Who is in Charge? You or the Robot?Sovereign AI — At the country level, at the firm level.gettyLast November, I argued in these pages that AI has become the new steel and the new oil, and that every nation needs sovereign AI—the capacity to produce intelligence rather than simply import it. After that column ran, the question I heard most from executives was a fair one: fine for France and the UAE, but what does sovereignty mean for my company?On August 24, Thomson Reuters offered the most serious answer yet.The company behind Westlaw and Reuters launched Thomson, its first proprietary large language model. It did not build a frontier model from scratch. It began with a powerful open-weight foundation, then used its own content, experts, tools and training methods to create a model it controls. Thomson Reuters is also exploring deployments that would place the model inside a customer's controlled cloud environment, where it could work with that customer's intellectual property without sending the information to a frontier-model provider.The economics deserve attention—but also precision. Thomson Reuters says it spent approximately $40 million developing Thomson after acquiring the startup Safe Sign Technologies for an undisclosed price. The final three-week training run incurred less than $450,000 in estimated GPU costs. That is a startling figure, but it is not the total cost of the model. It excludes the acquisition and should not be confused with the staff, infrastructure, data preparation, experimentation, expert evaluation and continuing operating costs behind the program.The early performance is impressive, although it too requires precision. According to the company's technical report, Thomson-1.0-Large's aggregate benchmark score exceeded GPT-5.4 and Claude Sonnet 5, but trailed Claude Opus 4.8. It did not win everywhere: coding, abstract reasoning, mathematics and some robustness measures remained weaker than the strongest competing models.Thomson Reuters also conducted a blind study involving 35 attorney-editors and more than 3,000 comparisons. Participants generally preferred the complete Thomson system over the OpenAI and Anthropic systems tested, particularly on legal work. But this was a comparison of systems, not just models. Thomson had access to Westlaw, Practical Law and Reuters; the competing systems had web search. The study was designed and reported by Thomson Reuters and has not yet been independently replicated.Those qualifications do not erase the result. They explain it. Thomson Reuters did not win simply by producing a cleverer model. It assembled a better intelligence-production system for the work it knows best.What Thomson Reuters Actually BuiltIn August 2024, Thomson Reuters acquired Safe Sign, a Cambridge-centered startup whose team included researchers and practitioners with backgrounds spanning Cambridge, DeepMind, Harvard and MIT. Thomson Reuters subsequently established a joint Frontier AI Research Lab with Imperial College London.The large version of Thomson ultimately began with Alibaba's Qwen3.5-397B open weights, mediated through the Snowdon work developed with Imperial. Thomson Reuters then subjected that foundation to extensive mid-training and post-training using selected material from Westlaw, Practical Law, Checkpoint and Reuters, together with synthetic and expert-generated data.The team was not enormous by frontier-lab standards, but it was hardly trivial: the technical report says it included no more than three dozen engineers and scientists and used as many as 368 Nvidia B200 GPUs during experimentation. The work also involved outside partners, including Imperial College London, DatologyAI and Lambda. Thomson Reuters rented and ring-fenced its compute in the cloud rather than constructing its own data center.The scarcest input was not the GPUs. It was expert judgment. Hundreds of subject-matter experts participated across the broader development program, helping to define objectives, create and refine examples, identify failure modes and build evaluations. Thomson Reuters says its experts curated more than 11,000 internal evaluation items. Thousands of hours of lawyer time went into deciding not merely whether an answer sounded good, but whether it was complete, supported and useful to a professional whose work could not afford a confident error.Two details should stop every CEO mid-scroll.First, Thomson Reuters says it changed the root model many times as the open-weight frontier advanced. It did not pay to recreate the original pretraining behind each generation. It captured advances in the open-weight ecosystem and concentrated its own investment on specialization. That creates switchability, although not for free: every migration still demands new training, safety testing, evaluation and deployment work.Second, the company says it has used less than 10 percent of its legal content so far. CEO Steve Hasker put the opportunity plainly in an interview with Artificial Lawyer: "We've applied less than ten percent of our legal content to the Thomson model to date. And yet it is performing extraordinarily well."The Five Layers, Translated For The EnterpriseIn November, I defined national AI sovereignty across five layers: energy, hardware and compute, data, models and talent. No serious sovereign strategy requires owning every layer. It requires knowing where dependency is acceptable, where control is essential and how difficult it would be to switch suppliers.Thomson Reuters made those choices explicitly. It rented compute. It remained dependent on advanced accelerator hardware. It started from a model whose original training data it did not control. But it exercised much greater control over the layers that differentiate its