How AI is reshaping data centres

There is a new reality rising. The limiting factor in AI data centers right now is not software. It is not silicon. It is the power, cooling, and physical infrastructure required to support AI at scale. 

Here are some startling facts. When GPT-4 launched in March 2023, processing one million output tokens cost around $60. By the end of 2024, models such as DeepSeek V3 were delivering GPT-4o-equivalent quality for approximately $1.10. Today, that same level of output can cost as little as $0.50 with models like DeepSeek V4, representing a cost collapse of more than 100x in under two years. When intelligence becomes this cheap to generate, people use exponentially more of it, and the pressure lands squarely on the physical world. 

That is the reality we are now living in. And it requires a new type of data center. One where chip physics, power delivery, cooling, and operations must be orchestrated as one system: the AI factory.  

Building around the chip 

For years, data centers were designed around workloads. You understood the applications you needed to run, estimated the compute requirements, and designed your power and cooling environment accordingly. Then AI came along and inverted that logic entirely. 

A decade ago, average rack densities sat comfortably between 2 and 5kW. AI environments today routinely exceed 30kW per rack, with some training clusters already pushing past 100kW. NVIDIA's roadmap points toward architectures approaching 600kW later this decade. At those densities, the facility is no longer a container for compute. It becomes part of the compute equation itself. 

Today, the processor dictates how the facility must be built around it. The power draw, thermal profile, and density of modern GPUs and AI accelerators are reshaping infrastructure decisions from the ground up. 

Suddenly, the mechanical and electrical engineers are having conversations that used to belong exclusively to the compute team. Cooling design becomes critical to system performance, power architecture becomes a competitive differentiator, and mechanical and electrical engineering decisions carry direct implications for AI capability, uptime, and economics.  

Mastering the orchestration 

An AI factory is a data center, but it produces intelligence as opposed to just storing and processing. It's a data center, but with a whole new challenge of scale and density ...a data center, but not as we know it (for the Trekkies out there). 

The relationship between compute, cooling, and power in these environments has become so tightly coupled that failures propagate at a different speed. In a conventional cloud environment, a cooling failure might have given operators twenty or thirty minutes before systems shut down. In a dense GPU cluster, that tolerance can collapse to seconds. A cooling issue is now an immediate operational event with direct revenue consequences.  

Most AI factories will also not operate as a single standardized environment. Operators will be managing mixed-density architectures simultaneously, with some zones running below 50kW and others exceeding 100kW, each with different thermal and power characteristics. Air cooling is rapidly approaching its practical limits for high-density AI workloads. Liquid cooling, including direct-to-chip and immersion technologies, is becoming an operational necessity. In many environments, the cold plate now defines GPU performance and reliability more than almost anything else. 

I think of modern AI infrastructure less as a capacity challenge and more as an orchestration challenge. Organizations that move from having the largest GPU footprint to focusing on synchronizing power systems, cooling loops, workload distribution, and operational controls as one coherent, integrated environment will define the next era of intelligence.  

Digital twins: from chip to grid 

The traditional approach to infrastructure deployment was linear: design, build, commission, optimize. That model is too slow for the AI era. AI factories are under pressure to deliver operational capacity in months, not years. Operators need to understand how entire environments will behave thermally, electrically, and operationally before construction begins. Digital twins make that possible, but only when they are built to the right scope. 

It is no longer sufficient to model only the white space inside the data hall. The power architecture, cooling loops, building management systems, and grid interactions all need to be represented as one synchronized operational model. That is what moves infrastructure management from reactive to predictive, and predictive operational intelligence is increasingly your competitive advantage. This is how you will protect your investment. 

The new era of AI infrastructure  

There was a world before electricity and a world after it. Industries were not simply changed by the existence of electricity, but by the organizations that learned to operationalize it at scale. 

AI is beginning to feel similar, and the companies that get there first will not necessarily be the ones with the best models. They will be the ones who figured out how to power, cool, and run them reliably at scale, day after day, without the wheels coming off.