Originally Posted At DCD

A proactive approach is essential, or risk falling behind in this AI race

Graphics Processing Units (GPUs) are pushing the limits of data center power infrastructure. AI workloads are driving the desire for GPU clusters that can deliver immense parallel processing power, as hyperscalers, colocation providers, and enterprise operators are feeling the strain.

At the same time, thermal demands have become a primary concern, and retrofitting and scaling data centers to support these workloads have emerged as a significant hurdle. The AI revolution driving GPU adoption means that uninterruptible power supply (UPS) strategies must be a top IT and boardroom priority.

GPUs are fueling the shift

AI, machine learning, and high-performance computing have revolutionized data centers, and with this advancement comes the issue of GPUs. A GPU is an electronic circuit designed to rapidly manipulate and alter memory to accelerate the creation of images, videos, and other visual content. Because they can perform multiple calculations simultaneously, GPUs are utilized in machine learning, video editing, and high-speed advanced computations. This leap in capability comes at a cost. GPUs demand more power and generate significantly more heat than traditional processors.

A Central Processing Unit (CPU) was once the primary processor for handling general tasks. However, legacy systems built for predictable, low-density CPU workloads are struggling to adapt. This shift to GPU-intensive computing is altering data center design and power supply strategies. Today, GPU racks can consume up to 700W, with some requiring 1,200W or more.

Most existing UPS systems, built before the AI explosion, weren’t designed for these new workloads. As a result, these older, legacy systems lack the capacity, responsiveness, and scalability required for runtime and reliability under modern GPU loads.

Power supply constraints 

Over the past five years, AI workloads have grown more than tenfold. This growth necessitates GPU clusters that can deliver immense parallel processing power, but they also incur increased power and thermal costs. 

The increase in workloads has been evidenced in AI inferencing, which uses a trained AI model to generate outputs from new, unseen data. Inference is critical for deploying AI in real-world applications. 

Take, for example, Meta’s announcement of a $1 billion upgrade to its AI infrastructure. Plans will include the need for dedicated GPU clusters. Upgrades of this magnitude require scaling power demands in hyperscale environments. The ability to handle this demand is crucial. 

It’s not just hyperscalers that are being pushed; colocation providers and enterprise operators are also feeling the strain. Power density has become a critical constraint. UPS runtimes are shrinking dramatically under GPU loads. During an outage, the average coverage was 10 to 15 minutes; however, older systems are now having difficulty achieving three to five minutes of coverage. This indicates that redesigns are needed.

Thermal demands are now a whole other consideration, often superseding even the best cooling systems. The heat from GPU racks is overwhelming airflow models designed for older, lower-power deployments. Without proper cooling or intelligent load-shedding capabilities, UPS systems could become failure points.

The nightmare for digital infrastructure managers is that these problems tend to show up after GPUs are deployed. Retrofitting a power infrastructure can cost two to three times more than planned upgrades and increase the risk of service disruptions. Unplanned outages can average more than $100,000 per incident, with the cost attributed to cascading power or cooling failures.

Solving these issues isn’t as simple as replacing old hardware. A future-ready UPS strategy requires holistic thinking. Designs must be scalable.

UPS strategies need to be a boardroom decision

Solving these issues isn’t as simple as replacing old hardware. A future-ready UPS strategy requires holistic thinking. Designs must be scalable. This means utilizing modular UPS units that can scale with GPU loads while integrating real-time power monitoring and load balancing to handle fluctuating demand.

This shift isn’t just a facilities conversation; it’s a board-level decision. This includes a plan for aligning power supply concepts with thermal strategies such as liquid cooling and hot aisle containment. A good example of proper planning is Google’s AI-dedicated data centers, which pair advanced UPS configurations with automated load-shedding systems. This allows for adjustments to power distribution without impacting service. 

Giving AI the energy it needs requires more than computing horsepower. Innovative infrastructure is needed at every layer. Without resilient, scalable power, even the most advanced AI stack can’t operate at its full potential. 

Proactive approach

The GPU revolution is reshaping data center design, and legacy UPS strategies are no longer adequate. Predictable CPU workloads are being replaced by GPU clusters that demand entirely new approaches to power and thermal management.

GPUs are calling for a complex transformation that really can’t be accomplished as a standalone upgrade. Partnerships are required with data center operators and builders that are leaders in power infrastructure innovation. What’s needed is not just advanced UPS hardware but also the strategic expertise to design modular, scalable systems that anticipate AI workload growth rather than react to it.

Operators who take a proactive approach will be the ones powering tomorrow’s AI breakthroughs. Those who dither will realize that they are facing expensive retrofits and risk service disruptions. 

The question isn’t whether you should act, but rather how quickly you can align with the right strategic partners to future-proof your infrastructure. The AI revolution isn’t waiting. You either adapt or become a footnote in history.