business: proprietary content, expert judgment, evaluation standards, tools, post-training methods, deployment decisions and the resulting model weights.The company's own technical report makes an important admission that corporate boards should remember: sovereignty is a spectrum, not a threshold. Thomson Reuters has not eliminated dependency. It has reduced and rearranged it.That is the enterprise translation of the "third way" strategy I described for countries such as France and Singapore. A company does not need to match the capital expenditure of OpenAI, Anthropic or Google. It needs to preserve options at the layers becoming commoditized while protecting control over the assets that make the company distinctive.Hasker's proposed customer architecture extends that logic. Thomson Reuters is discussing placing Thomson inside a client's controlled cloud environment, connecting it to the client's document systems and updating the model through a one-way feed. That is more precise than saying the model is simply "behind the firewall," and it remains a potential offering rather than a broadly proven deployment. But the strategic idea is clear: give the customer greater control over where its knowledge lives, which model uses it and who benefits from it.Why Some Companies Should Consider ThisThree reasons, in descending order of urgency.Your data may be a moat—but only if you can turn it into one. Some insurers, asset managers, hospital systems and industrial companies possess decades of proprietary documents, decisions and expert judgment that never reached the open web. But possession is not enough. The material must be rights-cleared, cleaned, structured, deduplicated and connected to experts capable of grading the model's work. Medical records, client documents and employee communications may be sensitive or legally restricted. A disorganized archive is not a training advantage. It is a storage bill.Thomson Reuters' advantage is not merely that it owns a great deal of legal content. It has spent generations editing that content and employs roughly 1,500 attorney-editors whose daily work is distinguishing what is authoritative from what is merely plausible. The frontier labs are not technically incapable of working with proprietary data. What they cannot easily reproduce is the combination of a protected corpus, accumulated editorial decisions and professionals able to convert tacit judgment into training and evaluation signals.The economics have changed—but they have not disappeared. Training a frontier foundation model from scratch still requires extraordinary capital. Continual training of a strong open-weight model can cost dramatically less. The sub-$450,000 final GPU run demonstrates that point, but the more relevant number for a board is the approximately $40 million development program, plus the undisclosed acquisition and continuing maintenance costs.That is not a casual line item for most mid-cap companies. It is, however, within strategic range for some large and data-rich institutions—especially when a high-volume, high-value workflow can support the investment. The right comparison is not $40 million versus zero. It is $40 million versus years of frontier-model fees, slower performance, strategic dependency and the value of improving a core professional product.Dependency is a strategy, whether you chose it or not. A company that imports all its machine intelligence accepts someone else's architecture, roadmap, pricing, release schedule and usage rules. Building or controlling a specialized model can reduce that exposure. It can also reduce inference costs when utilization is high enough and keep sensitive client information away from frontier-model providers.But ownership does not abolish dependency. Thomson still relies on cloud infrastructure, GPU suppliers, an open-weight lineage and an extensive open-source software stack. CoCounsel also remains multi-model by design, using Thomson where it has an advantage and other leading models elsewhere. That may be the most mature form of sovereignty: not autarky, but the ability to choose.The Demanding PartDo not mistake this for advice to build a model next quarter. Thomson Reuters earned this outcome with ingredients most firms lack: genuinely differentiated and legally usable data, an acquired research team, hundreds of participating experts, rigorous evaluations and the willingness to operate and update the system for years.Before authorizing a proprietary model, a board should ask four questions:Do we possess a corpus competitors cannot readily obtain—and the right to use it?Do we have experts who can define and repeatedly measure what "correct" means?Is there a high-volume, high-value workflow where specialization could repay the investment?Are we prepared to maintain, reevaluate and migrate the model as the open-weight frontier changes?If the answer to any of those questions is no, retrieval over a frontier model, a smaller fine-tune or a multi-model architecture may be the better strategy.In November, I wrote that sovereignty is no longer only about borders. It is about who controls the intelligence that runs the nation. Substitute "company" for "nation" and the sentence still holds—but with an important qualification.Thomson Reuters has not proved that every company should build its own model. It has demonstrated something narrower and more consequential: for institutions possessing proprietary knowledge, expert evaluators and the discipline to maintain a model over time, dependence on frontier providers is becoming a strategic choice rather than a technical necessity.That is what corporate sovereign AI looks like—not independence from everyone, but control over what matters and the power to choose the rest.