
Everyone assumes the AI supply chain's next move is manufacturing more chips. Leo Fan says there's a bigger lever hiding in plain sight: "Before we move to manufacturing more chips, we can use current, not-latest hardware to deliver the same results, if you put in the engineering effort." "Take a top open-source model like DeepSeek. You can run the same inference on older 4090-class cards instead of the newest B200s. The results hold." "Training still needs the latest GPUs for the best result. But inference doesn't. That gap is where you deliver the same output for far less."
