Edge Inference Is the Next Capacity Bottleneck

Most of the market's attention on compute scarcity has focused on frontier model training runs. We think the more investable bottleneck over the next four quarters is inference capacity at the edge.

Why it matters

Enterprise deployment of smaller, fine-tuned models is accelerating faster than edge silicon roadmaps can absorb. That's a setup we've seen before in prior compute cycles, and it tends to reward the picks-and-shovels layer over the application layer.

Our view

We favor semiconductor equipment and packaging names with the clearest path to shipping edge-optimized silicon in the next two product cycles.