On the weird economics of semiconductors and GenAI (gauthierroussilhe.com)

The semiconductor industry, long driven by Moore’s Law's economic principles, faces unique challenges as generative AI (genAI) demands exponentially growing computing power. Moore’s Law, originally an economic guideline for maximizing cost efficiency in chip design, is now intersecting with Rock’s Law, which highlights escalating costs for advanced chip fabrication plants—making cutting-edge manufacturing accessible to only a few players. This creates a delicate balance: semiconductor firms must produce increasingly powerful chips at scale while managing soaring production costs, a dynamic intensified by genAI’s hunger for GPUs and specialized hardware. GenAI providers like OpenAI and Microsoft are mirroring the historical “Wintel” alliance's role in software-hardware synergy by needing to consume the computing power supplied by the semiconductor industry to sustain growth. However, unlike traditional software, genAI incurs direct marginal costs per inference (electricity and hardware use), complicating profitability and growth strategies. Providers must navigate how to rapidly scale user adoption—often relying on free tiers—to drive demand, while aggressively optimizing inference efficiency through innovations like lower-precision computations and sparsity to reduce costs. The core economic tension lies in aligning efficiency improvements with the semiconductor industry’s need to absorb escalating production capacity. Excess efficiency risks underutilizing silicon supply, potentially triggering industry downturns, while insufficient efficiency hampers profitability and slows user adoption. This interplay will shape the future trajectory of AI hardware development and investment, influencing both global semiconductor cycles and the sustainability of genAI business models.