AI labs segment pricing to sustain demand amid capacity shortage

Hyperscalers' Q2 2026 earnings calls painted a picture of sustained scarcity: Sundar Pichai (Alphabet), Andy Jassy (Amazon), and Jensen Huang (NVIDIA) all affirmed that AI compute supply will lag demand not just through 2026 but into 2027, with 2028 demand already heavily booked. This creates a classic pricing dilemma: capacity shortage typically forces prices up, but higher prices destroy the consumption growth that justifies further infrastructure investment. AI labs appear to be solving this via segmentation rather than uniform pricing. Instead of raising all prices or cutting all prices, they are launching distinct model tiers—premium (Anthropic's Fable 5 at $50/M output tokens), mid-market (GPT-5.6 Sol, Kimi K3), and value (GLM-5.2 down to DeepSeek V4 Flash at $0.03/M tokens)—each capturing a different cost-sensitivity cohort. The mid-market tier is the novel entry, positioned to convert price-sensitive buyers who would otherwise forgo consumption entirely. By keeping the value tier cheap and the mid-market tier performant at a fraction of premium cost, labs aim to sustain Jevons' Paradox—the dynamic that lower cost drives higher total consumption—even while raising premium prices. This hinges on a new layer of competition: routers, the software that directs customer queries to the economically optimal model for each workload. Routers become the gatekeeper between buyer and model, and labs that lack competitive offerings in all three tiers risk losing router selection battles to rivals or startups that can undercut them on price or capability at a specific tier.