Thomson Reuters launched its own large language model this week, called Thomson, built from an open-source base and trained on decades of Westlaw, Practical Law, Checkpoint, and Reuters content. The company put the program at around $40 million in talent and compute. Buried in the coverage is the number that actually matters: the final training run for the version they shipped cost about $450,000.
Four hundred fifty thousand dollars, for a frontier-grade model in a domain where being right is the entire product. That number is a lesson, and it is not the one the headline suggests.
The compute was cheap. The data was the point.
A common belief is that building a serious model means a nine-figure compute bill only a few labs can afford. Thomson Reuters just built a frontier-quality legal model where the decisive training run was smaller than a mid-size engineering team's quarterly salary. The expensive parts were talent and the many runs it took to get the recipe right. The thing that made it possible was not compute at all. It was owning Westlaw.
That is the insight worth taking from this. You cannot out-general the frontier labs, and you should stop trying. What you can do, if you sit on decades of proprietary, authoritative, domain-specific content, is build a model that beats the generalists on your turf for a price that would not fund a Super Bowl ad. The moat was never the GPUs. It was the corpus, and Thomson Reuters has been assembling that corpus since before machine learning was a phrase anyone used.
What this means for people without a Westlaw
Most of us do not own decades of curated legal content, so the direct move is off the table. The principle underneath it is not.
The generic model layer is a commodity, and it is getting cheaper by the week. The durable advantage is proprietary data that a general model has never seen and cannot buy: your users' anonymized behavior, your niche's hard-won documentation, the labeled edge cases you accumulated by operating in a corner of the world nobody else bothered with. A general model plus a small amount of data nobody else has beats a bigger general model with none. That is the shape of a defensible AI product now.
It also reframes what "data strategy" means for a small team. Every time you operate in your niche you generate signal, and most teams throw it away. The Thomson lesson is that the signal is the asset. If you are keeping it, structuring it, and treating it as the thing you will one day train or fine-tune on, you are building the only moat that survives the model layer commoditizing under you.
The honest caveats
Two of them. First, $450,000 is one clean training run, and it hides the years of accumulated data and the $40 million of surrounding work that made that run possible. The cheap number sits on top of an expensive foundation, and the foundation is the part you cannot skip. Second, a domain model is only as trustworthy as its evaluation, and legal work is a domain where a confident wrong answer is worse than no answer. The interesting test for Thomson is not its launch benchmark. It is whether the lawyers who depend on CoCounsel trust it after six months of real cases.
Even with those caveats, the signal is clear and a little liberating. Owning the right data is worth more than owning the biggest cluster, and the right data is often something a focused team can accumulate without permission from anyone. The frontier labs are building the general engine. The defensible businesses are being built on top of it, out of data only one company has.
If you have built something valuable on proprietary data a general model could never reach, I would like to hear what the data was, and how you knew it was the moat before it paid off.
Top comments (0)