Cloud and cluster computing platforms have become standard across almost every domain of business, and their scale quickly approaches $mathbf{O}(10^6)$ servers in a single warehouse. However, the tier-based opto-electronically packet switched network infrastructure that is standard across these systems gives way to several scalability bottlenecks including resource fragmentation and high energy requirements. Experimental results show that optical circuit switched networks pose a promising alternative that could avoid these.

However, optimality challenges are encountered at realistic commercial scales. Where exhaustive optimisation techniques are not applicable for problems at the scale of Cloud-scale computer networks, and expert-designed heuristics are performance-limited and typically biased in their design, artificial intelligence can discover more scalable and better performing optimisation strategies.

This thesis demonstrates these benefits through experimental and theoretical work spanning all of component, system and commercial optimisation problems which stand in the way of practical Cloud-scale computer network systems. Firstly, optical components are optimised to gate in $approx 500 ps$ and are demonstrated in a proof-of-concept switching architecture for optical data centres with better wavelength and component scalability than previous demonstrations. Secondly, network-aware resource allocation schemes for optically composable data centres are learnt end-to-end with deep reinforcement learning and graph neural networks, where $3times$ less networking resources are required to achieve the same resource efficiency compared to conventional methods. Finally, a deep reinforcement learning based method for optimising PID-control parameters is presented which generates tailored parameters for unseen devices in $mathbf{O}(10^{-3}) s$. This method is demonstrated on a market leading optical switching product based on piezoelectric actuation, where switching speed is improved $>20%$ with no compromise to optical loss and the manufacturing yield of actuators is improved. This method was licensed to and integrated within the manufacturing pipeline of this company. As such, crucial public and private infrastructure utilising these products will benefit from this work.