Chinese AI chip maker Cambricon has announced a more than 120% rise in net profit for the first half of the year, with revenues doubling, despite remaining on the US blacklist. Meanwhile, NAVER and Nvidia are building around 55 megawatts of compute capacity in South Korea, explicitly framed as sovereign. Two signals saying the same thing: control over compute is becoming an attribute of statehood.
We have covered the compute race from an economic and environmental angle. Here is the political angle, which is becoming central.
What Cambricon's growth means
This figure deserves to be read correctly. A company kept on a US restrictions list, and therefore cut off from the most advanced technologies, doubles its revenues and sees its net profit more than double.
Two readings clash, and both hold some truth.
The first: export controls have failed, since they have mainly created a captive market for local manufacturers. A Chinese customer who cannot buy elsewhere buys from Cambricon, which funds its climb up the capability curve.
The second: these controls have worked, since the chips available remain behind the cutting-edge generation, and the company itself acknowledges supply risks. Its growth measures constrained demand, not necessarily technological parity.
A warning signal appears in its own accounts: inventories representing a very high share of its assets, exposing it to write-downs if demand falters. Rapid growth and fragility at the same time.
What "sovereign" actually means
The word is used everywhere, often without precision. It actually covers four very different levels, which are worth distinguishing.
Data sovereignty: your information stays on your territory, under your law. This is the most common and weakest level.
Model sovereignty: you hold the weights and can run them yourself. This is what open models enable.
Compute sovereignty: the machines running the model are on your territory and under your control. This is what NAVER is building.
Silicon sovereignty: you design and manufacture the chips. This is the level almost no one reaches, and the real stakes behind Cambricon.
A nation can have the first three and remain dependent on the fourth. That is precisely the position of most European countries.
It became urgent because of a concrete precedent we covered: the global suspension of two models on a government's decision. Before that episode, dependency was theoretical. Since then, every technical director knows a service can disappear for reasons that have nothing to do with technology or contracts. Sovereignty has stopped being a political slogan and become a line in a risk assessment.
The European position, honestly
Europe has the first level, with a solid regulatory framework on data. It has credible players on the second, with open models published by European companies. It is building the third, with compute capacity projects.
On the fourth, the situation is more mixed than is often said. Europe hosts absolutely critical links in the global semiconductor chain, particularly in manufacturing equipment, without which no advanced chip is produced anywhere. That is not nothing, and it gives Europe real leverage.
What is missing is the ability to design and mass-produce AI chips on its own soil. Leverage over the global chain is not the same as autonomy.
What to take away
What is at stake around these announcements goes beyond chip performance. We are witnessing the constitution of compute as a strategic national resource, much like energy in the last century.
The consequences are already visible: export controls, national projects, states funding open models as we saw with the US scientific programme, and now capacity explicitly presented as sovereign.
The question for a country is no longer whether it will have the best models. It is what it can still do the day someone else decides to turn off the tap. That is a less exciting question than benchmarks, and far more decisive.