The Tony Blair Institute published a framework in January for how nations navigate AI dependency. Three positions: Control, Steer, or Depend. Every country sits somewhere on that spectrum, across different layers of its AI stack. No country has full-stack independence, not even China or the US. That’s not a case of failure. The failure is not knowing where you sit.

Enterprises are making the same calls right now. Mostly without realising it.

Call it the dependency default: the condition where your AI vendor posture isn’t the product of deliberate choices, but the accumulated result of individually reasonable and independent procurement decisions. Your cloud contract. Your model API. Your data pipeline. Your governance tooling. Each one made sense at the time. Nobody stepped back to look at the full picture. Together, they result in a posture you never agreed to.

This week, Palantir and NVIDIA announced a sovereign AI operating system. A complete, production-ready stack, from Blackwell Ultra GPUs through to a full application suite covering data management, AI orchestration, and deployment. Hardware to software. Turnkey. On your premises. It’s the Control posture, sold as a product for the first time to governments, defence agencies, and critical infrastructure operators. These are customers who can’t afford to find out mid-operation that their vendor has changed the pricing, deprecated a model, or shifted its roadmap.

But here’s what that announcement actually reveals. The ‘Control’ posture has a real operational cost. Running this stack requires internal engineering depth, GPU infrastructure management, and Kubernetes expertise that most enterprise teams don’t staff for. You can buy the architecture. That’s not the same as having the capability. Sovereignty isn’t a procurement activity. It’s what you find out you have, or don’t have, when you try to change something in your stack.

The TBI framework makes this legible beyond geopolitics. Across the AI stack, compute, data, models, governance, most enterprises have never looked at all four layers as a single picture. Each one is owned by a different team, a different budget, a different renewal cycle. The dependency default accumulates in the gaps between those conversations.

The answer isn’t to own everything. Even countries with serious resources don’t try that. Japan maintains selective fallback capability on specific critical layers. India built sovereign public infrastructure at the foundational layer and opened the rest. France uses regulatory leverage rather than direct ownership. In every case, the goal isn’t maximum control. It’s deliberate choice. Knowing what and where you’ve conceded, and why.

Most enterprise AI strategies don’t have that map. And the map only matters when something changes. When you try to switch a vendor. Renegotiate a contract. Move a workload. That’s when you find out which dependencies were a decision and which were just the path of least resistance.

What this means if you’re architecting AI

The question isn’t whether to depend on vendors. Every enterprise does, and for most layers of the stack, that’s the right call. The question is whether you know which layers you’ve conceded deliberately and which ones you drifted into. One is strategy. The other is drift.

Runtime Check

Three observations for CXOs before your next AI strategy review:

• Before your next vendor renewal, map your stack against Control, Steer, and Depend. Compute, data pipelines, models, governance. One layer at a time. If you can’t answer that question for your organisation today, the contract negotiation will answer it for you. So does almost every enterprise that hasn’t done this exercise.

• The sovereign AI architecture announced this week is real, and most enterprises aren’t ready to operate it. That’s not a reason to dismiss the Control posture. It’s a reason to understand which layers you’d actually need to own before it becomes a genuine option, rather than something a vendor pitches you on.

• The enterprises with the most leverage in the next two years won’t be the ones that moved fastest. They’ll be the ones that knew what they traded away to do it.

The dependency default doesn’t present itself, not in the near term. It shows up when something you thought was flexible but it turns out you’re a victim of lock-in. That’s the moment most organisations realise their AI posture was never a choice. It was just what happened. The good thing is that it’s not too late to start building towards a progressively sovereign approach to AI-driven transformation.