Why companies are moving from renting AI to owning models

My notes

Summary

Companies are moving through three stages of AI adoption: renting frontier models from Anthropic, OpenAI or Google; switching to cheaper open-source models; and now, bringing model training fully in-house on proprietary data. CNBC cites Meituan (China’s food delivery giant) as an example already running its own in-house model built on open source, and notes Together AI just raised $800 million betting more companies will follow.

Key Insight

• Three-stage adoption curve: rent frontier model, adopt open source (cheaper), train and own a custom model in-house.
• Banks and financial services were early movers on internal AI platforms, even while still relying on outside models for some workloads, a hybrid stage before full ownership.
• Meituan (China) already trained its own competitive model in-house starting from open source, and developers are adopting it internally, a concrete signal the “own it” stage is reachable, not just theoretical.
• Together AI raised $800 million specifically to serve this shift (infrastructure and tooling for companies that want to train and run their own models).
• Stated business rationale for owning the model rather than renting:
  • Data control, not handing your data to a third party that could use it to train competing models.
  • Efficiency, a smaller base model fine-tuned for a narrow task can beat a large general model on cost per token while matching task-specific quality.
  • Lower inference cost, more “intelligence per dollar” once the model is right-sized for the actual use case.
• Framed as a threat vector for frontier labs (Anthropic, OpenAI, Google): open source already eroded price leverage; in-house training threatens the “rent our model” business model entirely for large, well-resourced customers.