Thomson Reuters spent $40 million to own the layer above the open weights

Run the two disclosed figures against each other and the training bill comes to about 1.1 percent of the stated program cost ^[16]. That ratio is why the smaller number travels so well in board decks, and it is also why the deck is incomplete: Thomson Reuters says the $40 million leaves out what it paid for Safe Sign Technologies, which it has not disclosed, along with staff, infrastructure, data preparation, experimentation, expert evaluation and continuing operating costs ^[5]. A firm reading this as a $450,000 entry fee has misread the invoice.

Where the money went is more useful to a buyer than the total, because it separates the inputs that can be rented from the ones that cannot. The compute was rented and ring-fenced in the cloud rather than built, with outside partners including Imperial College London, DatologyAI and Lambda ^[11]. The technical team was modest by frontier-lab standards, no more than three dozen engineers and scientists, drawing on as many as 368 Nvidia B200 GPUs during experimentation ^[10]. The input that was neither modest nor rentable was judgment: hundreds of subject-matter experts took part across the program, and the company says its experts curated more than 11,000 internal evaluation items, with thousands of hours of lawyer time spent deciding whether an answer was complete and supported rather than whether it read well ^[9].

A skeptic will say the flattering result is a company grading its own work, and on that narrow point the skeptic is right. The blind study of 35 attorney-editors and more than 3,000 comparisons was designed and reported by Thomson Reuters, has not been independently replicated, and set the full Thomson system, with Westlaw, Practical Law and Reuters behind it, against rival systems equipped with web search ^[7][8]. At the stated minimum that is roughly 86 comparisons per participant ^[17]. What the study supports is narrower than a model win: a model plus a licensed professional corpus was preferred over a model plus the open web on legal work, which is what a buyer would in fact be buying ^[7].

The architectural detail with the longest half-life is that Thomson Reuters says it changed the root model many times as the open-weight frontier advanced, with the large version ultimately starting from Alibaba's Qwen3.5-397B weights mediated through work done with Imperial ^[13][12]. That makes the base model a depreciating input and the evaluation apparatus the durable one, since the 11,000 items and the expert panel survive a swap that the weights do not. It also puts a ceiling on the sovereignty claim. What the company owns is the layer above someone else's open weights, plus the corpus and the graders, which is a narrower and more defensible position than owning a frontier model.

The incentive behind the whole exercise is worth naming, because it does not transfer cleanly. Thomson Reuters sells professional information, and training on selected Westlaw, Practical Law, Checkpoint and Reuters material defends the value of an asset it already owns ^[15]. A bank or a manufacturer whose corpus is purely internal gets control out of the same architecture rather than a product line, and should price it as an operating decision instead of a revenue one. The recurring cost is the graders, not the GPUs, and that is the line a build approved this quarter will still be paying next year.