Thomson Reuters builds own AI model instead of renting GPT

What happened

Thomson Reuters has built its own large language model, named "Thomson," rather than relying solely on third-party providers such as OpenAI or Anthropic. According to The Decoder, the model is built on Alibaba's open-weight Qwen architecture and represents an investment of roughly $40 million over two years.

The model's strongest benchmark results appear when it is paired with Thomson Reuters' own proprietary content, such as its Westlaw legal database, rather than operating as a general-purpose system. Chief Technology Officer Joel Hron has framed the strategic logic not as a bet on raw model intelligence, but as a decision about which parts of the AI stack a company needs to own outright versus rent from external vendors.

Why it matters

The move signals a broader shift in enterprise AI strategy: rather than treating large language models as interchangeable commodities accessed via API, some organisations with deep proprietary data assets are choosing to internalise the model layer itself. For Thomson Reuters, whose value proposition rests heavily on curated, authoritative content like case law and regulatory filings, owning the model that sits closest to that data may be seen as a way to protect differentiation and control how that content is surfaced and monetised.

For digital transformation leaders more broadly, this points to a maturing question in AI adoption: it is no longer simply "build versus buy" for an entire AI capability, but a more granular calculation about which specific layers — data, retrieval, fine-tuning, orchestration — justify direct ownership, and which can safely remain rented commodity infrastructure.

By the numbers

• $40 million — approximate investment in developing the Thomson model over two years, per The Decoder
• 2 years — development timeframe cited for the project

The Renascence take

The detail that will get skipped over in most coverage is the benchmark caveat: Thomson only excels when it can draw on Thomson Reuters' own proprietary content. That is not a footnote — it is the entire thesis.

What Thomson Reuters has effectively proven is that a proprietary model is only as valuable as the proprietary data behind it — the intelligence itself is increasingly a commodity, but the content it retrieves is not. For experience and service leaders, the lesson isn't "build your own model"; it's "map which of your data assets create defensible advantage when paired with AI, and own the connective layer around those, not the whole stack." Most organisations chasing model ownership for its own sake will spend heavily to rebuild something the market already rents cheaply, while the few with genuinely unique content moats — regulatory data, transaction history, service logs — stand to gain the most from doing exactly what Thomson Reuters just did.