Training and serving cutting-edge language models is getting pricier and less profitable. GPT-3 cost about $55 million to build and generated $200 million in revenue over a 30-month window, netting 3.6× returns. By contrast, GPT-4.5’s $5.3 billion total cost produced $3.9 billion in revenue before being overtaken in five months—a 0.7× recovery. Inference bills now dwarf training costs; OpenAI spent more on inference in 2024 than it earned. Each new model lives at the frontier for a shorter time, with margins eroding under GPU-heavy usage and collapsing API prices. To escape that trap, labs won’t just sell tokens—they’ll build discovery engines. The first frontier is science. AI-driven drug discovery slashes R&D costs by 25–40% and accelerates timelines by 30–40%, saving up to $1 billion per drug and cutting failure rates in half. A typical FDA-approved drug brings in $6.7 billion over its lifetime; Keytruda hit $31.7 billion in 2025 alone. Companies like Isomorphic Labs and Google’s Gemini Deep Think are already running AI-led design and publishing papers without human co-authors. Those discoveries retain value long after the underlying model is outdated. The second frontier is physical instrumentation. OpenAI and Ginkgo Bioworks ran 36,000 cell-free protein synthesis reactions, cutting protein production costs by 40% through iterative lab-model loops. DeepMind’s UK materials lab synthesizes hundreds of new compounds daily, creating data no competitor can replicate. Owning both the strongest closed-door models and the instruments that generate proprietary data creates two reinforcing flywheels: better models drive better research, and unique data feeds back into model improvements. That dual approach, not token sales, is where lasting value lies.