Headlines: $NVDA trading at its cheapest valuation in a decade That’s tempting, but when you dig a bit deeper, it makes sense. Here is why: $NVDA is the king of training. However, as performance gains from training plateau, AI workloads increasingly shift to inference. We have already seen this with pre-training as its compute share declined from 67% in 2024 Q1 to 13% last quarter. Yet total share of training has remained somewhat stable around 62-63%, as post-training replaced pre-training. Gains from post training will inevitably plateau as well, and we’ll see inference expand its market share. Over the medium term, possibly by 2030, we’ll see inference take up more than 50% of compute share. ASICs are expected make up for 50% of inference compute while training will keep being 8
View on X ↗