Download the Model. Become the Provider.

Most of us first met generative AI through a chat window.

The experience felt like software. Type something. Get something back. Move on.

Then the useful work started accumulating. A few prompts became research, writing, analysis, code, customer support, and workflows that other people began to depend on. That was when the meter became visible: token limits, subscriptions, usage tiers, and the uncomfortable realization that the tool was not simply installed software. It was a service being consumed.

Open-weight models appear to offer a clean escape. Download the model. Run it yourself. Stop paying somebody else for every answer.

That promise contains something real. It also hides the most important part of the decision.

You have not downloaded the service. You have downloaded the trained model at its core.

You received the engine

A hosted AI service includes far more than a model.

Inside the model are a tokenizer, an architecture, and billions of learned settings called weights. The tokenizer breaks a request into pieces the model can process. The architecture defines how information moves. The weights shape what the model predicts next.

Those pieces form the trained model. They do not produce a dependable business outcome by themselves.

The model still needs an inference system and computing hardware. It needs behavior and safety controls. It needs an interface, security, monitoring, scaling, updates, support, and people who know how to keep the entire system working.

A hosted provider assembles and operates those pieces, then sells access to the finished result.

The simplest analogy is a car service. You ask for a ride. The company supplies the vehicle, driver, maintenance, insurance, routing, and operations required to get you there.

When I first heard that I could download a model and run it myself, I thought I was getting the car.

What I received was the engine.

It was a valuable engine—and one I could possess and control. But the experience around it was now my responsibility.

“Open” answers more than one question

This is where the language starts causing trouble.

Open weights, an open-source license, and Open Source AI are often treated as interchangeable claims. They are not.

Open weights describe access to the model's learned parameters. Depending on the package and its terms, that access may let you choose where the model runs, test it, fine-tune it, add controls, and build a product around it.

The license answers a narrower legal question: what may you do with the files you received?

The Open Source Initiative's Open Source AI Definition asks a broader question. Do you have the freedoms and materials needed to use, study, modify, and share the complete system? OSI's own explanation of open weights versus Open Source AI makes the boundary explicit.

Possessing the weights does not necessarily tell you what the model learned from or how it was taught. If the model produces an unexpected result, the weights may not reveal which training material caused it. They may not give you what you need to recreate the model either.

That does not make open weights unimportant. It makes the claim specific.

Specific claims are useful. “Open” as a vague moral label is not.

Free files do not create free outcomes

The model builder has already paid for a major body of work: research, training, testing, and refinement. Downloading the resulting model does not require you to repeat that entire program.

Everything required to operate it still has to come from somewhere.

Hardware. Inference software. Electricity. Cooling. Security. Monitoring. Updates. Reliability. Governance. People.

The cost did not disappear. It changed location.

A hosted service bundles more of those responsibilities into a subscription or usage charge. A self-operated model moves more of them into infrastructure, engineering, security, and operations carried by the organization using it.

That shift can be worth making. Local operation may create meaningful advantages for sensitive data, customization, continuity, performance, or control over provider dependence. Those advantages are not free merely because they are valuable.

Ben Thompson made the economic point directly in “Who's Afraid of Chinese Models?”: open-weight models are not free to serve. The model artifact may be downloadable. Producing answers still consumes resources.

The honest comparison is not free versus paid. It is one operating model versus another.

Tokens measure activity, not value

Token prices look like a common unit for that comparison. They are not.

Different tokenizers can split the same request differently. Different models may require different amounts of computation. One model may produce an acceptable result on the first try while another creates retries, review, delay, or rework.

A token can meter usage. It cannot tell you whether the result was useful.

The enterprise measure that matters is total cost per acceptable outcome.

That includes the visible price of inference, but it also includes the system and human work required to produce something the organization can actually use. A cheaper token that causes three retries and a manual rescue may be more expensive than a higher-priced result that works the first time.

This is also why one permanent model strategy is unlikely to fit every workload. Some work belongs with a hosted provider. Some justifies a self-operated model. Some should not use AI at all.

The practical question is not simply whether a model is open or closed.

What outcome do I need—and where do I want its costs and responsibilities to live?

That is the decision open weights make possible. They give an organization another way to carry the engine.

They do not finish the car.

Sources

• Open Source Initiative — Open Source AI Definition 1.0
• Open Source Initiative — Open Weights vs. Open Source AI
• Ben Thompson — “Who's Afraid of Chinese Models?”

Robb Boyd spent nearly two decades at Cisco as Managing Editor of TechWiseTV—the company's highest-ROI marketing asset, reaching audiences in 65+ countries. Today he spends as much time thinking about how AI is changing the way organizations trust their own people as he does producing video for them—this piece is part of that thinking, done in public.

Want more analysis like this? Subscribe to ExplaiNerds. And if you're a marketing or content leader with a story that deserves a bigger audience—let's talk.