Most AI coverage argues about benchmarks. I think that is the wrong fight. The real story of this era is AI sovereignty, which is a polite word for a blunt question: who controls the compute, who owns the models, and who can get cut off. Strip away the demos and every headline below is really about leverage, where it sits and who is quietly losing it. Here is the case, laid out.

Compute is the choke point

Start with the physical bottleneck, because everything else flows from it. I wrote that your AI revolution was waiting in a GPU queue, and that was not a joke, access to silicon is the actual gate. The hardware fight matters for exactly that reason, which is why GPT-4 Turbo going live alongside AMD’s MI300X was a sovereignty story dressed as a product launch, a second source of accelerators is a second set of hands on the tap. And when MIT techniques cut training cost and time, that is not just efficiency, it changes who can afford to sit at the table at all.

On-device versus the cloud

The other front is where the model actually runs, your device or someone else’s data center. That is the whole subtext of the on-device sovereignty question iOS 18 would not answer, and two years later Siri getting smarter inside the garden but still behind outside it showed Apple choosing control over raw capability on purpose. The shift toward local got real with on-device AI code generation, the MIT and AMD shift, proof the frontier does not have to live in the cloud. Google, meanwhile, kept shipping Gemini 2.0 with better benchmarks and the same sovereignty problem, because a better model you do not control is still a model you do not control.

The capability backdrop

None of the control questions matter unless the models keep getting better, and they did, fast. There was Gemini 1.5 Pro and the multi-cloud scramble, then AI reasoning getting a real boost, and the pace is best captured by the thirty months that compressed a decade, from GPT-4 Turbo to GPT-5.5. That curve is the reason sovereignty stopped being an abstract worry and became something governments and companies actually budget for.

Control, capital, and getting cut off

Which brings it to the thesis. I laid it out directly in the AI power shift, control, capital, and who gets cut off, the argument that this was never really about chips, it was about leverage. Europe’s answer was an EU AI Act legislating a sovereignty it has no infrastructure to back, governance without the compute underneath it. The clearest signal came when open-source AI went from preference to imperative after the Fable 5 shutdown, the moment getting cut off stopped being hypothetical. And the most unsettling reframe was Nadella treating the learning loop itself as the new IP, because if the loop is the asset, then whoever owns it owns you.

The full reading list

If there is one thread running through all of it, it is that capability is necessary and control is decisive. The lab with the best model does not win, the entity that owns the compute, the weights and the off switch wins, and right now that is a very short list of companies and an even shorter list of countries. France figured this out, Europe legislated around it, and the rest of us are finding out the hard way that you do not actually own software you cannot run without someone’s permission.