AI may be advancing at extraordinary speed, but the infrastructure required to support it is moving on a very different timetable. Power, cooling, networks, equipment and skilled people are increasingly determining how quickly new intelligence can be deployed.
When Vantage Data Centers and Nebius announced in August that they would deploy high-density NVIDIA-powered AI infrastructure at Vantage’s Newport campus in South Wales, the significance went beyond another capacity deal. The project is the first announced commercial deployment in the South Wales AI Growth Zone and is intended to support training, inference, agentic AI and enterprise workloads. It also captures a wider change taking place across the industry. The race to build better models is increasingly inseparable from the race to secure the physical infrastructure capable of running them.
For much of the recent AI boom, attention has remained fixed on chips, models and applications. Yet the limiting factors are moving steadily down the stack. Compute can only be deployed where there is enough electricity, enough cooling, enough network capacity, suitable land, available equipment and people with the skills to design and operate increasingly complex facilities. The infrastructure behind intelligence is no longer a supporting story. It is becoming one of the main determinants of how quickly AI itself can scale. That changes the competitive equation, because access to infrastructure becomes a strategic advantage long before a customer ever switches on the first accelerator or serves an inference request.
Capacity is becoming a physical question
The numbers explain why. JLL expects global data center capacity to rise by 97 GW between 2026 and 2030, effectively doubling to around 200 GW. It estimates that the sector could require as much as $3 trillion of combined real estate and tenant capital expenditure by 2030. At the same time, AI could account for half of all data center workloads by the end of the decade, with inference overtaking training as the dominant AI requirement as early as 2027.
That forecast is striking not simply because of the volume of new capacity required, but because JLL makes an important assumption: energy innovation will have to mitigate persistent power constraints for even its base case to be achieved. Power has already displaced location and cost as the primary site-selection criterion in many markets, while the average wait for a grid connection in major data center locations now runs into several years. The problem is therefore no longer whether demand exists. It is whether the physical system can respond quickly enough to convert demand into operational capacity.
Capgemini’s 2026 research makes that tension even clearer. Seventy-seven per cent of electricity executives surveyed believe data center demand will grow faster than their ability to expand supply, while 68 per cent foresee electricity shortfalls in some regions within the next few years. Among data center executives, 78 per cent identify grid infrastructure construction timelines as a critical constraint. The mismatch is structural: digital infrastructure can be conceived, financed and built on a timeline that electricity networks were never designed to match.
AI also changes the nature of the load being requested. Capgemini found that almost 80 per cent of electricity executives expect greater volatility and more pronounced demand peaks as AI usage increases. Large clusters concentrate huge loads in specific locations, placing pressure not only on total generation but on substations, transmission, distribution and reserve margins. That means a region can have sufficient electricity in aggregate and still be unable to deliver hundreds of megawatts to the place and at the speed a new AI campus requires.
The constraint is no longer one thing
Power may be the most visible bottleneck but treating it as the only one risks misunderstanding what is happening. Uptime Institute’s 2026 survey of more than 800 data center owners and operators shows management concerns widening across the whole infrastructure chain. Costs remain the largest concern, but capacity forecasting, power availability, supply-chain disruption, staffing and meeting AI infrastructure needs all rank prominently.
The supply chain is particularly important because every new megawatt requires far more than an electricity contract. Grid upgrades need transformers, switchgear and substations. Facilities need generators, UPS systems, busways, cooling equipment, pumps, heat exchangers, racks and networking. Capgemini found that 74 per cent of electricity executives are dealing with rising costs and delays for critical infrastructure equipment. Some generation equipment is already subject to delivery schedules measured in years. Even when capital and power are available, the components needed to turn them into usable capacity can become the next constraint.
The pressure on physical infrastructure is being matched by a growing shortage of the specialist skills needed to build and operate it. Uptime reports that 53 per cent of operators now struggle to find qualified candidates for vacant roles, up from 46 per cent in 2025. Large AI facilities intensify that challenge because higher thermal densities and increasingly sophisticated electrical and cooling systems demand different expertise from conventional environments. A sector planning to almost double global capacity in five years must therefore expand its workforce almost as aggressively as its physical estate, while ensuring operators have the skills to manage infrastructure that is becoming more complex with every generation of AI hardware.
