Open Models Change The Economics of AI

Video summary

Ollama CEO Jeffrey Morgan says open models could handle 80–90% of enterprise tokens as cloud usage surged 150x.

Ollama co-founder and CEO Jeffrey Morgan says enterprise AI use is shifting toward open models, driven first by coding agents and then by OpenClaw and Hermes, which brought long-running agent tasks to non-developers. Ollama Cloud token usage has risen 150x since the start of the year, while AT&T has shifted 40% of its token consumption to open models. Morgan expects open models to process 80–90% of business tokens, though frontier models will remain important for the hardest tasks. He describes Ollama as an operating-system-like layer that connects models, hardware, inference providers, and agent harnesses, and predicts a hybrid future: run simpler work locally and send demanding coding tasks to the cloud. The conversation also traces Ollama’s path from Docker and years of searching for a product-market fit to its 2023 launch, when making local models easy to run quickly attracted users and enterprise adoption.

Chapters

1. 0:00Intro: Ollama serves 9 million developers as enterprise use shifts to open models
2. 2:13Is It All About Cost?: AT&T moves 40% of token use to open models
3. 3:31The Token Usage Explosion: Ollama Cloud token usage grows 150x
4. 5:31Fine Tuning: Faster DeepSeek Releases and Open Models for Security Testing
5. 8:26Launching Models at Scale: DeepSeek’s multimodal launch and Ollama’s day-zero playbook
6. 11:31Ollama as the OS Layer: Connecting inference, hardware, and developer runtimes
7. 13:58Hidden Layers Between Model and App: Knowledge, orchestration, and persistent memory
8. 17:28Open vs. Closed: Open models may handle 80–90% of enterprise tokens
9. 20:41Local vs. Cloud Models: Quinn 3.8 38B benchmarks match Opus 4.6 for coding
10. 24:00NVIDIA's Open Source Play: DGX Spark runs 20–120B models with 128 GB memory
11. 26:03The Return to Local: GB300 desktops could bring coding agents back from the cloud
12. 27:15Getting GPUs Is Hard: B200 and B300 Access Requires Provider Partnerships
13. 29:16The Flash Model Revolution: DeepSeek Flash Makes High-Volume AI Cheaper
14. 31:35God Model vs. Orchestration: Kimi’s Web-Development Results Intensify Competition
15. 33:37The Geopolitics Question: Model Origin, US Hosting, and Llama at a Finnish Power Plant
16. 37:12Applied to YC With the Wrong Idea: From Kubernetes Desktop to Ollama in 2023
17. 40:36Lost in the Wilderness for Two Years: A Team of 10+ Found Its Direction by Running Llama
18. 42:15The Pivot That Changed Everything: Ollama’s First Version Shipped in Two Weeks
19. 44:49100K GitHub Stars, No Revenue: Ollama Cloud Monetization Followed Enterprise Adoption
20. 47:40How Do You Monetize Open Source? Coding Agents Made Open Models Ready for Harder Problems
21. 49:43Why Do YC as a Second-Time Founder? Founder Peers and Lessons from Docker
22. 51:49What Seeing “Good” Actually Does for You: Product-Market Fit and Two-Year Software Quality
23. 53:38Old DevOps Rules That Don't Apply Anymore: Ollama's Curation of Fragmented Models