Why companies want an AI of their own

Most companies meet AI as a rental. A seat in ChatGPT or Claude, a key for an interface, a bill that grows with every user. The models are excellent, and for most of the work in a company that arrangement is the right one. Our own company brain works in Claude and ChatGPT every day.

Some work is different. A contract that nobody outside may read. The test reports for a product that has not been announced. Decades of engineering know how, finally written down in a form a model can use. For that work the question changes. It stops being which model is best and becomes whose building the model runs in, and who learns from what it reads.

What companies say they worry about

The worry can be measured. For its 2026 benchmark study, Cisco surveyed 5'200 IT, technology and security professionals across 12 markets. Of those, 86 percent said data stored locally is inherently more secure, down from 90 percent a year earlier, and 81 percent reported heightened demand to keep data local. Three in four organisations had a dedicated body for AI governance. Only 12 percent called theirs mature.

The Swiss picture looks similar. In an EY survey of 604 companies published in May 2026, 89 percent of respondents said they use AI in their daily work. For 51 percent it was business critical that AI systems comply with Swiss or European data protection requirements.

Behind the numbers sits a simpler point. Clem Delangue, who runs Hugging Face, put it to TechCrunch in July: “you don’t want to outsource your core capabilities to another company.” For most companies, know how is the core. It is the part that is hardest to copy, and the part a rented model knows least about.

What the enterprise terms already cover

Renting has become safer than it was. OpenAI states that data sent to its API has not been used to train its models since 1 March 2023, unless a customer opts in. Anthropic states that by default it does not use inputs or outputs from its commercial products for training.

The details are where the work is. At OpenAI, logs kept for abuse monitoring are retained for up to 30 days by default. Zero data retention exists for eligible customers, after approval. A European region covering the EEA and Switzerland keeps customer content there, and using it requires one of the approved retention controls. System data such as account, billing and usage data may still be processed elsewhere, and for models released since 5 March 2026 the regional endpoints cost 10 percent more.

None of that is a reason to avoid the rented models. It is a reason to read the terms as a list of conditions. The contract decides what the provider may do with your data. It does not decide where the provider sits, when the model changes, or whether the model ever gets better at your work. Those are questions of architecture.

Open weight models are close enough

Open weight models, which a company can download and run itself, have come close to the best closed ones. Stanford’s AI Index 2026 found that in March 2026 the top closed model led the top open model by 3.3 percent. In August 2024 the gap had been 0.5 percent, so it has opened again, slightly. Six of the top ten models on the Arena leaderboard are closed.

Even the largest labs publish open weights. In August 2025 OpenAI released two open weight models under the Apache 2.0 licence. The larger one runs on a single GPU with 80 GB of memory, the smaller one within 16 GB.

Switzerland has its own. Apertus, built by EPFL, ETH Zurich and the Swiss National Supercomputing Centre, appeared as version 1.5 in July 2026, now with image and audio understanding, under Apache 2.0 and open to research and commercial use. Its makers point to the Canton of Ticino, which runs a translation service of its own on Apertus, so that sensitive government documents are translated without relying on commercial providers.

For a narrow task inside one company, the leaderboard matters less than whether the model knows the company’s terms, rules and past decisions. That is something a company can teach a model it owns.

Where a private model can run

Private covers three places. A company’s own cloud account, with its own keys. Servers in its own building or data centre. And closed networks that never touch the internet.

Each has a price. Owning a model means owning its operation: the hardware or the cloud contract, the updates, the security, the testing of every new version, and a person responsible for all of it. Renting means a bill that grows with every seat and every query, and a model the provider can change on any day.

For most companies the honest answer is both. The best rented model for open work, where breadth matters most, and a model of their own for the work that must stay inside. Delangue expects most production workloads to run either on private models within companies or on open models. We put it more modestly. The line between the two is the first decision, and it is worth drawing on purpose.

What private should mean

Private is an easy word to claim. These are the questions we ask before we use it.

Where it runs. A named cloud account or a named building, in a named country. A region on a map is a start, and only that.

Who can read it. Prompts, files, logs and backups, each with a retention period and a list of the people and subprocessors who can see them.

What it learns from. Your data trains nothing outside your company. Ideally it trains the model that works for you.

Who changes it. A version stays until you decide to replace it, and a new one can be switched back in one step.

How it is tested. Every new version is measured against the last one, on your own tasks, before it goes live.

Who can leave. The model, its adaptations and your data can move with you if you change provider.

Who answers for it. One named person owns the model, its rules and its gates. A model can read every contract in the archive. Deciding which clause the company accepts stays with a person, and so does the responsibility for it.

Why we work with Eigna

Eigna is built by our partner Eigna Labs, and it answers most of that list by design. Each company gets a private model of its own: an open model, adapted to its domain and trained on its own files, rules and language. It runs in the company’s own cloud or on its own servers, and it is designed to run fully offline in closed networks. Company data never leaves that cloud.

Teams work with local experts for their own tasks, from legal and risk to operations and support. Their corrections are kept inside the walls, the model retrains there in steps, and a new version goes live only when it tests better than the last one. Otherwise it stays out.

crue was one of the first companies to test Eigna, and we use it for our own work. When the work in a company’s Twin turns confidential or critical, Twin runs on Eigna instead of a rented model.

Eigna is in early access, and we say so. If the work is sensitive, the architecture should begin there.