Microsoft Urges Yield-Led Co-Design to Ease AI Compute Bottlenecks

Microsoft says the rapid rise of agentic AI is exposing limits in power supply, memory bandwidth and system efficiency. Adding more chips and building ever-larger data centers can produce diminishing returns, making raw hardware capacity an incomplete measure of progress. The company is advocating a semiconductor-style “yield mindset” that focuses on how much useful computing output an AI system can reliably deliver from the resources deployed.

In its latest proposal, Microsoft called for co-design across model architecture, software orchestration, memory systems and chips to improve infrastructure utilization and lower the cost of each computation. The company did not disclose an investment figure, implementation date or quantified performance target. Its central argument is that the next phase of AI infrastructure competition will depend on affordable, widely accessible compute, rather than a continued race to construct the largest possible data centers.