Since generative AI became a business tool, the default arrangement has been simple: a vendor runs the model on its own servers, and your business sends a request and pays for the response, usually by the token. That's still how most companies use AI, for good reason — no hardware, no maintenance, access to whichever model is currently most capable. What's changed is that it's no longer the only serious option. Several labs building the best models, including OpenAI, Meta, and the teams behind Alibaba's Qwen and DeepSeek, now also publish open-weight versions: the model file itself, downloadable and licensed for a business to run on infrastructure it controls, not only through the vendor's own API. OpenAI publishing its own downloadable models in 2025 was a signal of how far this has moved from a hobbyist niche — a company built on API-only access decided the open-weight option was worth offering too.
It's worth being precise about what actually changes when a business runs a model itself instead of calling someone else's API. It isn't about which company's logo is on the technology; it's about where the computation happens and who controls the machine it runs on. An API call sends your prompt — a customer's message, a contract, an internal document — to a server the vendor operates. A self-hosted open-weight model runs on a server your business rents or owns, and the data involved never has to leave that boundary. Nothing about the arrangement makes the model smarter. It changes who has custody of the data going in, and what the cost looks like coming out.
The cost shift is the more visible one. A metered API bill scales with usage in a way that's easy to underestimate until a workflow moves from occasional to constant use — a support inbox running every message through a classifier, a document pipeline summarizing every contract that arrives. Each of those is a recurring expense with no ceiling other than usage. Running an open-weight model on infrastructure you control turns that into something closer to a fixed cost, largely the same whether it handles ten requests a day or ten thousand. For light use, that trade doesn't favor self-hosting. For a workload that has become genuinely high-volume and steady, the economics look different, and it's worth actually running the comparison rather than assuming the metered bill is cheaper because it was the only option when the workflow started.
The data-control shift matters even where the cost math doesn't clearly favor either side. Sending a customer's financial details, a patient's information, or a client's confidential contract to a third party's servers means that data's protection now depends on that vendor's security practices, breach history, and legal exposure to disclosure in whatever jurisdiction its servers sit in — regardless of how the contract is worded. That isn't a reason to avoid API-based AI broadly; most business use of these tools involves nothing sensitive enough to justify running your own infrastructure. It's a reason to treat the decision differently for the narrower set of workflows that touch data your business has a real obligation, contractual or regulatory, to keep inside its own boundary. For that work, a model you actually control is a more defensible answer to a question a regulator or a client's security review will eventually ask.
None of this makes self-hosting free in a different sense than the API bill was free. Someone still has to provision the hardware, keep the model and its serving software patched, monitor it, and handle the outage at an inconvenient hour — work a metered API vendor was quietly absorbing as part of the subscription. And open-weight models, while considerably more capable than a year or two ago, still generally trail the very best proprietary frontier models on the hardest reasoning and coding tasks. This is a trade of some raw capability and operational simplicity for cost predictability and data control, worth making deliberately rather than by default.
The practical way to decide is three plain questions, asked per workflow rather than once for the whole business. Does this task touch data your business has a real reason to keep off someone else's servers — by contract, regulation, or ordinary prudence? If yes, self-hosting deserves serious consideration even at a higher upfront cost. Is the volume high and steady enough that the metered bill has become a real, forecastable line item? If yes, run the actual comparison instead of assuming the API is cheaper because it always has been. And does the task genuinely need the single best available model, or a good, consistent answer at a known cost? Most recurring business automation — classification, summarization, routine drafting, internal search — falls into the second category, where a capable open-weight model is plenty.
Most businesses that work through these questions honestly end up with a hybrid: a capable open-weight model running on infrastructure they control for the bulk of routine or sensitive work, with an occasional call out to a frontier API for the harder task that genuinely justifies it. That's not a reason for a business running fine on a vendor's API today to rush out and buy servers. It's a reason to treat where a model runs, and who has custody of the data going into it, as a deliberate decision worth revisiting on its own schedule — rather than a default locked in by whichever tool happened to be easiest to try first.
- open-weight models
- ai infrastructure
- data privacy
- ai costs
- self-hosted ai