What we know

AI clusters with more than 100,000 accelerators can no longer fit within a single data center because of power and cooling limitations. Instead, these massive systems must be distributed across multiple facilities. Companies like Google, Cisco, and Marvell are developing architectures designed to scale AI clusters across several data centers.

Why it matters

As AI workloads grow rapidly, the demand for larger accelerator clusters increases. Traditional data centers face significant bottlenecks in power delivery and cooling capacity, limiting the size of single-site AI systems. By spreading clusters across multiple data centers, the industry can continue scaling AI performance.

What is still unknown

Details on the accuracy of these distributed AI systems, the range of supported languages, and timelines for broader deployment remain unclear. No independent evaluations of performance or energy efficiency have been reported.