Thomson Reuters has expanded its partnership with Cohere, the Toronto-based enterprise AI startup, together with a $40 million investment. According to The Decoder, the move is part of a strategy to bring core AI capabilities in-house rather than depend on cloud APIs from OpenAI or Anthropic for its most sensitive products.
That means Westlaw, Practical Law, and the CoCounsel assistant — the company's main legal and tax tools — will run on Cohere models that Thomson Reuters says it can shape, fine-tune, and deploy on its own infrastructure.
Owning vs. renting: what's actually different?
"Renting" AI is how most companies use it today: you call an API, pay per token, and your data travels to the vendor. For a legal research assistant that reads through thousands of court rulings, that per-token bill adds up fast — and the model underneath can change overnight, with no say from you.
"Owning" in this case means two things. First, equity: the $40M gives Thomson Reuters a stake in Cohere's roadmap and a reason to steer its model development. Second, control: the two companies will co-develop models focused on legal and tax work, and Thomson Reuters gets deployment flexibility that a pure API customer never has.
The contrast with competitors is telling. Harvey, the legal AI startup that dominated headlines this year, is closely tied to OpenAI and Anthropic. Lexis+ AI from LexisNexis leans heavily on Microsoft and OpenAI. Thomson Reuters is taking a different route: it bought Casetext for $650 million in 2023 to own CoCounsel outright, and now it's investing to own the model layer underneath its products too.
What $40 million actually buys
For a company with roughly $6.7 billion in annual revenue, $40M is about 0.6% of one year's turnover — a small check in absolute terms, but strategically placed. For European readers, that's roughly €35 million at current exchange rates. It is not a prepayment for tokens; it's a stake in the model supplier and a hedge against the pricing and policy whims of the big US API vendors.
That math matters for anyone running AI in production. API costs are variable and scale with usage; an owned model has high fixed costs — people, GPUs, training — but far lower marginal cost per query. For document-heavy legal work, where a single research session can consume millions of tokens, that difference is decisive.
The European angle: GDPR and the AI Act
This is where the deal gets interesting for European readers. Legal and tax data is among the most sensitive categories a company can touch: court filings contain personal data, and attorney-client privilege is non-negotiable. Processing that data through an API operated under US jurisdiction creates GDPR transfer issues that European law firms increasingly refuse to ignore.
Cohere has built its enterprise pitch exactly around this: data privacy, flexible deployment, and availability in Europe. Thomson Reuters' Westlaw and Practical Law are standard tools in European law firms, and the AI layer now being built on top of them will land on European desks too. For those firms, the question "where is my data processed and who owns the model?" is becoming a procurement criterion, not a footnote.
The EU's AI Act adds more weight. Legal research isn't automatically in the "high-risk" category, but the regulation imposes transparency, logging, and human-oversight duties on providers and deployers. A company that controls its models — rather than consuming black-box APIs — has a much easier path to demonstrating compliance when regulators ask.
What we see from our own rig
At AI Arena, we run local LLMs on an RTX 5060 Ti 16 GB alongside cloud models, and the tradeoff is visible every day. Local models give you privacy, predictable cost, and control — but you carry the infrastructure burden, and frontier cloud models still win on raw capability.
Thomson Reuters is making a scaled-up version of that bet: that a model it helps build, fine-tuned on legal corpora, can be good enough — and that control and data sovereignty are worth more than the last few points of benchmark score. For legal work, where provenance and consistency beat creative flair, that calculus is reasonable.
What European companies can take from this
The practical lesson isn't "everyone should invest in an AI startup." It's that enterprise AI procurement is splitting into two camps: those who rent general-purpose intelligence from API vendors, and those who buy or co-develop specialized models they can control. For regulated sectors in the EU — law, tax, finance, health — the second camp is becoming structurally attractive.
If your company processes sensitive European data, the questions worth asking are: Where are my tokens actually processed? Who can see my prompts? And what happens when the vendor changes pricing, policy, or the model itself overnight? Thomson Reuters just paid $40M to have better answers than most.
Is Cohere's technology available to European companies today?
Yes — Cohere's models are available via API from Europe, and the company emphasizes GDPR-aligned deployment options, which is one reason Thomson Reuters chose it for sensitive legal data.
Does this mean Thomson Reuters is abandoning OpenAI?
Not entirely. The company still has partnerships tied to Microsoft's Azure OpenAI for some products. The Cohere investment adds a strategically owned alternative for its most sensitive legal and tax workflows — and reduces dependency on a single vendor.
Will European law firms see different pricing because of this deal?
Probably not directly. Thomson Reuters sells outcomes — subscriptions to Westlaw, Practical Law, and CoCounsel — not tokens. The change is in its cost structure and control, not in how customers are billed.
