The AI Race Is About to Get Physical

For the past several years, the story of artificial intelligence has been told almost entirely in the language of software. Models are growing more capable by the minute, becoming faster and more specialised. Billions of dollars are flowing into foundation models, research labs and the startups promising to reshape entire industries. Governments have published national AI strategies and every boardroom in the country has debated integration roadmaps. The conversation has been widespread, forward-looking and, in important ways, incomplete.

AI is not magic. It does not live in the cloud in any metaphorical sense. Every model trained, every query answered, every image generated depends on physical infrastructure – data centres drawing enormous amounts of power, cooling systems running around the clock, fibre cables carrying data across continents, substations connecting it all to functioning electricity grids.

The High Energy Cost of Inference

We surveyed more than 200 senior AI decision-makers in the US, and the data hard numbers to what many inside the industry have been acknowledging. Almost 29% of organisations say energy costs are already limiting their ability to scale AI. Over a quarter (28%) say rising power prices have forced them to slow or pause AI training activity altogether. Grid connection delays, cooling limitations, connectivity shortages and planning bottlenecks are all cited as growing barriers. The infrastructure layer, long treated as someone else’s problem, has become everyone’s problem.

This is not entirely surprising in hindsight. The scale at which modern AI operates is genuinely extraordinary. Training a large frontier model can consume as much electricity as hundreds of homes use in a year – a single GPT-4 training run consumes roughly 50 GWh of electricity – equivalent to the annual consumption of 40,000 U.S. households. Running inference across millions of daily users requires data centers of a size and density that would have seemed implausible a decade ago. And demand is accelerating, driven by the proliferation of AI applications, the expansion of agentic systems and the increasing integration of AI into core business operations. Data center electricity consumption is growing at around 15% per year – more than four times faster than total electricity consumption across all other sectors combined. The physical world is being asked to absorb an exponential curve, and it was never designed to do so this quickly.

The Strategy Shifts to the Physical World

What makes the current moment particularly revealing is where the constraints are actually appearing. Land, it turns out, is not the core issue. Only around 14% of firms in our research identify land availability as their primary challenge. The harder problem is everything that needs to happen after you find the land. A site without grid capacity is useless; no planning approval means you’re stranded; and if there’s no resilient connectivity the AI build cannot function. The bottleneck is not space – it is the complex, slow-moving process of turning physical space into operational compute environments. That distinction matters because it is much harder to solve with money alone.

It also reframes how we should think about national AI competitiveness. The conversation has long focused on research talent, investment volume and the race to train the most capable models. But infrastructure strategy is emerging as an equally decisive variable. In regions where permitting moves faster, where energy policy is more aligned with industrial demand and where private investment in large-scale infrastructure is more mature, AI can simply scale more quickly. The US has structural advantages here that are often underappreciated. They are not glamorous advantages – they involve grid planning, zoning reform and utility coordination – but they are increasingly consequential.

Geopolitics is adding another layer of complexity. Around a third (33%) of businesses surveyed say they are actively considering relocating AI workloads to other regions or countries because of geopolitical concerns. Alongside the practical constraints of power and planning, organisations are now also weighing sovereignty, supply chain resilience and the political stability of the environments in which they deploy compute. The result is a reshaping of where AI infrastructure investment is flowing – towards regions with abundant renewable energy, available grid capacity and faster regulatory pathways, regardless of where those regions sit on a traditional technology map.

The first phase of the AI revolution was defined by what was possible in software. The next phase will be defined by what is achievable in the physical world – by the unglamorous, grinding work of building power infrastructure, securing energy supply and navigating the bureaucratic complexity of large-scale industrial deployment.

This is where the US will have an advantage that no amount of model capability can easily overcome. The future of AI depends on the grid holding up. That is less exciting than the headlines suggest. It is also more important.