NextFin News - The most important change in tech reporting is not a new app or a louder platform war. It is the migration of the AI story into the physical world, where the limiting factors are land, power, cooling, water and local consent. That shift matters because it changes artificial intelligence from a software race into an infrastructure race, and infrastructure obeys a different clock.
The Guardian’s latest reporting on datacentres captures that transition. A Scottish AI complex billed at £8.2bn was found to have overstated the feasibility of its renewable-energy promise. A separate investigation found that large datacentre projects around the world are increasingly running into energy, construction and community obstacles. In Slough, Europe’s largest datacentre hub, emerging research suggests nearby temperatures can rise by an average of 2C and as much as 9C. The story is no longer only about what AI can do online. It is about what it takes to keep the machines running offline.
The scale is already visible. The Uptime Institute has identified 250 global datacentre projects exceeding 100MW in energy demand that were announced between 2021 and 2024. It says roughly half of those projects will either not happen or be delayed. Even if those delays prove less severe than feared, the same analysis says planned projects announced last year alone, assuming they run at 25% of planned capacity, would consume 1.3% of the world’s projected electricity usage for 2025. That is not a niche infrastructure problem. It is a macro one.
The physical constraints are now shaping the narrative that once belonged almost entirely to software. Reporters who used to track launches, downloads and product updates now have to check grids, planning approvals, energy infrastructure and local environmental impacts. The industry’s center of gravity has moved from the browser tab to the substation. When a project depends on a site, a transformer and a power connection, the questions become slower, heavier and more concrete.
That also changes what counts as a meaningful AI announcement. The headline may still be model capability or funding, but the real determinant of success is whether the project can be built on time and at scale. In the UK, the North East AI growth zone is being built around 1.1GW of eventual capacity, with over 400MW due online in 2028. In Scotland, officials and developers have had to confront whether the power that underpins the promise can actually be delivered. In the US, similar projects are being tested by the same bottlenecks: grid access, chip supply, cooling systems and local resistance.
That is why the current datacentre boom looks less like a standard cycle and more like a structural shift. A cyclical problem would be a short-lived gap between demand and supply that closes once investment catches up. A structural problem is a regime change in which the bottleneck itself becomes a feature of the system. The current buildout has cyclical elements - supply shortages, long lead times, permitting delays - but the deeper driver is structural: AI demand keeps pushing upward faster than the physical networks that support it can expand.
The Tech Story Has Become A Grid Story
What changed first: the technology, or the economics around it? The answer is both, but the economics are now governing the technology. As model training and inference loads rise, the need for electricity, land and cooling rises with them. That means the AI boom increasingly behaves like an industrial expansion rather than a pure software cycle. The project pipeline can be announced quickly, but building a datacentre is a long, capital-intensive process that can be slowed by every stage of the chain.
The Uptime Institute’s numbers make the scale hard to dismiss. Two hundred and fifty projects above 100MW is a huge pipeline even before you account for the likelihood that many will slip or die. Roughly half of them, Uptime says, will not happen or will be delayed. And the aggregate electricity demand implied by planned projects announced last year alone - 1.3% of projected global electricity use in 2025 if they run at a quarter of capacity - shows why energy utilities, regulators and local governments are now core actors in the AI story. This is no longer a narrow technology debate. It is a resource allocation debate.
That resource question helps explain why local reporting matters so much. In Lanarkshire, the promised AI complex was not just a corporate land deal; it was a claim about energy realism. The government and developers publicly linked the site to a renewable-powered future, yet the investigation found internal acknowledgement that power provision was a problem. That gap between promise and deliverability is the crucial mechanism. A datacentre can only exist if the power system, the planning system and the financing system all agree at once. If one of them fails, the project stalls.
In the US, the same logic appears in a different form. The datacentre pipeline is so large that even committed projects can run into delays from transmission constraints, equipment shortages and community opposition. A mega-project is not only a software decision wrapped in steel. It is a long sequence of physical commitments. The longer that sequence becomes, the more likely it is that public scrutiny, capital cost and supply-chain friction will change the outcome. The AI boom is therefore becoming a test of coordination, not just ambition.
