If artificial intelligence is the new electricity, the United Kingdom has decided to build its own power plant rather than buying energy from overseas. With up to £60 million in funding, the British government is backing two new university research labs—at Oxford University and University College London (UCL)—with the goal of developing the “next generation” of AI systems. Not bigger or more powerful models in absolute terms, but technologies that are cheaper, more open, and capable of running on widely available hardware. The move is part of a broader strategy to strengthen national AI capabilities and reduce dependence on U.S. tech giants.

Two labs, two paths to efficiency

The two centers will pursue complementary approaches. UCL’s SOFAIR lab, led by Professor David Barber, will focus on creating open-source models that can run on common machines, without the expensive GPU clusters that dominate today’s landscape. “Many current AI systems suffer from basic issues like inaccurate responses,” Barber explains, “and they often use similar architectures. SOFAIR will bring together broader sciences and fresh ideas to create a new generation of models, reducing dependency on a small number of providers and boosting the UK’s sovereignty.”

At Oxford, the BOLD project headed by Professor Jakob Foerster will take an even more radical approach: moving away from the race for centralized computing power. “The UK cannot win the global AI race simply by trying to outspend big tech on data and compute,” Foerster says. “BOLD is about discovering fundamentally new ways to build AI that are more efficient, more open, and better aligned with human needs.”

Both labs will forge partnerships across academia, industry, and the public sector, marking the first major investment under the UK government’s AI strategy outlined by UK Research and Innovation (UKRI). The declared goal: making AI more affordable and sustainable for British businesses and citizens.

Efficiency is more than an academic exercise

For those tracking on-premise deployment dynamics, this news should be read beyond the headline. Research into less computationally demanding models has direct implications for running LLMs locally, without relying on third-party cloud services. Today, hosting a language model in-house means grappling with high hardware costs and specialist expertise. If the UK labs succeed, they could spawn solutions that run on mid-range servers or even workstations, slashing Total Cost of Ownership and easing compliance with regulations like GDPR.

The direction is clear: instead of endlessly scaling parameters and VRAM requirements, the focus shifts to smarter architectures, aggressive quantization, and modular designs. It’s the same spirit that animates many open-source frameworks used by those who self-host their models, a space AI-RADAR constantly monitors to provide evaluation tools for practitioners.

Digital sovereignty: the game is played on common hardware

The UK investment is part of a larger trend: digital sovereignty is becoming a priority for governments and companies that don’t want to hand over data and cognitive capabilities to external providers. Recent experience has shown the risks of depending on a narrow group of corporations for strategic technologies. Open, low-resource models promise to democratize access to generative AI, allowing even budget-constrained entities to train and run inference on sensitive data within their own data centers.

In this light, UCL’s SOFAIR lab, with its open-source mandate, could become an important building block. Models released under permissive licenses could be integrated into on-premise pipelines, fostering an ecosystem of alternatives to dominant vendors. Oxford’s bet, meanwhile, could influence basic research, leading to new architectures that rethink the relationship between data, compute, and learning.

What this means for those evaluating on-premise now

Beyond political announcements, the investment sends a concrete signal to IT decision-makers: AI’s future isn’t only in mega-datacenters, but also in local servers and edge devices. For those planning an LLM adoption strategy, the prospect of models optimized for local execution could shift CapEx and OpEx calculations, making feasible what today seems prohibitive. Analytical tools like those AI-RADAR provides in its on-premise trade-off section help navigate these choices, but in the meantime, the industry is sending a clear message: sovereign AI isn’t a utopia, but a trajectory backed by significant public investment.

Whether these labs will deliver truly disruptive innovations remains to be seen, but the direction is set. Artificial intelligence is increasingly becoming a commodity that can be managed in-house, without needing to ask for anyone’s permission.