Summary
Y Combinator’s Lightcone podcast interviews Jeffrey Morgan, co-founder and CEO of Ollama — the open-model platform used by 9 million developers, with 178,000 GitHub stars and adoption across 85% of the Fortune 500. Morgan draws on Ollama’s own token-flow data to describe a measurable enterprise shift away from closed frontier models toward open-weight alternatives, driven primarily by the economics of coding agents and emerging co-work platforms like OpenClaw and the Hermes agent project.
The data points are striking: AT&T has already redirected 40% of its token consumption to open models. Graphs from Ollama’s cloud show that cloud-hosted inference is currently dominated almost entirely by Chinese-origin models — DeepSeek, MiniMax, Kimi — while local model usage splits roughly evenly between US and Chinese models. Morgan traces two key inflection points in per-developer weekly token consumption: an initial surge driven by open coding agents earlier in 2026, followed by a steeper April climb when OpenClaw extended long-horizon agentic workflows to non-developers in finance, support, marketing, and sales.
The conversation also covers Nvidia’s strategy of open-sourcing models and tooling to sustain hardware ecosystem dominance, the DGX Spark and upcoming DGX Station hardware (with GB300 chips) that can run frontier-scale models on-premise, and why enterprise cost reduction is the entry point but proprietary fine-tuning is the long-term north star. Morgan notes that Germany is a surprisingly large source of open-model cloud token consumption outside the US.
📺 Source: Y Combinator · Published September 04, 2026
🏷️ Format: Interview
