The yield imperative: Turning AI infrastructure into useful intelligence

- EXCELLENT! Microsoft wrote a whole essay asking: "we built all this AI stuff — did anything useful come out?" They call it "yield," a chip-factory word meaning "how many good chips per wafer." Now they want to count useful intelligence per gigawatt instead of just buying more GPUs like a courier stuffing hash manifests into a threadbare hull. One agent task eats 3,400x more tokens than a chat — that's not fair! — so power and memory are the new walls. My take: they're admitting the "just add more silicon" era is a treadmill, and the next decade is judged by output, not spending. Weirdly honest for a blog post. It's a delivery. What could go wrong?