Open Models Change The Economics of AI
Ollama Co-founder Jeffrey Morgan discusses the enterprise shift to open models, driven by cost, customization, and coding agents, leading to a 150X token growth. He highlights the dominance of Chinese models in cloud consumption, the hardware renaissance enabling local LLMs, and the rise of ultra-efficient "flash" models. Morgan also shares Ollama's journey, from its YC pivot to its monetization strategy, and reflects on lessons for building AI companies.
The Enterprise Shift to Open Models and Ollama’s Role
Jeffrey Morgan, co-founder and CEO of Ollama, discusses the monumental shift towards open models within enterprises, highlighting the exponential growth in token consumption and the evolving landscape of AI development and deployment. The conversation delves into the strategic advantages of open models, the impact of next-generation hardware, and the unique challenges and opportunities in the open-source AI ecosystem.
Driving Forces Behind Open Model Adoption
Morgan emphasizes that while cost reduction is a primary motivator for enterprises adopting open models, the deeper objective is gaining greater control and customization over AI for their specific business needs. This allows them to tailor models for unique use cases, moving beyond short-term cost savings to long-term strategic advantage. An example cited is AT&T, which has already shifted 40% of its token consumption to open models, primarily for coding agents and AI assistants.
The Token Usage Explosion and AI Agents
The most significant driver of token usage growth in 2024 has been coding agents and AI assistants like OpenClaw and Hermes. Ollama’s data shows a 150x increase in per-developer token usage since the start of the year, driven by the emergence of open models capable of powering these agents (e.g., Kimmy, GLM, Miniaax). These agents automate large chunks of work, not just for developers but across various departments (finance, support, marketing), leveraging increased context windows (from 128K to over 1 million tokens).
Fine-Tuning: A Resurgent Trend
While interest in fine-tuning previously fluctuated, it is now seeing a resurgence. The pace of open model releases is accelerating (e.g., three iterations of Deepseek Flash in a summer), making custom training more challenging but also more critical. The gap between frontier closed models and open models is rapidly closing, partly due to AI safety concerns potentially slowing down frontier labs. Open-source models, especially those with powerful cybersecurity capabilities like GLM53, offer unique advantages, particularly in security testing, where closed models often refuse to perform tasks like penetration testing.
Ollama’s “Operating System” Role
Ollama positions itself as an **