Over a little more than two years, Thomson Reuters has spent about $40 million on a team and compute for its own language model, and the model is already live under the name "Thomson". Instead of renting models from OpenAI or Anthropic, the company is betting on full independence and uses Alibaba Qwen as the base.
The market is used to chasing the newest models from the big vendors. But this news is a serious signal for anyone building complex AI systems. It marks a shift from "buy or rent" to "own and control".
What does this change? Large corporations used to adapt their processes to the limits of public APIs. Now Thomson Reuters is building AI that fits directly into its legal and news archive, built up over many years. This is no longer fine-tuning. It is building an AI asset of their own.
For me, this means:
- Control over data: when the model runs inside the company, you decide for yourself how to further train it and on what data. That lets you keep confidentiality and accuracy as high as possible in specialized domains. This is critical for RAG systems in closed environments.
- Architectural flexibility: your own stack lets you integrate AI more deeply into existing workflows and agent pipelines, instead of piling wrappers on top of third-party APIs. Fewer workarounds, fewer surprises.
- Long-term cost and risk: $40 million is an investment, not an operating expense that can suddenly go up. It reduces dependence on vendor pricing and opens the way to competitive advantages that others don't have.
Takeaway: for large companies with huge volumes of sensitive data and specialized domain expertise, like Thomson Reuters, building their own AI can make more sense than renting forever. It is harder, slower and more expensive at the start, but it gives control, security and flexibility in the long run. For small and mid-sized teams that need to ship an MVP fast, cloud APIs will remain the main option.
You can read more about the Thomson Reuters decision in this article from The Decoder.