Microsoft turns 3,400x token gap into useful yield

Microsoft is urging AI infrastructure buyers to track “useful yield”—useful output per dollar and watt, plus throughput and latency—as agentic tasks can consume more than 3,400 times as many tokens as a typical chat interaction, Microsoft’s infrastructure essay says. Its SEMICON Taiwan keynote summary treats capability, deployment velocity, and utilization as compounding inputs, but this remains a Microsoft operating thesis rather than an industry-standard benchmark. For teams tracking hyperscaler AI capex, file the 3,400× figure as a warning against buying peak capacity: compare useful work per dollar, watt, and deployed hour before asking for more GPUs.