Executive summary
- The ‘comfortable financial rhythm’ of data centre economics is old news, faced with dual pressures of hungry AI workloads and volatile European energy markets.
- Now we’re past the ‘initial boom’, it’s time to face a new ‘commercial playbook’ considering alternative cooling systems, battery storage, energy, and how we charge for it.
- Code may dictate how smart a hybrid cloud can be but physical engineering will dictate whether it can actually run.
If software-defined architecture is the brain of modern AI data centres, then the physical infrastructure is the digestive system and oh boy, is the system struggling with severe indigestion.
The AI conversation over the last few years has been focused on software, the orchestration layers, virtualisation and algorithmic efficiency. However, whilst software controls optimise where the workload is, it can’t alter the laws of thermodynamics; AI chips demand relentless power, cooling flow and uninterrupted physical grid capacity.
We’ve done the experimentation with AI, run away with it a lot, but the unconstrained capital expenditure is over; the harsh realities of facility economics have settled in this year.
During the initial boom, operators managed compute loads that were more episodic, because training large language models (LLMs) needs gigantic spikes in power, but they had clear start and end dates, so facilities could schedule them around grid availability.
This has now changed as we shift from training LLMs to inference. Enterprise AI has left the nest, moved out of R&D into mission-critical, live operations. So now, real-time inference, which powers automated financial decisioning, autonomous logistics and instant multi-modal queries, now runs 24/7, 365 days of the year.
Gone are the scheduled spikes in power, gone is the “breathing room” when spikes weren’t occurring, and now there’s almost constant power withdrawal. This now flattens the usage curves that utilities traditionally use to manage grid balance.
It also means extreme heat is generated from continuous power intake, and if there is anything as much as a brief lag in the cooling system, it could prevent a complete shut down which is incredibly costly.
The industry is building on shifting sand
Data centre economics have worked like clockwork; a secure concrete shell is built, fitted with air vents and power cables, and leased out to corporate tenants for 15 to 20 years. Tenants brought in servers, swapped them every five years and paid a predictable, stable monthly rent.
This comfortable financial rhythm is old news. The current processing chips have been evolving so quickly that their power and cooling requirements double every 18 to 24 months, so now the operator of a state-of-the-art facility built three years ago is watching tenants demand four times the capacity in the same space.
If floor structure and internal piping can’t be adapted, then the operators are left with expensive digital real estate that is thermally obsolete, and the mortgage isn’t even paid.
The new commercial playbook for hybrid leaders
Cloud and collocation directors have to throw out the playbook; it’s outdated and isn’t going to help survive the dual pressures of hungry AI workloads and volatile European energy markets. The new commercial playbook looks a little like building modular plug-and-play spaces, essentially building facilities like Lego blocks instead of permanent, rigid layouts that can’t adapt.
This is so that air-cooled rooms can be upgraded to liquid cooling solutions without disturbing existing cloud tenants. Battery energy storage systems also play a key role and need to be installed on-site so energy can be stored when grid prices are low and be directed to be used to power heavy AI tasks during unavoidable peak hours.
Finally, modern hybrid contracts now have to prices around heat management, thus moving away from charging clients by the square metre or total wattage; this new shift seeks to guarantee how quickly and efficiently a facility can remove heat away from high-density hardware.
Software innovation grabs headlines and captures the imagination of the public and it will always be that way, but in boardrooms, it’s clear that code dictates how smart a hybrid cloud can be but physical engineering will dictate whether it can actually run. Thought leaders need to master the art and economics of feeding the beast.