Cooling provides perhaps the clearest illustration of how quickly the engineering assumptions are changing. JLL notes that AI training environments commonly demand rack densities of 40 to more than 100 kW, making liquid cooling increasingly necessary. Yet Uptime’s operator data shows a more complicated transition. The average of operators’ most typical rack densities moved above 11 kW for the first time in 2026, but without a relatively small number of extremely dense facilities the figure is only 7.8 kW. The industry is simultaneously operating a huge legacy estate and preparing for racks whose requirements look radically different.
The mismatch between the lifespan of a data center and the technology operating inside it is becoming one of the industry’s most difficult design challenges. Operators cannot rebuild entire facilities every time a new accelerator generation changes power or cooling requirements, while designing everything around the most extreme demands of today’s AI hardware risks creating expensive capacity that may never be fully used. The priority is therefore shifting from simply adding infrastructure to creating facilities with enough flexibility to accommodate successive generations of compute without requiring fundamental redesign or becoming obsolete before the asset has earned its return.
AI changes the building as well as the load
Arup’s Data Centre Futures work puts this problem in longer-term perspective. Data centers are capital-intensive assets designed and financed for long operating lives, while the technologies they support are becoming more fluid and uncertain. The report warns of the risk of lock-in and redundancy when permanent physical infrastructure is built around assumptions that may change far more quickly than the building itself.
The consequences are beginning to appear in the way operators think about the facility itself. AI training, inference, storage, edge computing and emerging forms of specialised compute place very different demands on power, cooling, latency and location, making a single uniform design increasingly difficult to justify. Large training workloads may favour centralised campuses with abundant power, while inference and edge applications can require proximity to users, lower latency or tighter sovereign control. Arup argues that this divergence is pushing the industry towards more differentiated environments, with modular and reconfigurable infrastructure allowing new generations of technology to be integrated without forcing a complete redesign of the site.
This is a more significant shift than simply replacing air cooling with liquid cooling or installing larger electrical feeds. It changes how facilities are planned. JLL expects enterprises to adopt hybrid portfolios spanning on-premise, colocation, hyperscale and edge environments. Arup similarly argues that operators should plan portfolios rather than individual sites, deciding which workloads belong at the edge, in regional aggregation facilities or in core campuses. Connectivity, power, water, security and local skills then become variables in a system rather than a checklist applied to every building.
The result is that data center strategy increasingly becomes an exercise in optionality. Facilities need upgradeable power, multiple cooling paths, room for new interconnects and layouts that can be reconfigured. Operators need relationships with utilities, fibre providers, equipment suppliers and local authorities established earlier in the development process. Procurement strategies may need multiple suppliers to reduce lock-in around accelerators, cooling systems and network fabrics.
The bottleneck will keep moving
There is a temptation to view the current infrastructure squeeze as temporary: grids will be upgraded, factories will increase equipment production, more engineers will be trained, and cooling technology will catch up with higher rack densities. Some of that will happen. But the deeper issue is that AI demand and infrastructure development operate on fundamentally different clocks.
The difficulty is that the technologies driving demand evolve on a far shorter cycle than the infrastructure built to support them. AI models can change within months and server generations within a few years, while data centers, substations and transmission networks are designed to operate for decades. Arup warns that decisions being made now will therefore shape exposure to technological, regulatory and environmental risk long after today’s dominant AI architectures have been replaced. The challenge is less about trying to predict which technology comes next than ensuring that sites retain enough flexibility to remain useful across multiple generations of compute.
This is why the infrastructure conversation needs to move beyond how many gigawatts can be announced. The meaningful question is how much capacity can actually be delivered, connected, cooled, equipped, staffed and adapted to whatever comes next. JLL’s near-100 GW expansion forecast illustrates the scale of ambition, while Uptime and Capgemini show the constraints already appearing across the system.
The next phase of AI will still be shaped by better processors and more capable models, but their impact will depend increasingly on an industry that has to turn unprecedented digital ambition into physical reality. The companies that solve that problem will not merely provide buildings for AI. They will determine where, when and at what scale intelligence can be deployed.
