Open vs Closed Models

Welcome back to the NEXUS AI Literacy Series. This episode is about one of the most important strategic choices a company makes when it adopts AI, and one of the biggest ongoing debates in the whole field: open models versus closed models. Should you use a closed model from a big provider — through an API, like renting — or an open model that you can download and run yourself? This is the Linux-versus-Windows question of the AI era, and the answer shapes your costs, your control, your privacy, and your flexibility. By the end of this episode, you'll understand the real trade-offs and how to think about which is right for a given situation.

Let me define the two clearly first, because the names are a little misleading. A closed model — sometimes called proprietary — is one owned and operated by a company, that you access over the internet through their service. You don't get the model itself; you send your request to their servers, their model processes it, and they send back the answer. Think of the big-name commercial AI services. You're renting access to a model you never actually possess. It lives in their house; you visit through a door they control.

An open model — you'll often hear "open weights" or "open source" — is one whose actual guts, the trained parameters, the billions of dialed-in numbers, are made publicly available for download. You can take that model, run it on your own computers, your own servers, your own cloud, modify it, build on it — without asking anyone's permission or sending your data anywhere. You possess the model. It lives in your house.

Now, the cleanest way to feel the difference is an analogy. A closed model is like ordering from a restaurant. It's incredibly convenient — you don't buy ingredients, you don't cook, you don't clean up. You just order, and a professional kitchen delivers a polished meal. But you eat what's on their menu, you pay their prices every time, and you have no idea exactly what's happening back in that kitchen. An open model is like getting the recipe and cooking at home. More work — you need the kitchen, the ingredients, the skill. But you control every ingredient, you can tweak the dish however you want, you know exactly what's in it, and once you're set up, each meal is far cheaper. Restaurant versus home kitchen. Convenience and polish on one side; control and ownership on the other.

Let me walk through the real trade-offs, because each side genuinely wins on different dimensions.

Start with ease and capability. Closed models win on convenience, and historically they've held the lead at the absolute frontier — the most capable models in the world have tended to be closed. You sign up, call the API, and you're using a world-class model in minutes, with none of the infrastructure headache. For most businesses getting started, that's a massive advantage. The flip side: open models have been catching up at a startling pace. The gap between the best closed model and the best open model has narrowed dramatically, to the point where, for a huge range of real tasks, a good open model is more than capable enough. The frontier crown is still usually closed, but you often don't need the absolute frontier.

Now control and customization. This is where open shines. With an open model you can fine-tune it deeply on your own data, shape its behavior, run it exactly how you want, with no rate limits imposed by a vendor and no risk that the provider changes the model or the terms out from under you. With a closed model, you're a tenant — if they update it, deprecate it, raise prices, or change policies, you adapt or you're stuck. Open means you own your destiny; closed means you're along for the provider's ride.

Now privacy and data — and this one is huge, especially for us. With a closed model, your data leaves your walls and goes to the provider's servers to be processed. For a lot of uses that's perfectly fine. But for sensitive data — health records, confidential business information, anything regulated — sending it to an outside service can be a serious problem, sometimes a legal non-starter. With an open model running on your own infrastructure, your data never leaves your control. You can process the most sensitive information without it ever touching a third party. This is exactly why, for regulated or highly confidential workloads, self-hosting an open model is often not just preferable but necessary. It's a big reason the open option matters so much in fields like healthcare.

And cost, which ties back to last episode. Closed models are pay-per-use — no upfront cost, you pay for every call. Convenient, and great at low or unpredictable volume. But at high, steady volume, those per-call fees add up fast. Open models flip it: there's real upfront cost and effort to set up and run the infrastructure, but once you're running, the marginal cost per use can be far lower. So the math often comes down to scale — low or spiky volume favors renting; high steady volume can favor owning. Like deciding whether to keep taking taxis or buy a car.

So how do you actually choose? Here's the framework. Reach for a closed model when you want maximum capability and convenience with minimal setup, your volume is low or unpredictable, and your data isn't especially sensitive — which describes a great many projects, especially early on. Reach for an open model when data privacy is paramount, when you need deep customization and control, when you're running at high enough scale that owning beats renting, or when you want to avoid being locked into a single vendor. And the sophisticated answer, the one that's increasingly common: it's not either-or. Many serious operations use both — a top closed model for the hardest, highest-value tasks, and self-hosted open models for the high-volume, the sensitive, and the routine. You match the tool to the job.

Let me connect this to how we actually operate, because it makes it concrete. The pragmatic posture is: start with closed models to move fast and prove value — don't stand up infrastructure before you know what you're building. Then, as specific needs emerge — a sensitive data workload, a high-volume task, a need for deep control — selectively bring in open models where they clearly win. That's not fence-sitting; that's using each for what it's genuinely best at. The worst move is treating it as a religious war and forcing everything onto one side.

So let's bring it home. A closed model is one you rent access to over the internet — convenient, polished, often the most capable, but you don't possess it, your data leaves your walls, and you pay per use. An open model is one you download and run yourself — more setup, but full control, deep customization, your data stays home, and lower cost at scale. Restaurant versus home kitchen. Closed wins on convenience and the frontier; open wins on control, privacy, and steady-state cost — and the two are closer in capability than ever. Choose by your needs — capability and speed versus control and privacy — and don't be afraid to use both.

When this debate comes up, you'll now be the person in the room who can lay out the real trade-offs instead of cheerleading for one side — which is exactly the kind of clear thinking that earns trust on a decision this consequential.

In the next episode, we revisit an idea we met in the Foundations track and look at it strategically: scaling laws. Does making AI bigger reliably make it better, and will that keep working forever? The answer shapes where this whole technology is headed — and whether the giants stay on top. See you there.