In computing, whenever a cheaper, more open alternative shows up, it’s often framed as a winner-takes-all fight to the finish. We saw it with Linux and Windows, and again with Android and iOS. In reality, both sides “won” those battles because they serve different markets, and seeing it as “either-or” was never the right lens to begin with.
A similar pattern is playing out with AI models. While proprietary models from OpenAI and Anthropic dominate enterprise use, inference token spend has been gradually shifting to open weight models. Andreessen Horowitz has reported that 37% of enterprises were running five or more models in production, including open and closed.
The real risk to a business isn't picking the wrong side, it’s getting so locked into a single model provider that switching later becomes too costly or impractical.
What companies actually want is control
Open-weight models are closing the performance gap with proprietary ones, but that’s only part of what’s driving adoption. Much of the enthusiasm is really about control -- over where data goes, how the model behaves, and a business’s exposure level if a vendor disappears or changes its terms.
Running a model on-prem or in a company's own cloud makes AI viable in regulated industries like healthcare, legal, and financial services, where sending data to an external API is often not an option. Post-training an open-weight model on a company's own data keeps that customization in-house, shaping how the model behaves without handing that process to a vendor. And owning the weights means a business isn't left stranded if a vendor disappears, gets acquired, or changes its terms. The model can't be taken away.
Control cuts both ways
But control comes at a cost; it puts more responsibility on the end user. Open-weight isn't the same as open source. Typically, you get the weights but not the training data, the code, or a real license. That matters because it determines who's on the hook if something breaks.
With a closed model, the vendor has done the important safety work. With an open model, that work is your team's responsibility, and it’s easier to undo than most people realize.
Researchers have shown that a model's safety training can be reversed with a small number of fine-tuning examples, and the guardrails on models like Llama and Gemma can reportedly be stripped away in minutes. That's not a reason to avoid open-weight models, but it's a risk companies need to know they're taking on.
Where open-weight models earn their keep
The clearest place I see open-weight models paying off is for narrow, well-defined workloads like embeddings, classification, and detecting personally identifiable information (PII). A purpose-built model tuned for one of those jobs, an embedding model or a classifier, can often outperform a frontier model for significantly lower cost, at least in terms of compute.
That’s because a large frontier model can handle a wider range of tasks, but you pay for that flexibility whether you use it or not. A model tuned for a narrow task avoids that overhead.
Who should actually run them
This question has parallels with open source software, where the code is superficially "free" but running and maintaining it is not. With software, most organizations pay for a managed service or a packaged distribution, because it's rarely worth the overhead of maintaining it yourself unless the function is core to the business.
Open-weight models follow the same logic. Pulling a model off Hugging Face and running it through a service like Ollama is easy. Matching the efficiency of a hosted frontier model, with the reliability your application needs, is harder. It requires experts who can manage GPUs, tune infrastructure, and keep the model performant under production load.
For most enterprises, the more realistic path is going through an inference provider like Fireworks, Together AI, or Base10, which specialize in running and scaling open-weight models. This provides the cost and control benefits without the ops burden. The exception is self-hosting when AI is core to the business and you have the scale and expertise to justify it. At a certain scale, the operational investment pays for itself.
Where frontier models still hold their ground
With frontier models, you're paying for a near-guaranteed level of performance without having to build it yourself. Open-weight models can get you similar reasoning capability for less, but the cost shows up elsewhere, in latency or in the work of getting the model to perform at scale.
Stanford's AI Index found that the gap between the best open and closed models actually widened slightly this year, which is worth remembering the next time someone claims open has basically caught up. Which trade-off makes sense depends on the workload. A customer-facing application often can't absorb the latency. An internal task like summarizing customer support tickets usually can.
The value moves up a layer
At some point, operating systems became a commodity; no one makes real money selling an operating system any more. That commoditization of the OS layer is partly what made the cloud possible, which in turn has driven much innovation since. The value didn’t disappear; it moved up a layer.
We may see something similar happen with models. As open-weight models close the gap on workloads that matter, "which model wins" stops being the interesting question. The real competition moves to what gets built on top -- the applications, the data, and the context that make a model useful.
That's the layer where the open-weight decision actually matters, not in the abstract debate over which type of model is "better." Open-weight models are a valuable part of a diversified model strategy, but you need to do the work of figuring out where they actually work for you.
Sean Falconer is VP of Product, AI Products and Strategy at Confluent
Welcome to the VentureBeat community!
Our guest posting program is where technical experts share insights and provide neutral, non-vested deep dives on AI, data infrastructure, cybersecurity and other cutting-edge technologies shaping the future of enterprise.
Read more from our guest post program — and check out our guidelines if you’re interested in contributing an article of your own!