Thomson Reuters is investing $40 million over two years to build and deploy its own large language model rather than licensing technology from OpenAI, Anthropic, or other external providers. The company calls its model "Thomson," and it runs on top of Alibaba's Qwen foundational model.
The decision reflects a broader industry shift among large enterprises. Instead of renting AI capabilities through API subscriptions, information-heavy companies now recognize that competitive advantage comes from controlling models trained on proprietary data. For Thomson Reuters, that proprietary asset is decades of legal, regulatory, and financial content accumulated across products like Westlaw, Westlaw+ and Refinitiv.
Chief Technology Officer Joel Hron frames the strategy clearly. Intelligence itself is a commodity now. What matters is owning the right intelligence for your specific domain. Thomson Reuters competes in markets where accuracy and legal precision carry real consequences. When Thomson accesses Thomson Reuters' own corpus of case law, legal precedent, and regulatory guidance, the model delivers measurably better results than generic alternatives. The company's benchmarks show top performance when the model operates within its native knowledge domain.
This approach differs fundamentally from the licensing model that dominated 2023 and early 2024. Firms like Microsoft and Google built scale by integrating OpenAI and Gemini into enterprise products. Thomson Reuters ran the same playbook initially, partnering with OpenAI for various AI features. But the company concluded that long-term differentiation requires vertical integration of AI infrastructure.
The financial math supports the strategy. $40 million over two years amounts to roughly $20 million annually, a manageable line item for a $7 billion revenue company. By comparison, ongoing licensing fees to OpenAI or Anthropic would likely compound over time, especially as usage scales across Westlaw, Westlaw+ and other Thomson Reuters products. Building once and owning forever costs less than perpetual rents.
Using Alibaba's Qwen as the foundation layer proved pragmatic. Thomson Reuters gains access to a competitive open-source model without building a foundational model from scratch, which would cost hundreds of millions or billions. The company then fine-tunes Qwen on its proprietary legal and financial content, creating a specialized model that outperforms generic competitors on domain-specific tasks.
The benchmarks matter less for general capabilities than for domain performance. Thomson's strong performance when accessing Thomson Reuters' own datasets tells the real story. The model falters on general knowledge questions without that proprietary context, which is expected and acceptable. Thomson Reuters does not compete with ChatGPT on trivia or creative writing. It competes on delivering accurate legal research, regulatory intelligence, and financial analysis.
Other major enterprises watch Thomson Reuters closely. Companies like Bloomberg, Capital IQ, and LexisNexis face identical decisions: license or build. Thomson Reuters' $40 million bet signals that building pays off for information monopolies. The model becomes a moat, a reason customers stay locked into Thomson Reuters ecosystems rather than switching to cheaper generalist alternatives.
Deployment timelines remain unclear, but the company likely integrates Thomson incrementally into existing products throughout 2026 and 2027. Westlaw users may see improvements in legal research results first, followed by broader rollout across other Thomson Reuters products. The real test comes when the model operates at scale handling millions of queries against billions of documents, maintaining accuracy while controlling costs.