Proprietary AI Model: A $40 Million Warning

A proprietary AI model is not a trophy for companies that want to look serious about AI. It is an expensive operating decision that only makes sense when a business owns something more valuable than the model: unique data, repeatable demand, and a workflow worth controlling.

That distinction matters now. Thomson Reuters announced that it spent $40 million to train “Thomson,” a model built from an open-source foundation and tuned for its professional markets. The company says it can run the model at a fraction of the cost of comparable frontier models. Most operators will hear that number and focus on the price. They should focus on why the price could be justified.

A proprietary AI model starts with an unfair advantage

The model is not the moat. Thomson Reuters already had the moat.

It owns decades of trusted legal, tax, accounting, and news content. It also has customers who use that information inside recurring, high-value workflows. Therefore, a better answer can affect research time, professional risk, and the value of an existing subscription.

That is very different from a SaaS company training a model because its cloud AI bill feels large. High spend alone does not make model ownership sensible. A proprietary AI model needs raw material that another company cannot easily buy. It also needs a narrow job where better performance creates measurable value.

If your training data is mostly public, your workflows are generic, and your customers can switch tools without friction, ownership gives you costs before it gives you leverage.

The real build-versus-buy test

Founders often treat build versus buy as a feature comparison. AI makes that framing dangerous. The decision affects capital, hiring, evaluation, security, deployment, and the speed at which the system improves.

I would ask five questions before funding a proprietary AI model.

1. Do we own data that changes the result?

More data is not enough. The data must improve performance on a job customers care about. It should also be lawful to use, clean enough to maintain, and difficult for a model vendor to reproduce.

2. Is the workload stable and frequent?

Ownership has a fixed-cost problem. Training, evaluation, and serving infrastructure arrive before the savings. Consequently, predictable inference volume matters more than an impressive demo. A model used across a core product every day has a chance to repay the investment. An occasional assistant probably does not.

3. Can we evaluate the model better than a vendor can?

Domain expertise only becomes useful when it produces a reliable test. You need to know which errors matter, which citations are acceptable, and when a plausible answer is still wrong. Without that discipline, a proprietary AI model can become a privately owned source of confident mistakes.

4. Does control improve the business?

Control has value when model behavior, data residency, latency, or unit economics affect the product. However, control for its own sake is overhead. Customers rarely reward infrastructure ownership directly. As I have argued before, your customers cannot see your tech stack. They can see whether the product works, whether it is trustworthy, and whether the price makes sense.

5. Can we support the asset for years?

The first training run is only the entry fee. Models need new evaluations, refreshed data, security reviews, deployment work, and people who understand failure modes. Meanwhile, outside models keep improving. The question is not whether your team can build once. It is whether ownership keeps producing an advantage after the announcement fades.

Renting intelligence is often the disciplined choice

There is no shame in buying model access. In fact, most companies should start there.

An API turns a large fixed investment into a variable cost. It lets a small team test demand before committing infrastructure and specialized talent. It also gives the company access to rapid model improvements without funding every step.

Still, renting creates exposure. Prices can change. Model behavior can shift. Usage limits can interrupt a workflow. Sensitive data may require additional controls. Moreover, a vendor will optimize for a broad market, not your exact failure cases.

The right response is not immediate ownership. Instead, design for replaceability. Keep your product logic outside the model. Build evaluations early. Record where the system fails. Route different tasks to different models when that improves cost or quality. Then let real usage show whether a proprietary AI model would create durable value.

This approach also prevents a common founder mistake: converting anxiety about dependency into an engineering project. Dependency is a business risk to manage. It is not automatically a reason to build everything yourself.

The $40 million number is a warning, not a benchmark

According to the Thomson Reuters announcement, the company began with an open-source foundation and invested $40 million across talent and compute. That figure sounds small beside frontier-lab budgets. For almost every SaaS company, it is still enormous.

More importantly, the investment sits on top of assets built over decades. The content already existed. The customer relationships already existed. The workflows already existed. The distribution already existed. The model connects those advantages. It did not create them.

That is the part founders should copy.

Before building core AI infrastructure, identify the proprietary advantage it will amplify. If you cannot name that advantage clearly, keep renting. Improve the workflow. Collect better feedback. Learn which errors cost customers money. Build the evaluation set that nobody else has.

Then revisit ownership when the evidence becomes uncomfortable to ignore.

Own the advantage before you own the model

A proprietary AI model can make economic sense. However, the winning sequence is easy to reverse.

First, own a valuable problem. Next, own differentiated data and a reliable way to measure results. Finally, consider owning the model when scale, control, and cost all point in the same direction.

Most companies do not need a $40 million model strategy. They need a sharper answer to a simpler question: what do we know about our customers’ work that a general model provider never will?

If the answer is weak, buying is discipline. If the answer is strong and the usage is real, ownership may become a business advantage instead of an expensive identity project.