Open-source AI vs proprietary labs: why one isn't killing the other, explained

For the past two years, a confident prediction has circulated in the technology industry: as capable open-source AI models proliferate and become free to download, they will hollow out the business of the closed labs that charge for access to their systems. A recent analysis by TechCrunch pushes back on that narrative, arguing that so far the rise of open models is not coming at the expense of frontier labs, and that the two may actually capture different phases of the same life cycle.

To understand the argument, it helps to define the two camps. Proprietary or closed models are developed by companies that keep the underlying system private and sell access, usually through an interface that lets other software send requests and receive responses. Open-source or open-weight models, by contrast, are released for anyone to download, run and modify, often at little or no cost.

The intuitive case for open models displacing closed ones is straightforward. If a free model is good enough for a task, why pay for a proprietary one? That logic has held in many parts of software history, where open alternatives eventually commoditised once-premium products. Applied to AI, it suggested the closed labs were selling something that would soon be available for nothing.

What that reasoning misses, the analysis suggests, is that open and closed models tend to be used at different stages of a project's life. Frontier proprietary models are often where teams start, because they offer the highest capability and the least friction: no infrastructure to manage, immediate access to the most advanced systems, and rapid iteration. That is the exploration phase, where capability matters more than cost.

Open models, meanwhile, come into their own later. Once a company knows exactly what it needs, an open model that it can host itself, fine-tune on its own data and run at predictable cost becomes attractive. This is the optimisation phase, where control, privacy and unit economics matter more than raw frontier capability, and where a cheaper self-run model often wins.

Seen this way, the two are less rivals than complements. A team might prototype on a top-tier proprietary model, prove that an idea works, and then move a mature, high-volume workload onto an open model to cut costs, while still returning to the frontier for the next hard problem. The same organisation can be a paying customer of a closed lab and a heavy user of open models at once.

There is also a moving target that protects the frontier labs. The most advanced proprietary systems keep improving, so the capability gap between the newest closed models and the best open ones tends to persist even as open models catch up to where the frontier was a year earlier. For customers who need the most capable system available today, that gap is exactly what they are paying for.

None of this means open models are not a competitive force. They exert real pressure on pricing, push closed labs to justify their premium, and give companies leverage and alternatives they did not have before. The analysis frames the situation as "not hurting yet", a deliberate qualifier that acknowledges the dynamic could shift if open models close the capability gap faster than the frontier advances.

The economics underneath explain why the labs are not panicking. Building a frontier model requires enormous investment in computing power and research talent, costs that open releases do not eliminate but redistribute. Someone still has to fund the training of the most capable systems, and the closed labs argue that access revenue is what makes that continued investment possible, which in turn produces the advances that open models later build upon.

For businesses choosing between the two, the practical lesson is that it need not be a binary decision. The most sophisticated AI adopters increasingly use both, matching the tool to the phase: proprietary models for cutting-edge capability and fast experimentation, open models for control and cost at scale. Rather than one killing the other, the market appears to be settling into a division of labour, and for now, at least, both sides are growing.