The logic of scale is seductive. Organisations like OpenAI, Google DeepMind, and Anthropic argue that bigger models bring predictable gains in reasoning and creativity. But evidence is emerging that this assumption is starting to falter and that generative AI may not sustain the exponential demand for GPUs that so much global investment is built upon.<\/span> <\/p>\n Today\u2019s hyperscale language models are our new mainframes: powerful, but centralised, fragile, and controlled by a handful of global firms.<\/span> <\/p>\n True sovereignty means controlling how data is processed, protected, and governed. That\u2019s impossible when relying on models hosted or operated abroad. Even \u201cUK instances\u201d of foreign models cannot change the fact that model training and governance policies are decided elsewhere.<\/span> Finally, sovereignty must mean accountability.\u00a0 In recent weeks, both AWS and Azure have experienced global scale outages created from errors incurred in the US.\u00a0 Other countries have had to accept the problems but have had no ability to identify or mitigate the issues. This issue of being a passenger in someone else\u2019s technology \u201cvehicle\u201d will only get worse in an age of hyper-scale GenAI.<\/span><\/p>\n <\/p>\n Fortunately, a new alternative is emerging. Advances in model compression now allow smaller, domain-specific models to run securely on local devices. These systems can operate offline, respect privacy boundaries, and collaborate through encrypted orchestration layers – like the UK\u2019s orchestration challenger; Great Wave AI – which are the digital equivalent of an AI operating system.<\/span> <\/p>\n
\n<\/span>
\n<\/span>If that proves true, history offers a useful analogy. In the 1950s and 60s, computing was dominated by vast, centralised mainframes that only the wealthiest institutions could afford. The assumption then, too, was that bigger central terminals would always be better. The real revolution, of course, came later when compute power migrated from research labs to the desks of workers around the world.<\/span><\/p>\nHyperscale models are the new mainframes<\/h3>\n
\n<\/span>
\n<\/span>According to RAND (2024), the economics of foundation models are already \u201cprecarious,\u201d with soaring compute and energy costs delivering ever-smaller performance gains. Beyond the economic and environmental implications of this, they also raise deeper questions about sovereignty.<\/span><\/p>\nAI sovereignty: Beyond data location<\/h3>\n
\n<\/span>
\n<\/span>As Harrison Kirby, Author of \u201c<\/span>Exploring GenAIOps\u201d<\/span><\/i> and co-founder at Great Wave AI, has argued:<\/span>
\n<\/span>
\n<\/span>\u201cAI sovereignty isn\u2019t just about data location\u2014it\u2019s about control, governance, and accountability. Just think back to COVID-19, when even the closest allies competed for scarce vaccine resources. What happens to our \u2018sovereign\u2019 instances of overseas tech when priorities diverge?\u201d<\/span>
\n<\/span>
\n<\/span>Beyond sovereignty, there are structural security risks. Large models must process data in unencrypted form \u2014 meaning every sensitive input, from defence intelligence to patient records, is briefly visible inside a foreign-controlled system. How long can that remain acceptable?<\/span><\/p>\nThe rise of distributed, sovereign AI<\/h3>\n
\n<\/span>
\n<\/span>This is an important historical parallel: control of the operating system has been the foundation of US technology dominance for four decades. Owning the orchestration layer, rather than the models themselves, could be Britain\u2019s strategic advantage.<\/span><\/p>\nPractical examples of sovereign AI in action<\/h3>\n