"Open AI" is often used as a single label for very different realities.
In software, open source had a relatively clear meaning: source code under a recognized license, inspectable and reusable with known rights. In AI, the word "open" now covers a spectrum. This ambiguity creates bad strategy and bad procurement.
When evaluating an "open" AI offering, separate the layers:
- Are weights available?
- Is training code available?
- Is training data available?
- Are evaluations transparent and reproducible?
- Is commercial usage unrestricted?
- Can the model be self-hosted?
- Can behavior be audited and version-pinned?
Many models are open weights but closed data. Others are "open for research" with restrictive commercial terms. Some are technically downloadable but practically inaccessible at scale due to deployment complexity.
None of this means open is fake. It means open is often partial.
Why this matters:
- Strategic clarity: founders need to know which part of the stack they are opening and why.
- Buyer protection: enterprise teams need to understand what can be inspected, moved, and governed.
- Honest competition: comparing "open" and "closed" products without layer-level definitions is mostly noise.
A useful operating rule: do not ask "is it open?" Ask "what exactly is open, what is portable, and what remains controlled by one vendor?"
This framing improves decisions. It prevents teams from overestimating sovereignty and underestimating lock-in. It also avoids lazy dismissals of open ecosystems that are effective even when not fully open end-to-end.
The next step is evaluating the case for closed systems on its own terms, because the closed argument is often stronger than open advocates admit.
This is part 2 of 10 in Open vs Closed AI.