The analyst contends that the profitability of training better frontier AI models recursively accrues value to AI labs, justifying massive investments in semiconductor capex and R&D for HBM, interconnects, photonics, and packaging. Conversely, a world of cheap, open-source intelligence would lead to a cost-cutting 'knife fight' among inference providers, disincentivizing large training runs and undermining the semiconductor market.
The economic model for AI infrastructure investment hinges on the continued profitability of frontier model training, not just widespread inference.