Orbital data centers powered by abundant solar energy could help solve AI's growing power crisis, but only if engineers redesign how chips handle heat in the vacuum of space. A new research paper from arXiv reveals that standard GPU architectures struggle under the thermal constraints of space, while emerging compute-in-memory accelerators could deliver 10 to 40 times better performance in these extreme conditions.

Why Is Space Suddenly Attractive for AI Computing?

The explosive growth of artificial intelligence has created an unprecedented demand for computing power. Data center workloads are projected to account for 30 to 40 percent of all new electricity demand in the United States by 2030, according to McKinsey and Company. This energy crisis has forced researchers and companies to explore unconventional solutions, including deploying AI infrastructure beyond Earth.

Space offers two compelling advantages: virtually unlimited solar energy and infinite physical space for large-scale deployments. The cold background temperature in the vacuum of space initially seems ideal for cooling hot computing chips. However, this apparent advantage masks a fundamental physics problem that terrestrial data centers never face.

What Makes Cooling in Space So Difficult?

On Earth, data centers rely on convection, where fans and liquid cooling systems rapidly move heat away from chips. In the vacuum of space, convection is impossible. Instead, chips must shed heat through thermal radiation, a much slower process that requires enormous radiators. This constraint severely limits how much power any space-based computing system can draw.

The research team developed a "radiator-in-the-loop" design methodology that directly links the permitted computing throughput with the practical cooling capacity available in space. Their thermal simulations revealed a critical problem: traditional GPU architectures with separate graphics processing units and high-bandwidth memory (HBM) create severe thermal hotspots under limited radiator capacity, forcing the system to throttle performance and reduce its effective computing power.

How Could Compute-In-Memory Accelerators Change the Game?

Compute-in-memory (CIM) accelerators represent an emerging alternative to traditional GPU designs. Instead of separating the computing logic from memory, CIM architectures integrate memory directly into the processor. This fundamental redesign addresses the "von Neumann memory wall," a well-known bottleneck where most energy consumption comes from moving data between memory and compute units rather than from computation itself.

The research demonstrates that CIM accelerators exhibit much more uniform heat distribution across the chip compared to GPUs. This uniform thermal profile eliminates the hotspot problem entirely, allowing the system to maintain higher performance under the same radiator constraints. Across various artificial intelligence workloads, CIM consistently outperformed traditional GPUs in terms of computing operations per watt (TOPS/W) across a wide range of radiator budgets.

Steps to Evaluate AI Hardware for Space Deployment

• Thermal Modeling: Use finite-element-methods (FEM) based system modeling to assess how limited cooling impacts AI hardware performance under realistic space conditions.
• Radiator Budgeting: Establish a fixed radiator thermal rejection power budget and evaluate how different chip architectures perform within that constraint.
• Workload Testing: Benchmark candidate hardware across diverse representative AI compute workloads to identify which architectures maintain performance under thermal throttling.
• Heat Distribution Analysis: Map spatial thermal profiles to identify hotspots that could trigger performance degradation or system failures.

The research team's evaluation framework, based on roofline-like modeling, explicitly links radiator availability to the impact of thermal throttling on peak performance. This methodology could become essential as space-based computing transitions from theoretical concept to practical deployment.

What Do These Findings Mean for AI Infrastructure?

The magnitude of CIM's advantage is striking. Under tight thermal constraints, CIM accelerators delivered over 10 to 40 times improvement in TOPS compared to GPU-plus-HBM systems, depending on the specific workload and radiator budget. This performance gap opens what researchers describe as "an exciting venue for developing AI hardware for space computing."

Meanwhile, on Earth, the power infrastructure challenge continues to drive innovation in terrestrial data center efficiency. Navitas Semiconductor, a power-chip specialist, recently showcased collaboration with NVIDIA at Computex 2026, highlighting an 800-volt-to-6-volt power delivery board using gallium nitride (GaN) technology that achieves 97.5 percent peak efficiency. The company estimates the AI data center market alone could represent a serviceable addressable market opportunity of 1.4 billion to 2.5 billion dollars by 2030, with GaN and silicon carbide (SiC) adoption expected to grow at a compound annual rate of 66 to 87 percent between 2025 and 2030.

The convergence of these developments, from space-based orbital data centers to advanced power delivery systems on Earth, reflects the industry's urgent need to solve AI's power consumption challenge. Whether the solution comes from the vacuum of space or from more efficient terrestrial infrastructure, the next generation of AI computing will look fundamentally different from today's GPU-dominated data centers.