QuoteTesla's boast that its future AI5 chip will run inference 10x cheaper than Nvidia's Blackwell architecture was pretty shortlived. The next-gen Nvidia Rubin AI platform offers a tenfold reduction in costs per token.

10x? I believe it when I see it (the energy efficiency has probably improved by ~30% (not, if it's still on the same node), which is normal for a gen-over-gen full node improvement, but this would only be the raw chip improvements, and there might be more). Are they comparing apples to apples (e.g. FP4 vs FP4), or as usual, apples to dingleberries?