Abstract: Hyperscale datacenters provide critical infrastructure for the information economy, refining data to extract value. Sharing datacenters improves energy efficiency but whether strategic users participate depends on management policies. I address these challenges by integrating practical insights from computer architecture with rigorous methods in algorithmic economics and machine learning. I illustrate this approach for power allocation and performance diagnosis. First, I allocate server power by anticipating strategic behavior and incentivizing participation. I design mechanisms in which users selfishly draw power to boost performance yet avoid oversubscribing the shared supply. Second, I diagnose performance anomalies by inferring interpretable causal relationships. I design statistical frameworks that learn performance models and extract recurring semantic structure from these models with natural language processing. Finally, I conclude by describing future directions for efficient and secure architectures in datacenters and beyond.