Today’s radar revolved around money and data: Nvidia backing an open model to rival China, Hugging Face reportedly valued at US$13 billion, and Thomson Reuters showing what it costs — and what it is worth — to build a proprietary model. I picked out what matters for people who work in product.

A question comes up often when I work with credit products: is it better to build something in-house or buy from someone who already has scale? There is no easy answer, but every so often a case appears that helps us think about it more clearly.

In brief

  • Thomson Reuters launched Thomson, a proprietary AI model trained on an open-source base combined with the legal and tax data the company already owned.
  • The model comes close to frontier models in performance and cost roughly US$40 million to develop — a fraction of what the major labs spend.
  • The advantage is not only the architecture: it is the combination of available technology with decades of proprietary data in a specific domain.
  • For vertical products, the build-versus-buy decision needs to account for the data asset that only the company itself can provide.

The proprietary model that starts with data

Thomson Reuters has just launched Thomson, its own AI model, trained on an open-source base combined with the legal and tax data the company already had. The result comes close to frontier models in performance, but cost roughly US$40 million to develop. That is a fraction of what the major labs spend.

What stands out most is not the model itself. It is the strategy behind it.

Instead of trying to compete head-on with companies that have billions to spend on compute, Thomson Reuters bet on what only it has: decades of proprietary data from a specific domain. That is extremely valuable when paired with the right base.

The Thomson Reuters announcement about the model launch makes the logic behind the bet easier to see.

Build or buy for vertical products

I draw a direct parallel with my day-to-day work in credit product. Receivables, invoices, payment history, and each borrower’s behavior over time. That kind of data is unique to each institution, built year after year, and it is exactly the kind of asset that becomes a real competitive advantage when used well with AI.

This is a central decision in AI product management: evaluate available technology alongside data, operating cost and the ability to turn information into outcomes. For PMs, AI for Product Managers helps structure that analysis without reducing the decision to a model benchmark.

For people building products in vertical sectors—finance, healthcare, legal services or any field with accumulated proprietary data—the message is clear. The AI race is not only about who has the most capable model on the market. It is about who knows how to combine available technology with what only the company itself has to offer.

That makes me optimistic about this moment. It shows that it is possible to build something competitive without relying only on the deepest pockets, as long as you know where the real differentiator is.

The rest of the radar

Nvidia bets US$6B on an open model to rival China — stronger open-weight options reduce vendor lock-in and put pressure on the price of closed APIs. Read more

Hugging Face may be sold for US$13B — if the central platform of the open-source ecosystem changes hands, it affects roadmaps, model hosting and the confidence PMs rely on when prototyping. Read more

Headlong: agents that think all the time, not only when they reply — it proposes a different paradigm for human-agent interaction through chat or Slack, beyond the standard prompt-response pattern. It is a useful example of the kind of system discussed in the guide to AI agents. Read more

Open Qwen3.8-27B model solves reverse engineering in 30 minutes — shows that complex technical tasks are no longer exclusive to closed frontier models. Read more

Google cuts Gemini Spark from US$100 to US$20 per month — a direct pricing reference for anyone defining tiered plans for agentic features. Read more

Anthropic launches Claude Cowork, a US$20/month desktop agent — another case of the desktop productivity agent becoming a category, with a low entry price. Read more

OpenAI tests ads in free ChatGPT and retires Custom GPTs — a sign of a possible shift from subscriptions to advertising in a mass-market AI product, and a reminder that platform features disappear when they do not support the business. Read more

Margin pressure is redefining AI pricing in 2026 — 70% of product leaders say delivery costs erode profitability; hybrid charging models have risen from 20% to 25% in 18 months. Read more

That is all for today’s radar. See you in the next edition.