AI Model Costs Diverge as Performance Converges
Source: Tomasz Tunguz
When five competing frontier models cluster within five intelligence points but differ 3.8x in cost-per-unit, pricing becomes the primary battleground. Google's willingness to undercut on inference suggests the profit margins that justified premium positioning are evaporating. This mirrors previous infrastructure commoditization cycles: once core capability gaps close, switching costs collapse and unit economics determine market share. Vendors compete on utilization efficiency rather than model superiority. For enterprise buyers, negotiating power shifts dramatically. For frontier labs, defending margin requires proprietary applications or vertical integration rather than model-only sales.