Power and cooling limits push 100,000+ accelerator AI clusters past single data centers — Google, Cisco, and Marvell target ‘scale-across’ networking across regions

Tom's Hardware
Read full story· 1 min
AI cluster growth is exceeding the power and cooling capacities of single data centers, prompting companies like Google, Cisco, and Marvell to develop networking solutions that connect compute resources across multiple regions. Researchers at Supercomputing 2026 emphasized that scaling distributed computing efficiently requires more than just faster networks, highlighting the need for architectures that span geographical boundaries.

What happened

The expansion of AI clusters is outpacing the power and cooling capabilities of individual data centers. As a result, companies including Google, Cisco, and Marvell are working on networking technologies to connect computing resources across multiple data centers and regions. At the Supercomputing 2026 conference, researchers discussed the need for new architectures to support distributed computing at this scale. They emphasized that simply increasing network speed is not enough for efficient operation across geographically separated facilities.

Key details

  • AI clusters now exceed the power and cooling limits of single data centers.
  • Google, Cisco, and Marvell are developing networking solutions for multi-region AI clusters.
  • Supercomputing 2026 featured research from The Ohio State University on this topic.
  • Efficient scaling requires new architectures, not just faster networks.
  • Workloads must be distributed across data centers in different regions.

Why it matters

Operators of large AI clusters, such as cloud providers and enterprises, must now consider distributing workloads across multiple data centers, rather than relying on a single facility. This shift requires investment in new networking technologies and architectural approaches to maintain performance and efficiency. Companies building or expanding AI infrastructure face new technical requirements and costs as they adapt to these physical and operational limits.


More in Chips & Compute