Front model token prices have fallen sharply over the past two years, yet most enterprises are watching their AI bills climb, not shrink. The reason is structural, not pricing: agentic AI doesn’t just answer a question, it plans, calls tools, retries failed steps, checks its own work, and re-sends accumulated context at every step along the way. A single agentic workflow can burn through hundreds of thousands, even millions, of tokens to complete one task. This is far more than the simple chatbot exchanges during the early days of LLMs. Cheaper tokens turned out to be an invitation to use more of them, and “tokenmaxxing”, i.e., throwing larger context windows and more reasoning steps at every problem, has become the default rather than the exception.
This is the gap our portfolio company Granica is targeting with Myelin, a new product sitting alongside its existing data-cost platform, Granica Crunch. Myelin sits between a company’s AI agents and the underlying model, reducing token consumption through intelligent compression and memory management. And it does this without summarizing or altering meaning, so teams get the same output for a fraction of the spend. Granica reports token consumption reductions of roughly 55-60% in production use, with an early live demo across 55 billion tokens showing cost reductions in the 39-60% range. Moreover, Myelin is designed to improve the more it runs. Paired with Crunch’s 30-60% storage-cost reductions, Granica is positioning itself to address two of the fastest-growing cost centers in enterprise AI today.
