From the paper:

"Our CXL solution achieves substantial gains for diverse workloads, including up to a 25% reduction in server count for disaggregated ML inference"

How does using worse RAM result in 25% reduction of server count for given workloads?

CXL is adding "slow" RAM over PCIe, basically. Not replacing.

Because it's used in addition to, not in place of, the better RAM.