- VentureBeat's column says open-weight models suit narrow jobs such as classification, and that the vendor's safety work becomes your team's job.
- a16z's 2025 survey found 37% of enterprises use five or more models, and agentic workflows make switching harder.
- Before moving a workload, ask who tests safety after fine-tuning, and whether your prompts can move to another model.
Cheaper AI models look like a price story. For the people who sign the contract, they are really a story about who holds the risk when something goes wrong.
A VentureBeat column published on October 31 makes the case. It says inference spend, the money paid to run models, has been gradually shifting toward open-weight models. It also says proprietary models from OpenAI and Anthropic still dominate enterprise use. The lesson here is that open weights do not remove dependence. They change where it sits.
Multiple models are now ordinary
The column leans on Andreessen Horowitz (a16z), which published its enterprise AI survey in 2025. The firm drew on its earlier interviews with more than twenty-four enterprise buyers. It also polled 100 CIOs spread across 15 industries.
In that survey, 37% of respondents were using five or more models, up from 29% a year earlier. a16z says the main reason is that different models suit different jobs. Avoiding vendor lock-in is a secondary reason.
a16z also reports that open source adoption was higher at larger enterprises, where running software on company premises is still a major consideration. Most enterprises, it says, mainly choose open source models for highly cost-sensitive uses. The survey is from 2025, and a16z is an investor, so treat it as one view of buyer behaviour.
The cost story runs both ways. a16z also found that closed small and medium models have become much more attractive on price for performance. Some customers now choose closed models for that reason. One customer put it in three words: "Gemini is cheap."
Open weight is not open source
The column draws a line many buyers miss. With an open-weight model, you typically get the trained numbers that make the model work, called weights. The training data, the code and a full licence usually stay with the maker.
That matters because of who is responsible. The column says a closed model arrives with the vendor's safety work already done. Choose an open model, and the safety testing falls to your own staff.
The column adds that this protection is easy to remove. It cites researchers who showed that a handful of fine-tuning examples can reverse a model's safety training. It also says Llama and Gemma guardrails can reportedly be removed within minutes. The column does not say to avoid open weights. It says companies should know the risk they are taking on.
Where the savings are real
The column sees open weights paying off on narrow, well-defined jobs. Its examples are embeddings, classification and spotting personal data. An embedding model turns text into numbers so software can search by meaning.
The author argues a purpose-built model can often beat a frontier model on cost, at least in compute. The reasoning is simple. A large general model charges you for flexibility "whether you use it or not". A tuned small model skips that overhead.
a16z notes that prices for models in general fall about tenfold every 12 months. That claim covers all models, not open weights alone. The sources do not say how the price gap between open and closed models will change.
Running it yourself is the expensive part
Downloading a model from Hugging Face and running it with a tool like Ollama is easy, the column says. Matching a hosted frontier model's efficiency and reliability is harder. You need specialists who can run GPUs, tune the infrastructure and keep performance steady under real traffic.
The author compares this to open source software. Unless a function is central to the business, the column says, most organisations buy a managed service instead of maintaining the code. For most enterprises, the column says, the realistic route is an inference provider such as Fireworks or Together AI.
Lock-in moves into the workflow
The column warns against getting locked into one model provider. a16z's survey shows why that is harder to avoid than it sounds. In an earlier a16z survey, many enterprises treated models as interchangeable. Agentic workflows, where a model completes a task in many steps, have made switching more difficult.
One leader told a16z: "all the prompts have been tuned for OpenAI." That example involves a closed model. Changing models can take a lot of engineering time, a16z says, because one change can affect every later step. a16z also says prompts port more easily than fine-tuned models.
a16z does not say open-weight users face this lock-in. This is the publication's own argument. Owning the model does not free you if your instructions, tests and guardrails only work on that model. The same dependence on prompts and testing would apply to any model, open or closed.
Questions to put to your team
Ask which workloads are narrow enough to move, such as classification or personal-data detection. Ask who tests safety after any fine-tuning, and how often. Ask whether your prompts and quality checks can run on a second model without a rewrite.
Ask what a managed provider would cost compared with the experts and GPUs needed to run the model in-house. Ask where your data goes under each option, since that was a core reason the column gives for choosing open weights.
Control is a real benefit. It is also a job someone must be named to do.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Andreessen Horowitz (a16z).
