Google's next-generation TPU cluster has officially adopted NPO (Near-Package Optics) technology as a standard solution.

In the long-term competition of the AI computing power track, the ultimate battle is never about who can pile up more computing power, but about who can solve the underlying power consumption and bandwidth bottlenecks. The core advantage of Google's implemented NPO solution is that it reduces power consumption by 50% compared to traditional solutions. For large-scale model clusters with tens of thousands of cards, power consumption costs and cooling pressure have long become the core bottlenecks restricting the expansion of computing power scale.

This is also why all leading global technology companies are investing heavily in the iteration of underlying optical technology — this is the real long-term moat for the AI computing power track, not short-term capacity hype.