Open models are winning on economics. The bottleneck is now orchestration and

Open models are shifting from hobbyist curiosity to enterprise infrastructure, driven by cost pressure and the rise of coding agents. Ollama's data shows token usage has grown 150x this year, with the surge coming from open models powering agentic workflows. Chinese models dominate cloud deployments, while US and European models compete locally. Cost reduction is

the entry point, but customization and control are the long-term goals. The real unlock is a new class of 'flash' models—good enough for 80% of tasks, ultra-cheap, and fast. These enable unlimited-token usage patterns, similar to early ChatGPT. The bottleneck has moved from model intelligence to coordination, security, and workflow design. Enterprises like AT&T have

already shifted 40% of token consumption to open models, primarily for coding agents. The key tradeoff is between frontier closed models for hard problems and open models for volume work. Most enterprises will use a hybrid model—80-90% of tokens from open models, but only 10-20% of budget. The scarcity is shifting from tokens to the

layers above: orchestration, memory, credential management, and safety tooling. Open models lack the out-of-box safety infrastructure of closed providers, creating both risk and opportunity for new tooling businesses. The geopolitical dimension is real but manageable. The 'Manchurian Candidate' concern exists theoretically, but enterprise security teams are experienced with open-source supply chain risks. Model provenance matters