The conversation about AI in your company today is probably stuck on 'do we pay for ChatGPT from OpenAI or Gemini from Google?'. That is a small decision inside a much bigger decision that just opened up in 2026 and almost nobody is noticing: you can have a proprietary AI agent, trained on your company's data, running the way you want, without depending on an American giant charging a monthly fee per user. This is not theory. It is a product. And the tooling to make it happen is open, free, and ready to use. This post explains what changed, why it matters for mid-size and large businesses, and what the practical path is to get started.
The thing is called Nemotron 3 Ultra. It is an open AI model, released by Nvidia, that delivers quality comparable to ChatGPT and Gemini, but with a difference that changes everything: you take the model, train it on your company's data, and it becomes yours. Nobody outside sees what it learned. Nobody charges a monthly fee. Nobody can change the usage rules from one day to the next.
1. The difference between 'using someone else's AI' and 'having your own AI'
When your company pays for ChatGPT or Gemini, it is renting capacity. It works, but it has invisible costs that keep adding up:
- A monthly fee per user, every month, forever. A company with 200 employees pays more than a small one. A company that uses it a lot pays more than one that uses it little. The cost grows with the success of the adoption, it does not go down.
- Your company's data passing through the American vendor. A question with a client's name, a contract number, a proposal amount. All of it leaves your control the moment you send the message. In some sectors this is illegal. In others, it is just uncomfortable.
- The vendor changes the rules whenever it wants. Raises the price, removes a feature, changes the terms of use. This has already happened several times in 2024 and 2025. Whoever depends on it, suffers.
- The AI does not learn about YOUR business. It is good in general, but it does not know who your 50 main clients are, what your internal process is, what the rule is that only applies in your market.
Your own AI solves all four. Predictable fixed cost, data kept in house, rules that you control, and most importantly: the agent understands YOUR business because it was trained on it. It is not a giant for everyone. For a mid-size or large company, with enough volume, it comes out cheaper and more useful.
2. The example that opens your eyes: chip design
Nvidia showed a real case with Cadence (a company that makes software for designing chips). They took Nemotron, trained it on all of Cadence's proprietary knowledge (internal manuals, solved cases, in-house standards), and created a specialist agent that nobody else in the world has. That agent started verifying chip designs 40 times faster than the previous method, and with superior quality, because it was trained on the exact peculiarities of Cadence's work.
Notice the sequence: they took an open model (free), added their own knowledge (which nobody else has access to), and created a proprietary competitive advantage. This same pattern applies to any mid-size company and up that has accumulated internal knowledge.
A law firm has briefs, internal case law, petition templates refined over decades. An engineering firm has projects, standards, technical decisions on record. A healthcare company has clinical protocols, treated cases, diagnostic patterns. An industrial company has operating manuals, technical data sheets, equipment failure history. That knowledge, trained into your own agent, becomes an asset. It keeps its value even if the employee leaves, even as time passes, even as the company grows.
3. What an 'open model' is and why it matters
An AI model can be delivered in two ways:
- Closed model: ChatGPT, Gemini, Claude. You only use it over the internet, you pay a monthly fee, you have no access to the 'core' of the model. It works well, but you control nothing.
- Open model: Nemotron, Llama (from Meta), DeepSeek, Mistral. You download it, install it on your server (or on your PC, if it is smaller), modify it, train it on your data, control it completely.
Until 2024, open models were inferior. The quality gap did not justify the work of installing them. In 2025 and 2026 that changed. The best open models today are at 95% of the quality of the closed ones, in some tasks they are ahead, and they cost much less to run. For a company that has volume and proprietary data, the numbers now add up.
Nemotron is particularly interesting because Nvidia delivers not only the model, but also the dataset used for training and the step-by-step of the training. It means you can take everything, add your own data, retrain, and end up with a model that is better than the original for YOUR case. This is not a favor. It is how Nvidia chose to play in order to win volume against the closed competitors.
4. Who should consider it, and who should not (yet)
Your own model is not for everyone. The rough rule:
- Consider it: a company with 50 or more active AI users per day, with relevant proprietary data (years of accumulated internal knowledge), with technology people capable of running it or a hired technical partner to drive it.
- Wait a bit longer: a small company that does not have the volume or the data, and that is using AI only on the surface (answering email, summarizing text). For those, someone else's model is cheaper and simpler.
The expensive mistake is to keep paying for ChatGPT for 300 employees for years, without ever evaluating that the same money invested once in your own model would buy three years of independent operation. That calculation needs to be done, with a serious vendor, based on YOUR real usage, not on what shows up in a magazine.
5. The practical path for the owner who got curious
Three steps, in order:
- Measure current usage. How many employees use AI today, how many times per day, for what kind of task. Without this, any cost comparison is a guess.
- Identify the usable proprietary knowledge. What does your company know that others do not? Is it written down somewhere? Is it worth consolidating to train an agent?
- Ask two or three serious technical partners for a proposal, with the implementation cost AND the annual operating cost, compared side by side with the 'keep paying OpenAI' scenario. The decision becomes obvious pretty quickly. One way or the other, but obvious.
The important reframe: AI has stopped being an off-the-shelf product and has become an internal capability of the company. The same way a serious company has its own systems (ERP, CRM, finance), within a few years it will have its own agent. Whoever starts thinking in that direction now arrives well positioned in 2027. Whoever leaves it for 'when I need it' arrives late, pays a lot, and on top of that competes with those who are already up and running.