In the past year, governments and large enterprises have started building their own infrastructures dedicated to training and inference of LLMs, driven by needs of digital sovereignty and data control. This trend, often labeled as 'Sovereign AI,' is now emerging as the primary growth vector for the semiconductor industry, according to recent sector analyses, including insights from Nvidia. The additional demand for GPUs and accelerators, no longer solely from large cloud providers but also from state bodies and local consortia, is shifting balances: for the first time, chips and systems are explicitly required to remain within defined geopolitical boundaries.

The flip side is that not everyone can take a seat at this table. The production of advanced chips, those needed for large-scale AI, is concentrated in very few global nodes — essentially TSMC in Taiwan, with a smaller share by Samsung in Korea — and requires multi-year investments and know-how that are difficult to replicate. Countries without fabs or that have not forged strategic alliances with manufacturers find themselves in a position of dependence, exactly the scenario that digital sovereignty aims to avoid. This is the sovereignty paradox: to gain control, you must purchase components from those who control global supply, at least until you develop domestic capacity.

This scenario has structural implications for those evaluating on-premise or self-hosted deployments. On one hand, regulatory pressure on data residency (GDPR in Europe, emerging regulations in Asia and the Americas) accelerates the building of local data centers. On the other hand, the hardware supply chain remains fragile and concentrated, making the planning of local AI projects a matter of TCO and risk management rather than mere economic convenience. Companies that today design their own AI infrastructures must evaluate not only the cost of GPUs and energy consumption but also the real availability and geopolitical stability of the supplier. It is no coincidence that more and more organizations are exploring long-term booking options or direct agreements with manufacturers, bypassing traditional distributors.

Who wins in this race? The large chipmakers — Nvidia first and foremost, which supplies H100 GPUs and the necessary software platforms — and advanced-node chip manufacturers. Also winning are providers of cooling and networking solutions, because on-premise AI demands higher compute densities compared to traditional enterprise workloads. Conversely, players that base their entire offering on public clouds and lack a hybrid or on-premise strategy stand to lose: the demand for 'sovereign clouds' will grow, but it will not erase the need for physical hardware control for the most sensitive workloads. Moreover, countries without foundries risk paying a strategic premium, diverting resources from local innovation.

In the AI-RADAR landscape, this tension between digital autonomy and manufacturing dependency is a critical factor in evaluating any on-premise stack. Choosing to run LLMs locally means embracing an architecture that, while ensuring privacy and control, remains tied to a global supply chain that is far from distributed. The 'next wave' of chipmaking powered by sovereign AI is already here: it's not a question of if, but of who can actually sit at the table.