Meta has deployed a custom bridge chip across millions of new servers to reuse older RAM modules. This hardware innovation reduces the number of machines needed for specific inference workloads by 25%.