The field of AI research is moving towards developing scalable and cost-effective infrastructure to support increasingly large AI workloads. Researchers are exploring novel network architectures and load balancing algorithms to optimize performance and reduce costs. A key direction is the design of flexible and reconfigurable networks that can efficiently interconnect thousands of chips, enabling hyper-scale AI training systems. Another important area of focus is the development of strategies to effectively utilize available computing resources, including the optimization of job submission policies to maximize revenue. Noteworthy papers include:

• RailX, which proposes a scalable and low-cost network architecture for hyper-scale LLM training systems, achieving better scalability and cost-effectiveness than existing architectures.
• Age of Estimates, which investigates the optimal times to submit jobs to a Markov machine to maximize revenue, developing a threshold policy and a switching policy based on the age of the estimate of the state of the machine.