This white paper looks at the metalens opportunity in AI datacenters. One of the most consequential open questions in AI datacenter design right now is how to keep GPUs supplied with data fast enough to make effective use of the computing capacity. GPU utilization varies widely with workload and system design, and GPUs can spend substantial time waiting for data. Network and interconnect bottlenecks can contribute to utilization issues, alongside storage, data-pipeline and workload-scheduling limitations. Low utilization means the GPUs are idle, which implies lower efficiency and higher customer costs. This paper addresses the interconnect part of that problem.