That is why the shift feels structural rather than cyclical. Cycles are about inventory and timing. Structure is about rules, assets and irreversible commitments. Here, the relevant assets are power lines, cooling systems, land parcels and local social license. Those do not scale at the speed of software. Once the industry needs them, it inherits their constraints. That makes the bottleneck durable unless and until the physical system itself is rebuilt.
“The global supply chain just cannot support the level of projects out there, on the timeline that is projected. The scale is such that it’s going to slow things down,” said Jay Dietrich, research director at Uptime.
That is the mechanism in one sentence. The supply chain cannot absorb the pace of project announcements, so delays are not aberrations but a predictable outcome. The second-order implication is more important than the first-order one. First-order, AI keeps getting more compute-hungry. Second-order, the cost of that compute becomes a function of infrastructure scarcity. Third-order, AI economics start to look less like software margins and more like industrial margins.
The strongest counter-thesis is that this is still just a temporary bottleneck in a rapidly expanding sector. Under that view, today’s delays are the byproduct of an immature market. More capital will draw in more suppliers, and over time the grid, the equipment makers and the developers will catch up. That argument has real force because the biggest technology companies are still pouring money into AI, and because infrastructure markets often respond to shortages by investing more.
But the counter-thesis weakens where the binding constraints are hardest to solve. Money can move quickly. Transformers, transmission lines, permits, water rights and local approvals cannot. If the current boom were only a short-term imbalance, you would expect project completion rates to improve quickly as more money enters the system. The falsifying signal for the structural thesis is therefore concrete: if the share of announced >100MW projects completed on schedule rises materially above Uptime’s roughly 50% delay-or-cancel expectation, and if grid-connection times shorten across major hubs instead of lengthening, then the bottleneck would be looking cyclical rather than structural.
For now, the evidence points the other way. The reporting shows not a temporary inconvenience but a new industrial reality. The AI race is increasingly being decided by who can build and power the machines, not just who can design them.
What The Physical World Means For AI’s Next Phase
The short-term winners are the firms and contractors that can turn digital demand into physical capacity. That includes developers with ready land, companies with secured power, suppliers of electrical equipment, cooling systems and grid hardware, and operators that already have the approvals to expand. Their advantage is simple: they can move faster because they are less exposed to the bottlenecks that now define the sector.
The exposed group is broader. It includes AI firms that have promised aggressive rollout schedules, speculative developers that assumed cheap and easy power, and governments that have sold AI infrastructure as a fast route to growth without fully pricing in the engineering burden. It also includes local communities that must absorb heat, noise, traffic and land-use trade-offs long before any promised economic upside arrives.
The time horizon matters. In the short term, the market can still reward AI enthusiasm because demand for models, cloud services and AI features remains strong. In the medium term, physical bottlenecks can slow deployments, push up costs and change which projects survive. In the long term, the industry is likely to look more like a regulated utility-and-infrastructure stack than a pure software sector. The firms best placed to benefit are the ones that can control both the digital layer and the physical layer beneath it.
The base case is continued AI expansion, but at a slower and more expensive pace than the public narrative suggests. The upside case is that grid investment, better cooling technologies and better coordination between developers and utilities allow more of the announced projects to get built. The downside case is that planning fights, local opposition and power scarcity cause a larger share of projects to slip or die, forcing a reassessment of how quickly AI can scale.
The next data points to watch are practical, not theatrical: how many announced sites reach construction, whether power connections arrive on schedule, whether developers keep matching public claims with actual infrastructure, and whether the cost and timing of grid access keep rising. If those measures improve, the buildout can still accelerate. If they do not, the physical world will keep slowing the digital one.
The clearest lesson is that the AI revolution now runs on concrete as much as code. The companies that ignore that fact will discover that the real bottleneck is not the model - it is the machine room.
Explore more exclusive insights at nextfin.ai.
