Thomson Reuters built its own $40M frontier model

Thomson Reuters spent $40 million training its own frontier model rather than renting one, per Thomson Reuters' announcement on August 24.

$40 million is a small sum next to what a corpus like Westlaw and Practical Law is worth. Feeding that corpus through a rented model's inference risks turning the moat into the vendor's capability.

The model, called Thomson, specializes an open-source base on Westlaw, Practical Law, Checkpoint and Reuters content, with input from hundreds of subject matter experts. First deployment is Tabular Analysis inside CoCounsel Legal for document review; the rest of the legal and tax portfolio, plus sovereign AI options, follow later.

Satya Nadella made a related argument from the buyer's side. In an interview published August 25, he said any firm without model control "will not remain a firm because you've essentially outsourced your thinking." He pitched model portability, the ability to swap out any one model, as the fix, Windows Latest reported. He runs the biggest AI reseller in the world, and he's arguing publicly against single-vendor lock-in at the model layer, the same stance Thomson Reuters backed with cash a day earlier, per Windows Latest.

Revolut launched a standalone AI research lab built on Pragma, its own foundation model, according to Finextra. Pragma trained on about 40 billion banking events from roughly 25 million users.

None of this means solo builders should train their own models. It means the build-versus-rent question has moved from can we afford it to what does our data become if we don't.