More than 100 UK data centre projects have applied to burn gas to generate their own electricity. The reason? The national grid cannot keep up with AI demand. The dominant approach to AI development treats every challenge as a scaling problem. Need better performance? Add more compute power. Need a smarter model? Train on more data. Need more energy? Build a power plant. Need faster results? Build a bigger data centre.
This logic has produced an industry that now consumes 5.8% of the UK’s national electricity and is asking for more power capacity than the entire country uses at peak demand. Around 140 proposed data centre projects in Great Britain are seeking grid connections with a combined requirement of roughly 50 gigawatts – against a national peak of 45 gigawatts. The industry’s answer to this self-inflicted crisis has been to build its own power stations.
The world cannot afford this path
This is a global crisis in the making. In Ireland, data centres already consume 21% of the country’s electricity – more than all urban homes combined. Singapore imposed a three-year moratorium on new data centre developments in 2019 because the grid could not cope. Parts of Virginia, home to the world’s largest concentration of data centres, are facing electricity shortages that threaten to constrain economic growth across the region.
In emerging economies, the challenge is even starker. Countries across Africa, Southeast Asia and Latin America risk missing the AI opportunity because they lack the energy infrastructure to support hyperscale data centres. The numbers make the global position painfully clear. UK electricity for an energy-intensive industry costs $111.65 per megawatt hour. In Germany, it is $88.97. In France, $44.19. In the United States, $28. That cost gap is reshaping where AI infrastructure gets built – and while solar and battery combinations are now the cheapest source of electricity on a standalone basis, the practical challenges of grid integration mean that most countries are still struggling to unlock that advantage at scale.
US hyperscalers control nearly 70% of the European cloud market. Europe has produced just three foundation AI models, against 40 from the US and 15 from China. The response from much of the industry has been to call for more energy supply, more grid capacity, faster planning permissions, dedicated power generation. In other words: apply their brute-force logic by removing the constraints.
The problem is waste, not supply
The current generation of large language models and AI systems are extraordinarily inefficient. They process vast amounts of data whether or not it is relevant. They activate billions of parameters for tasks that require a fraction of that capacity. They treat computing power as if it were unlimited and free.
The roots of this waste are structural, not incidental. Much like Uber’s early strategy of burning cash to capture market share before worrying about profitability, the dominant AI providers have prioritised capability at any cost – flooding models with compute and data to outpace competitors. The result is that inefficiency is embedded in the architecture of the models themselves. These systems were not designed to be lean. They were designed to win a scaling race. That means they cannot simply be optimised at the margins – they need to be rebuilt from first principles to become truly scalable.
There is a different approach. Nature-inspired algorithms take their cues from how natural systems that evolved over millions of years of testing have solved problems – think of how ant colonies find the shortest path to food or how birds coordinate their flight in defensive patterns. These methods do not process every possible option. Instead, they explore promising directions, learn from what works, and quickly converge on effective, low-energy solutions.
Applied to AI systems, these techniques can dramatically reduce the computing resources needed. Rather than activating billions of model parameters for every task, the system learns to identify which parts of the model are actually relevant and focuses computing power where it matters. The result is AI that delivers the same capability while using a fraction of the energy. The question has never been whether you can make AI more efficient. The question is why so few companies are trying.
The answer is economic. Hyperscalers make money when organisations consume more compute, not less. Efficiency is a threat to the current ecosystem.
Sovereign AI needs efficient AI
The conversation around sovereign AI – nations building their own AI capabilities rather than depending on foreign infrastructure – is gathering pace across Europe, the Middle East, Africa and Asia. But sovereignty pursued through brute-force scaling simply replaces dependence on foreign technology companies with dependence on foreign energy markets.
A country with limited data centre infrastructure does not need to match the US gigawatt for gigawatt. It needs to maximise what it gets from the infrastructure it already has. If an AI system can deliver the same capability on one hundredth of the computing power, a small solar installation could deliver the same value as the French nuclear fleet. It does not need to burn gas to power its AI ambitions. It does not need to choose between its climate commitments and its technology strategy.
The real efficiency dividend
It is important that we are precise about what efficiency means in this context. In addition to environmental benefits, it is a cost story: running the same AI workloads for a fraction of the price. It is a speed story: AI that runs on local infrastructure responds in milliseconds rather than making round trips to distant data centres. It is a sovereignty story: AI that runs where your data lives, on infrastructure you control. And it is an auditability story: efficient, well-structured AI systems can show you exactly how they reached a decision, which matters enormously in regulated industries.
For a financial institution that cannot send customer data to a foreign cloud provider, or a trading firm where milliseconds determine profitability, or a healthcare system that needs to explain AI-assisted diagnoses – efficiency is what makes AI deployable in the real world rather than just impressive in a demo.
A different thesis
There is a growing body of evidence that smarter search and reasoning beats simply adding more computing power. That rigorous mathematical approach can achieve what the industry assumes requires billions in funding and entire power stations worth of energy. This is not a popular thesis in an industry that has raised hundreds of billions of dollars on the promise that scale is everything.
But the evidence is becoming hard to ignore. When more than 100 data centre operators in a single country are applying to burn fossil fuels because the grid cannot support their energy demands, the scaling thesis is proving its limits. The AI industry has a choice. It can continue building its own gas supplies. Or, it can build AI systems that do not need them.
By Mathew Haswell, Co-Founder, Refiant
Stay ahead of the stories shaping our world. Subscribe to Impact Newswire for timely, curated insights on global tech, business, and innovation all in one place.
Dive deeper into the future with the Cause Effect 4.0 Podcast, where we explore the ideas, trends, and technologies driving the global AI conversation.
Got a story to share? Pitch it to us at info@impactnews-wire.com and reach the right audience worldwide
Discover more from Impact Newswire
Subscribe to get the latest posts sent to your email.
