What a $40M legal AI says about building with your own data

Thomson Reuters, the company behind Westlaw legal research and Reuters news, just launched its own AI model called Thomson. On legal tasks, it holds its own against the most powerful AI models in the world — and the whole thing cost $40 million to build. The actual training run, the step where the model learns from data, cost roughly $450,000 to complete. Compare that to the billions OpenAI and Google spend, and the difference is striking.

Thomson Reuters did not start from zero. They took a strong open-source AI model — freely available online — and specialized it deeply. They trained it on decades of their own content: Westlaw case law, legal documents, Reuters news, and financial records from Checkpoint. Hundreds of their own subject-matter experts helped shape how it answers questions. Its first job is helping lawyers review documents faster in their CoCounsel legal tool.

The key shift is this: you do not need the biggest budget to build a great AI model. You need the right data. Frontier labs spend billions training on almost everything on the internet. Thomson Reuters spent far less, but used some of the best-curated legal content in the world. On legal tasks, Thomson performs as well as or better than models that cost far more. The right data beats the biggest budget.

Thomson Reuters has barely started. The model was trained on less than 10% of their content library, and there is more to come. If you run a business with years of structured, specialized data — legal files, medical records, financial reports, customer conversations — this is a clear signal: your data may be worth building on. Take a good open-source foundation, specialize it for your domain, and you may end up with something more useful than any off-the-shelf API.