Microsoft pursues composable memory for AI workloads

Memory has become the scarcest resource in AI data centers, with GPUs sitting idle waiting on it. Microsoft's principal architect Samir Rajadnya says the company is exploring composable memory for Azure, a model that pools memory beyond individual servers and assigns it dynamically to workloads. Rather than locking memory to a single CPU or GPU, the approach treats memory as a shared resource. Rajadnya advises enterprises to identify their top workloads and build vertically integrated systems around them, while noting that optical interconnects will be critical to making composable memory work at scale.