Open Source vs Proprietary AI: What Actually Differs

Most models marketed as open source are open weights, not open source — you can download and run them, but under a bespoke licence with use restrictions rather than one that meets the Open Source Definition. The Open Source Initiative published an Open Source AI Definition precisely because the term had drifted this far. Getting this straight matters, because the business decision underneath is not really about licensing at all.

Read the actual licence of any model you plan to deploy. Terms vary by model and change between releases; nothing below substitutes for that, or for your own counsel.

The two decisions hiding under one label

Where does inference run? In a vendor's cloud, or on infrastructure you control. This is the decision that drives data residency, latency, unit economics at volume, and whether you can run at all in an air-gapped environment.

Who controls the model's future? A hosted model can be deprecated, re-priced or changed underneath you. A downloaded set of weights sitting on your own storage cannot.

Almost everything people actually want when they say "open source AI" is one of those two things. Neither requires an OSI-approved licence, which is why the licensing argument generates so much more heat than clarity.

What open weights genuinely buy

Deployment inside your own boundary, which for some regulated workloads is the difference between viable and not. Version stability — the model you validated is the model you keep. Predictable cost at high, steady volume, since you are paying for compute rather than per token. And the ability to inspect and adapt, though far fewer organisations use this than claim they will.

What they genuinely cost

Someone has to run it. Serving infrastructure, capacity planning, upgrades, monitoring — the vendor was doing this and now you are. Evaluation becomes yours too: no vendor is testing each release for you, so you need the harness described in the AI vendor evaluation framework. Safety tooling that arrives bundled with hosted APIs has to be assembled. And the capability gap at the very frontier is real, though it narrows and reopens with each release cycle — Stanford HAI's AI Index tracks that movement year over year.

How to actually decide

Start with data residency, because it is the only genuinely binding constraint. If your data cannot leave your boundary, self-hosting is not a preference. If it can, you have a real choice.

Then look at volume. Self-hosting has high fixed cost and low marginal cost; hosted APIs are the reverse. There is a crossover, it is specific to your workload, and it is a calculation you can do — the method is in AI total cost of ownership.

Then be honest about capability required. Many production tasks — classification, extraction, routing, summarising your own documents — do not need frontier performance, and this is where open weights are most often the correct commercial answer.

Finally, be honest about the team. Self-hosting is an operational commitment, not a one-time deployment. Companies without anyone who wants to own that should buy hosted and use the saved attention elsewhere.

The answer most companies land on

Hosted APIs for frontier work where capability matters and volume is spiky. Open weights for high-volume, well-defined, latency-sensitive or residency-constrained tasks. A thin internal interface so either can be swapped — which is the lock-in posture argued in AI vendor lock-in and the middle path in build vs buy AI.

"Open source therefore free" is the assumption to kill early. Compute, operations and evaluation are the cost; the licence was never the expensive part.

Frequently asked questions

Is open source AI actually open source? Usually not. Most popular models are open weights under bespoke licences with use restrictions, rather than licences meeting the Open Source Definition. Read each model's actual licence before deploying it.

What do open weights really give you? Deployment inside your own boundary, version stability, predictable economics at high steady volume, and the ability to inspect and adapt. What they cost is operations, evaluation and safety tooling you now own.

Which is cheaper? It depends on volume. Self-hosting is high fixed cost and low marginal cost; hosted APIs are the opposite. Calculate the crossover for your own workload rather than trusting a general claim.

What do most companies end up doing? A mix: hosted APIs for frontier work with spiky volume, open weights for high-volume or residency-constrained tasks, and a thin internal interface so either can be replaced.

Further reading

• AI vendor lock-in
• Build vs buy AI
• AI total cost of ownership
• The AI vendor evaluation framework
• AI strategy advisory