Sovereignty means owning your AI choices, says Mindware’s Mostafa Kabel
“If one technology provider disappeared tomorrow, how much of your AI capability could you keep running?” That is the test Mostafa Kabel, chief technology officer of Mindware, puts to every country and company building a sovereign AI strategy. Many national programmes rest on open-weight foundation models built by outside developers. Kabel argues that holding the weights does not remove dependency on the people who built them.
Kabel is based in Dubai. He has spent 17 years in technology distribution and holds postgraduate qualifications in artificial intelligence, cybersecurity and infrastructure. He recommends that organisations put their investment into proprietary data, language, industry knowledge and use cases, and treat any single model as one component in a multi-model architecture. Speaking to The Source Code, he explains where sovereignty sits across data, infrastructure and models. He also covers when training a model from scratch makes sense and which capabilities an organisation must hold itself.
What does sovereignty actually mean at each layer of the stack, from data and infrastructure through to model weights, and where does real control sit today?
Sovereignty spans more than one layer, so I look at it across three. The first is data sovereignty, which is about which datasets you hold and who controls them. The second is infrastructure sovereignty, which is about where the compute is hosted and who operates it. These two are the types of sovereignty most people already recognise, and organisations have been discussing them for years in the context of cloud and data residency.
AI adds a third layer, which is the model itself. With models, the hosting location matters far less than how much control and flexibility you have over the model. Can you change it, adapt it, move away from it, and understand how it behaves? I would call that a challenge more than a problem, but it is one every organisation needs to take seriously. If we agree that infrastructure, data and the model are the three main layers, the discussion around data and infrastructure is relatively mature. Model sovereignty is much more complicated.
To show where real control sits, I use a simple test. If one technology provider disappeared tomorrow, how much of your AI capability could you keep running? The answer tells you how dependent you really are. That is why everyone is talking about sovereignty now. Given what is happening in the world, organisations are asking very practical questions: what am I using, where does it sit, and what happens to my operations if it disappears? The market is also full of new entrants, which makes the question more pressing. Many organisations have built on providers they adopted quickly, and they are only now mapping how much of their capability depends on them.
Most national programmes are built on existing open-weight models and foundations. What are the strengths and limitations of that approach?
Building on an existing open-weight foundation is a practical approach, and in many cases it is the right one. Not every organisation or country needs to train a foundation model from scratch. Doing so takes enormous investment, time and specialist capability. Most of the value comes from elsewhere anyway. Building on an existing foundation lets an organisation put its investment where it creates the most advantage.
From my perspective as a CTO, I recommend investing in four things: your own data, your language, your industry knowledge and your use cases. I would put a line under use cases, as that is where the real game in AI is played. Value comes when a model is applied to a defined business problem, with your data and your knowledge behind it. Those are assets no external provider can replicate, and they are the parts of the stack you should own.
The limitation is that open weight does not mean dependency free. Even with an open-weight model, you may still depend on the original developer for updates, improvements, security fixes and the direction the model takes. You also depend on the community and tooling built around it. Many organisations assume that once they have the weights, they are independent. That is only partly true.
My recommendation is to use global innovation while building the architecture around your own assets. Take advantage of the best models the world is producing, and make sure your architecture, data and knowledge remain yours. That has been Mindware’s message in the market for some time, and it is now becoming the message of the big AI providers as well.
What should organisations think through before committing to a base model?
The first thing I would avoid is starting with the question of which model is number one today. The model leading today may well be overtaken tomorrow, as this market changes very quickly. We have seen it repeatedly. ChatGPT introduces a capability, open-source models follow with something comparable, Gemini comes in with its own strengths, and new players keep entering with impressive models. If your strategy depends on picking the winner, you will be revisiting it every few months.
I would start instead with security, data governance, performance and infrastructure requirements. What data will the model touch, and under what rules? What level of performance does the workload need? What infrastructure can you run it on, and where? Once you understand those requirements, choosing a model becomes an architectural decision.
Each model offers benefits the others do not, and you can combine them depending on what you need. A simple example from everyday use: many people use ChatGPT for emails and organising their work, or for legal and planning tasks, while many use Claude when they are building presentations, as it performs better there. Individuals are already working with several models, and enterprises should approach it the same way.
Remember what we learned in the cloud business. We told customers not to put everything in one cloud: be multi-cloud, and work with AWS, Microsoft and others. The same thing is happening with AI. Cost is part of this as well, as different models carry very different costs for different tasks. Enterprises will probably use multiple models for different workloads. That is how we approach AI architecture across the Mindware portfolio. The model should sit inside the architecture as one component, replaceable when something better or more suitable appears. That is our recommendation to every customer, and I can see large customers starting to move in this direction.
How do you see the balance between adapting existing foundations and training from scratch developing over the next few years?
For most enterprises, I expect adaptation to dominate. Training from scratch makes sense when there is a strategic requirement behind it, usually around national security, language or unique intellectual property. A country that needs a model to handle its language with full depth, or an organisation whose intellectual property is its core advantage, may have good reason to build its own. For everyone else, adopting and adapting an existing model is faster and usually more economical, and it lets the organisation focus on its own data and use cases.
I also expect more specialised and smaller models. Organisations will increasingly ask which is the most efficient model that can deliver the required business outcome. A smaller, specialised model can often do the job better, faster and at lower cost than a large general one.
That brings me back to the use case, which should guide every decision. If you understand what you need and what output you expect, that is your use case. Delivering a use case means more than saying you will automate a process or bring intelligence into a model. Once the work is complete, you need to go back to what you originally asked for. Did it deliver the ROI you wanted? Did it shorten the time frame you set out to reduce? Did it grow your business by the amount you expected? Those measures need to be agreed within the use case before you begin, and they should drive the choice of model, the level of adaptation and the investment. When organisations start from the outcome, the decision between adapting and building becomes much clearer.
Which capabilities does a country or company need to hold itself for a sovereign strategy to be real, and what can safely be sourced?
For me, sufficiency does not mean building every component yourself. No country or company can build everything, and trying to do so would slow innovation. What you need to control is the strategic layer: your data, security, architecture, governance and critical operational knowledge.
Critical operational knowledge is the part people underestimate. I describe it as mastery of the data. You need to know which data you want to train on and which data you must never train on. You need to understand the behaviour of the model you are using, how it responds, where its limits are and what it might do that you did not intend. There are many cases where a model did things nobody expected. Organisations that understand their data and their models deeply can spot those issues and correct them. Organisations that have handed that understanding to someone else cannot.
The technology itself can absolutely come from global providers. There is no reason to avoid the best models, platforms and infrastructure available. The key is keeping the ability to change technology without losing control of your business and your data. If a provider changes its terms, its direction or its availability, you should be able to move without starting again.
Sovereignty is about owning your choices, and that does not require owning everything. It is the line I use in every presentation I give, and I believe it is the most important principle for any country or organisation building an AI strategy today.
