Behind the excitement of AI infrastructure, from hyperscale data centres to GPU compute mega-campuses, lies a growing concern. While long-term demand for AI is widely expected to grow, the shape and timing of that demand is still taking form. The market is shaped by rapid chip cycles, fluctuating pricing and shifting workload demands, and the old ‘build it and they will come’ approach has resurfaced. Only now, the stakes are far higher and there is an expectation to get more certain on when ‘they will come’.  

Financial analysts are beginning to flag the tension between rapid AI infrastructure expansion and the pace at which demand is materialising. Some reports suggest AI infrastructure spend is outpacing realistic demand projections, and there is an increasing focus on how quickly demand can convert into committed, revenue-generating workloads.  

Yet projections do point to strong long-term growth. While questions are emerging about who bears the risk if utilisation levels fall short, the answer is not to halt or slow the development of vital infrastructure. Instead, utilising more flexible and phased deployment models will enable operators to better align capacity with usage and meet AI ambitions. Therefore, infrastructure can be deployed, adapted and expanded in line with realistic, evolving workload requirements. 

The demand complexities  

There is a widely held assumption that demand for AI compute will be effectively unlimited. That if capacity is built, it will inevitably be used. This assumption is likely because projections point to sustained growth in training and inference workloads, with data centre capacity demand expected to grow by around 20–25% annually through 2030.  

In reality, AI demand is complex and still evolving. Enterprise adoption remains uneven, many organisations are still in pilot (or experimentation phases) and not all workloads require large-scale, always-on infrastructure. Recent research shows that only around 20% of companies have scaled AI capabilities across their organisations, highlighting the gap between investment and real, sustained usage. At the same time, AI interest is undeniably accelerating, driven by rapid advances in generative models and increasing enterprise confidence. Most importantly, there is clear risk of over-cautious investment that could leave markets underprepared and unable to keep pace with national and commercial AI ambitions.  

Speed versus commercial discipline  

Historically, large data centre developments were underpinned by long-term agreements with hyperscalers or enterprise tenants before construction began. These commitments provided a clear route to secure revenue, reducing exposure to market volatility. Across the last few years, however, projects are increasingly moving forward while long-term customer commitments and workload requirements are still taking shape. 

This shift in commercial models is partly driven by the speed of the market. The pressure to deliver capacity quickly has intensified, particularly as new entrants such as GPU-as-a-service providers compete alongside established hyperscalers. In this fast-paced environment, the ability to deploy infrastructure rapidly can outweigh the discipline of securing customers in advance. But this comes with risk that needs addressing.  

Without committed users, developers are exposed to two key challenges. First, the utilisation risk; the possibility that capacity cannot be filled. Yet this is only likely if the facility is designed around specific architectures or hardware configurations that do not align with actual customer requirements. Where data centres were once built to operate for 10 to 15 years with minimal change, operators are now dealing with hardware refresh cycles closer to three to five years. The right technology for today’s infrastructure may not be fit for tomorrow.   

The second challenge is slower-than-expected customer adoption. Capacity may ultimately be utilised, but not at the pace required to support the investment case. This issue can extend timelines, delaying revenue generation and placing pressure on project economics. Delayed customer adoption, whether due to economic conditions, regulatory constraints or shifts in AI development priorities, can leave newly built facilities underutilised for extended periods. 

Rethinking deployment models 

How does the industry respond to these risks? In short, we must rethink how infrastructure gets deployed. Instead of committing to large, monolithic builds, there is a growing case for more upgradeable and phased approaches. Upgradeable data centres, built offsite with advanced manufacturing processes, allow operators to align capacity deployment more closely with confirmed demand. This technique reduces the risk of overbuilding; organisations can start small and add capacity as required, rather than building for a future that may not play out in the way expected. Delayed customer adoption suddenly becomes less of a concern. Critically, upgrades can also be performed while the rest of the site remains operational, avoiding prolonged downtime and revenue loss. 

As AI hardware evolves with increasing power densities and shifting cooling requirements, this upgradability also allows infrastructure to adapt and remain relevant as technology changes unfold. AI is evolving much too fast for rigid designs and, instead, flexibility should be built in to ensure this infrastructure can cater for the needs of tomorrow. In other words, we should be manufacturing building blocks that have the capability to evolve and scale, therefore reducing risk around utilisation. Rather than designing and building an AI factory that works only for the hardware we know, developers should be looking at power and structural systems that can scale to higher densities, or support a transition to new cooling methods, without the need for a full re-build or retro fit.  

From a financial perspective, this level of upgradability means the model can shift from large upfront capital commitments towards more incremental investment. Developers can then scale in line with both technological developments and customer demand. 

Demand-led infrastructure 

How effectively capacity is matched to real, sustained demand will be crucial in the coming years. The risk of building without guaranteed users is not new, but the scale at which it is now occurring is. The industry must quickly learn the art of balancing speed with discipline. 

Developers should align deployment with evolving demand, while also adapting flexibly. Understanding who is committed to using these facilities is critical to determining the sustainability of the AI boom. But it’s also key that we design facilities not just for today’s technology, but for tomorrow’s unknowns, ensuring they can evolve at the same pace as AI itself. 

Learn more at infrapartners.llc