> ## Content Index > Fetch the complete content index at: https://katecarruthers.com/llms.txt > Use this file to discover other available public pages before exploring further. # Open weights, closed exits: the next AI lock-in problem - URL: https://katecarruthers.com/open-weights-closed-exits-the-next-ai-lock-in-problem/ - Published: 2026-11-03T19:31:00.000Z - Updated: 2026-11-03T19:31:00.000Z - Description: Open-weight models offer more choice, but not necessarily more control. For Australia, AI sovereignty depends on capability, governance and the ability to leave. - Author: Kate Carruthers - Tags: ai, open weight models, ai sovereignty, ai governance, vendor lock-in, autarky Not every organisation wants to rent intelligence indefinitely from someone else’s platform. They want useful models. But they also want some say over where those models run, what happens to their data, what the whole exercise costs, and whether they can change suppliers without rebuilding everything. **The big AI suppliers appear to be getting the message.** As [The Register](https://www.theregister.com/ai-and-ml/2026/10/21/nvidia-woos-wary-europeans-with-more-open-weights-models/5304552?ref=katecarruthers.com) reports, Nvidia is courting European demand for sovereign AI with [open-weight models](https://en.wikipedia.org/wiki/Open%5Fweights?ref=katecarruthers.com) for enterprise search, document processing and autonomous driving. These are not just chatbots. They are building blocks for systems that work with an organisation’s own information and operational needs. And I would not mistake this for philanthropy. It is a recognition that the market for AI is bigger than the market for access to the largest closed model. There is money to be made helping organisations run models, not just charging them to use one. ## A benchmark score is not an operating model The frontier labs have spent years competing over intelligence benchmarks. That competition matters. But a benchmark score does not tell an organisation whether a model is suitable for the work it needs done. Can it extract information reliably from our documents? Can we evaluate it against our requirements? Can it operate within our security environment? Can we afford to run it at the volume we need? **But what happens if the supplier changes the deal?** A model can be impressive and still be the wrong choice. As I have argued before, [AI models are not interchangeable infrastructure](https://katecarruthers.com/ai-models-are-not-interchangeable-infrastructure/). Forrester’s analysis of open-weight models makes a related distinction. Enterprises need governance, security, integration, observability and operational support. Access to the weights expands their choices. It does not turn those choices into a reliable service. This is why the less glamorous parts of Nvidia’s announcement interest me. Embedding models, rerankers and document parsers do not attract the same attention as a chatbot producing a convincing essay. But they help organisations find and use information they already hold. There is a substantial difference between demonstrating that a model knows a lot and building a system that can work reliably with an organisation’s records. That is where [context and control](https://katecarruthers.com/context-control-ai-enterprise/) become operational concerns rather than features in a sales presentation. ## Follow the money Nvidia has a commercial reason to support this market. It sells the infrastructure needed to run models, not just access to the models themselves. Open weights fit that business quite comfortably. But the opportunity is not confined to infrastructure suppliers. OpenAI released its gpt-oss models in August 2025, making the weights available under the Apache 2.0 licence and supporting local deployment and customisation. That did not make its flagship proprietary models open. It did recognise that some customers want models they can download and operate themselves. Google’s Distributed Cloud announcement also accommodates open models, including Gemma and Mistral, within its platform. None of this amounts to a collective abandonment of closed AI. Suppliers can sell proprietary models while making money from the infrastructure, hosting and tooling around open-weight alternatives. ## Open at one layer, locked in at another Being able to download a model does not mean we can readily move the system built around it. We may still depend on proprietary orchestration, hosting, security services, monitoring and developer tools. Switching the model might be straightforward. Moving everything around it might not be. As Forrester argues, the platform remains a significant source of competitive advantage even when the underlying weights are available. This is part of the [sovereignty problem in a platform world](https://katecarruthers.com/the-hidden-politics-of-ai-sovereignty-in-a-platform-world/): control over one component does not give us control over the whole system. “Can we download it?” is therefore only the first procurement question. The harder question is whether we could actually **leave**. Can we move our data, evaluations and workflows? Can we obtain and support alternative infrastructure? Do we have the people who could make the move? Have we tested it? Otherwise, we risk congratulating ourselves on avoiding model lock-in while building platform lock-in around it. ## Open weights are not necessarily open source The terminology matters too. Open weights generally means the trained model parameters are available. It does not automatically mean the training code or sufficient information about the training data is available. Nor does it tell us what the licence permits. The Open Source Initiative’s Open Source AI Definition goes further. It requires freedoms to use, study, modify and share, supported by access to parameters, code and sufficient information about the training data. **Downloadable weights alone do not meet that definition.** The Register reports that Nvidia releases training data and recipes alongside the weights for many of its models. That deserves attention, but each release still needs to be assessed on its own terms. “Open” is not a substitute for reading the licence and examining what has actually been released. These distinctions determine what we can inspect, change, reproduce and maintain. ## Sovereignty requires capability For Australia, whether a model is local or foreign, open or closed is only part of the assessment. But the more useful question is what we can actually decide, change and sustain. Hosting a model in Australia may address a requirement. It does not, by itself, resolve questions about ownership, software dependencies, hardware supply, licensing or operational capability. As with cloud services, [data residency is not the same as sovereignty](https://katecarruthers.com/us-cloud-act-sovereignty-and-why-you-might-need-to-care/). Equally, using technology developed overseas does not automatically mean we have surrendered all control. We cannot sensibly pursue self-sufficiency across the entire AI supply chain. We can understand our dependencies and avoid making every important one irreplaceable. **That takes people.** People who can evaluate models, secure deployments, manage data, monitor performance and maintain alternatives. These are among the [skills we need in the AI age](https://katecarruthers.com/the-skills-we-need-in-the-ai-age/), not optional extras once the technology is installed. **A downloaded model without those capabilities is an asset on a server, not an operational sovereignty strategy.** ## We still have to do the governance Open weights can address some concerns about proprietary AI. They do not remove our responsibility for the systems we deploy. If we operate a model ourselves, we need answers to some basic questions: - What evidence shows it is suitable for the task? - What does the licence allow, and what restrictions remain? - What do we know about its training and subsequent modification? - Who maintains it and approves changes? - How will we detect failures and respond to incidents? - When would we stop using it? - What would we move to, and have we tested that move? These questions apply to closed services too. What changes is the allocation of responsibility. This is why I have argued for [treating AI products as governed assets with a clear lifecycle](https://katecarruthers.com/bringing-ai-products-to-life/), rather than perpetual experiments. Deployment is not the end of the work. Neither category gets us out of understanding what we are taking on, and whether we can do it. ## Become more capable customers The opportunity is not to pick a side in an open-versus-closed argument. It is to stop treating access to a frontier model as a complete AI strategy. Some tasks will justify a proprietary frontier service. Others may be better served by an open-weight model. [Some systems will use both](https://katecarruthers.com/the-two-layer-ai-stack-cost-discipline-meets-strategic-leverage/). For Australia, I would put the effort into shared evaluation methods, reusable deployment patterns, competent assurance, and procurement arrangements that preserve an exit. The suppliers are making room for customers who want to operate and adapt models, not simply consume them. More models to choose from will not make us more capable customers. The ability to evaluate them, operate them and leave when necessary will.