The scale of big tech’s billion-dollar investments in AI leaves little doubt about what’s regarded as the most transformative technology of our time. What makes it more than an expensive hunch is the proliferation of business use cases, with AI embedded everywhere, from production lines and supply chains to the back office and customer service. However, one aspect of this technological revolution is far from settled.

The pressure that power-hungry AI servers place on data center infrastructure is well documented, whether it’s Goldman Sachs predicting a 165 percent increase in demand by 2030 or the US Department of Energy weighing up the pros and cons of nuclear power as the solution. While small nuclear reactors might be viable for hyperscalers who are already developing these, other operators will look for a more practical way forward.

The good news is that there are ways to retrofit a data center to support the demands of AI servers. It’s important to remember that the fundamentals of designing and operating a data center stay the same – you’re still working within budgetary constraints around space, power, and cooling. What changes dramatically is how you configure the cooling and power infrastructure to give AI servers the resilience they need.

Three key strategies to consider:

Optimize liquid cooling

Research shows that AI workloads dramatically increase power density in data centers, typically exceeding 120-136kW per rack – several times that of traditional servers – and often require a hybrid cooling approach, with 22-25 percent air cooling and 75-78 percent liquid cooling. This results in higher heat output and sometimes requires advanced cooling solutions, such as liquid cooling.

Effective systems include direct-to-chip liquid cooling supported by coolant distribution units (CDUs) engineered for long-term reliability, safety, and efficiency. This can involve using treated water to prevent corrosion and conductivity problems, routed through robust infrastructure such as overhead steel piping. In well-developed setups, most chilled water may be allocated to AI servers, with a smaller portion reserved for air-based cooling systems that support ancillary equipment like networking gear.

Upgrade the electrical distribution system

AI accelerators are highly sensitive to power quality. Sub-cycle power fluctuations can cause bit errors, data corruption, or system instability. Older uninterruptible power supply (UPS) systems may struggle to handle the dynamic loads AI can produce, often involving three MW sub-cycle swings or more.

Updating the electrical distribution system (EDS) is an opportunity that includes replacing dated UPS technology, which often cannot handle the dynamic AI load profile, redesigning power distribution for redundancy, and ensuring that power supply configurations meet the demands of high-density computing.

For instance, it may be necessary for the server chassis to operate with six out of eight power supplies active rather than traditional A / B rack power distribution. Backup generators and updated battery systems also play a crucial role in ensuring power stability and uptime.

Deploy an energy and power management system

With the high cost of AI downtime, risk mitigation becomes paramount. Energy and power management systems (EPMS) are capable of high-resolution waveform capture, which allows operators to trace and address electrical anomalies quickly. These systems are essential for identifying the root cause of power quality issues and coordinating fast response mechanisms.

For example, without a responsive EPMS during power transients, cooling systems in high-density setups may fail to react quickly enough, resulting in hardware stress. Integrating EPMS with building management tools ensures real-time responsiveness and supports a proactive stance on energy efficiency, stability, and uptime assurance.

One size doesn’t fit all

No two mission-critical facilities are the same regarding space, power, and cooling. Add the variables of each AI deployment, and what works for one facility may not be the best fit for another.

That said, there are some universal truths about retrofitting for AI. You will need engineers who are well-versed in various equipment configurations, including cooling and electrical systems connected to the network. You should plan beyond the initial design to reduce future risk and operational expense, and you will need close working relationships with vendors. This is a fast-moving sector with constant innovation, and market knowledge is essential.

At Ascent, we have the experienced teams and partnerships to take our customers to the leading edge of AI deployments, and – crucially – keep them there.

Learn more about Ascent.