When startups should stop renting AI

Open models can lower costs and keep sensitive data closer. They also turn an API decision into an infrastructure function.

A founder I know tried to run DeepSeek locally on a Mac Studio. The model worked. So did the heating. What began as an experiment in AI independence quickly became a lesson in memory limits, model sizes and the infrastructure work that disappears behind an API call.

The appeal is obvious.

API bills can grow with usage. Providers can change prices, rate limits, available models or product policies. Some companies need sensitive data to remain inside their own cloud or physical environment. A model they can deploy and adapt gives them options that a closed service cannot.

The language is messier than the marketing. Many models described as open source are really open weight. You can download and run the trained parameters, but you may not receive the full training code, detailed data information or unrestricted rights to use and modify everything. The Open Source Initiative makes that distinction clear. Open weights give you access to a model. Open source should give you the information and freedoms required to study, modify and share the system.

That may sound like licensing trivia until a startup builds a product around a model and later learns what the licence does not allow.

Privacy is another reason founders consider bringing AI closer.

Self hosting can reduce the amount of data sent to an external provider and may help with residency, internal policy or customer requirements. It does not create privacy by itself. The company now owns access controls, logging, patching, security testing, incident response and every careless copy of production data that finds its way into a development environment. Running the model yourself transfers responsibility. It does not make the risks disappear.

The opposite claim is also too simple. Renting a model does not automatically mean the provider trains on your business data. OpenAI says its API and business products do not use customer data for training by default. API inputs and outputs may be retained for up to 30 days for most eligible services, while qualifying customers can request zero data retention for certain endpoints. The contract, retention settings and architecture matter more than the slogan.

Cost has the same problem. An API looks expensive because the bill arrives with a visible price for every token. Self hosting spreads the cost across GPUs, idle capacity, inference software, monitoring, evaluation, engineers and upgrades. At high and predictable volume, an open model can be cheaper. With low or uneven demand, the rented service may still win because you pay only when the model is used. Production deployment also involves model loading, GPU memory, scaling, workload distribution and reliability. The model download is the beginning of the work.

Then there is quality. Frontier providers keep improving their models. A small team can spend months tuning and operating a cheaper system, only to find that a new API release performs better by Tuesday. When the model itself is not central to the product, this can become an impressive piece of engineering that customers never asked for.

The case for ownership becomes stronger when the workload is large and steady, the task is narrow enough to evaluate, and one of the constraints is commercially important. You may need local deployment, predictable latency, strict data residency, deeper control over model behaviour or protection against one provider changing the economics of your product. You also need people capable of running the system after the demo works.

Stay rented when the company is still testing the use case, demand changes sharply, model quality matters more than unit cost, or the team has no desire to become an AI infra company by accident. The fastest route to learning is often the API because it removes work that has nothing to do with the customer problem.

Most companies will end up somewhere between the two. They may use a frontier model for difficult reasoning and smaller deployed models for repetitive classification, extraction or internal tasks. They may switch providers behind an orchestration layer rather than commit everything to one model. The useful architecture will follow the economics and risk of each workload, rather than an ideology about ownership.

Open models increase choice, reduce dependence and let companies put AI into places a public service cannot reach. They are not free intelligence. They are infrastructure, with all the maintenance, security and organisational responsibility that word quietly carries.

Owning your AI stack is not proof that the company has grown up. Sometimes it is the right operating decision. Sometimes it is a founder spending three weeks becoming an accidental GPU administrator to save a bill that was never the real problem.

Most startups should keep renting while they are still learning what the AI should do. Ownership becomes sensible only when a real cost, risk or product advantage justifies everything that arrives with it.

See you out there.

Martin