AI labs segment pricing to sustain demand amid capacity shortage

In Q2 2026 earnings calls, major cloud and AI companies acknowledged that AI compute capacity remains the binding constraint on growth. Sundar Pichai of Alphabet stated simply, "We continue to be supply constrained, a sign of momentum & rapid adoption." Andy Jassy of Amazon was more explicit: "We will still not have enough capacity to meet all the demand we have in 2026, & I believe this dynamic will also be true in 2027 too. The demand we already have for 2028 is striking." Jensen Huang at NVIDIA added that "Blackwell sales are off the charts, & cloud GPUs are sold out." These declarations set up a paradox: persistent scarcity typically triggers price increases, yet AI labs—facing capacity constraints—are pursuing a mixed pricing strategy rather than raising prices uniformly across the board. Anthropic moved first, launching Fable 5 on July 24 at $50 per million output tokens, double the price of its prior flagship Opus 5 and a new ceiling for model pricing. Google's trajectory was similar: Gemini's flagship climbed from $1.50 to $12 per million tokens across four generations. But the picture reversed when OpenAI cut prices sharply, reducing GPT-5.6 Luna by 80% just five days after Fable 5's launch, an aggressive move attributed either to market capture or a fundamental breakthrough in cost. The labs' apparent solution is segmentation into three tiers, each serving a different customer segment and use case. The premium tier—Anthropic's Fable 5, Anthropic's Opus 5, and others—commands the highest price. The mid-market tier is newly ascendant: GPT-5.6 Sol (OpenAI) and Kimi K3 deliver 96% of frontier intelligence at 40% of premium cost, making them attractive for workloads where maximum capability is less critical. The value tier ranges from GLM-5.2 down to DeepSeek V4 Flash at $0.03 per million tokens, offering 84% of frontier intelligence at 1 to 5% of premium cost. The bet is that this segmentation sustains Jevons' Paradox—the economic principle that as the cost of a resource falls, total consumption rises—by absorbing workloads that would otherwise drop out of the market at higher prices. As this segmentation becomes standard, the router—software that intelligently routes a query to the most appropriate model for a given task and budget—emerges as the strategic battleground. Routers will live in three places: embedded inside models themselves, built into customer applications, and in middleware harnesses. The major labs each need competitive entries across all three tiers, or they lose the router auction to competitors who can undercut them on price. Startups have an advantage: by owning a single point on the frontier—such as DeepSeek V4 Flash's existence proof of extreme cost efficiency at $0.03 per million tokens—they can capture router selection even if they cannot compete across all tiers.