worldwide. According to research from Duke's Nicholas Institute for Energy, Environment and Sustainability, large data center power demands are expected to rise approximately 130 percent by the end of this decade, with about one-half of that growth driven by AI. This represents a fundamental shift from the relatively flat electricity growth patterns of previous decades.

Duke researchers have identified a potential solution: load flexibility. By allowing data centers to shift their computing workloads to times when electricity is abundant and cheaper, the power system could absorb new demand more quickly while reducing the need for costly grid expansion. This approach could help maintain reliability, control costs, and reduce emissions across the broader energy infrastructure.

How Are Researchers Making AI Hardware More Efficient?

Beyond facility design, researchers are reimagining the fundamental building blocks of AI computing. Duke's Electrical and Computer Engineering Professor Tania Roy is developing neuromorphic, or brain-inspired, semiconductor devices that can perform AI tasks without relying on energy-intensive data centers. Her laboratory creates tiny experimental circuits that mimic the efficiency of the human brain, enabling future smart devices to process information locally rather than sending every request to a server farm.

"The transistor is the basic device that has revolutionized everything. Our computers, mobile phones, everything has them. We want to change that basic building block for AI, so that it can do AI more efficiently," explained Tania Roy, Electrical and Computer Engineering Professor at Duke University.

Tania Roy, Electrical and Computer Engineering Professor, Duke University

Other Duke researchers are pursuing complementary approaches. Helen Li and Yiran Chen are developing magnetic random access memory technology to reduce water and energy consumption in AI systems. Both Roy and Chen recently received Department of Energy funding to develop robotic AI processing hardware that is significantly faster and more efficient on local devices, reducing the amount of energy-intensive data transfers to and from centralized data centers.

Steps to Understanding AI's Power Challenge and Solutions

• Recognize the Scale: AI data centers now account for roughly one-half of all new electricity demand growth, with power consumption expected to rise 130 percent by 2030, fundamentally changing how utilities plan grid infrastructure.
• Understand Orbital Advantages: Space-based data centers eliminate cooling costs, generate continuous solar power without weather interruptions, and can process data in orbit before transmitting results to Earth, reducing bandwidth demands.
• Evaluate Efficiency Innovations: Brain-inspired chips, magnetic memory technologies, and load-flexible computing systems represent three distinct approaches to reducing AI's energy footprint without sacrificing performance.
• Consider Facility Design: Modern AI computing facilities can minimize environmental impact through heat recovery systems, closed-loop water recycling, and integration with existing campus infrastructure.

The convergence of orbital and terrestrial approaches suggests that AI's power challenge will not be solved by a single technology. Instead, a portfolio of solutions is emerging: space-based infrastructure for massive-scale computing, efficient ground facilities for research institutions, and fundamentally redesigned chips that reduce energy consumption at the hardware level. As AI continues to reshape industries from healthcare to climate science, the infrastructure supporting these systems is becoming as important as the algorithms themselves.