The expansion of artificial intelligence is forcing a rethink of what to do with data centers built before the arrival of high-density GPUs. In many cases, the next leap in capacity doesn’t come from building a new facility from scratch, but from upgrading existing sites with more electrical power, new cooling solutions, higher-capacity networks and systems ready for AI workloads.
Data center retrofits in 20 seconds
- Existing facilities can gain AI capacity through power, cooling and connectivity upgrades.
- Liquid cooling allows higher-density racks to be introduced without necessarily rebuilding the entire facility.
- Google has already developed Brazos to bring liquid cooling to facilities originally designed for air cooling.
- Schneider Electric frames the decision as a comparison between retrofitting, outsourcing and new construction based on power, costs and timelines.
- The main constraints remain available electricity, cooling, space, structure and grid connection timelines.
The idea of turning older facilities into AI-ready infrastructure is gaining traction for a simple reason: building new capacity takes time. Securing land, electricity, permits, equipment and grid connections can become a bottleneck while demand for compute keeps growing.
Schneider Electric frames this exact decision as a choice between retrofitting, outsourcing or building from scratch. Its 2026 analysis notes that certain existing facilities can be adapted for AI workloads if they have an adequate baseline of power, cooling, space and connectivity.
A retrofit doesn’t necessarily mean tearing down an entire data center, either. In some cases it can mean working on specific racks, adding liquid cooling, increasing a room’s electrical capacity, or gradually modifying distribution systems.
Cooling is the biggest change for AI data centers
The arrival of increasingly powerful accelerators is raising the thermal density of racks. Facilities designed for conventional servers can fall short once GPU clusters are introduced, packing far more power into the same space.
Schneider Electric notes that the most demanding AI workloads require greater power and cooling capacity and, in some cases, liquid cooling. That doesn’t mean every AI server needs liquid cooling, but it does mean facilities that want to host high-density configurations need to look at alternatives to conventional air cooling.
One such alternative is direct-to-chip cooling. There are also rear door heat exchangers, known as RDHx, coolant distribution units (CDUs), and systems that transfer heat to water loops.
The goal is to let an existing facility support more power per rack without having to completely rebuild its climate control system.
Google has taken this idea a step further with Brazos, a liquid cooling system designed to introduce liquid-cooled equipment into facilities that originally use air cooling. The approach allows installation rack by rack, avoiding the need to convert the entire infrastructure at once.
This is especially useful for operators who can’t take a live facility offline. A phased retrofit allows capacity to be added while the rest of the facility keeps running, although each project depends on the specific characteristics of the site.
Retrofitting isn’t always better than building new
Retrofitting has an obvious advantage: it can make use of assets that already exist. But not every data center is ready to take on high-density AI workloads.
Electrical availability is one of the first things to check. Cooling capacity, power distribution, floor load capacity, ceiling height, backup systems and available space for new equipment all matter too.
Schneider Electric insists the decision has to start from the characteristics of the workload. Infrastructure intended for AI inference can have different requirements than a cluster dedicated to model training.
That’s why a retrofit can be a quick fix in a building with a good electrical connection and cooling capacity, but unattractive in an older facility that requires rebuilding nearly all of its critical systems.
The electricity problem has become even more pressing as data centers keep growing. In the United States, for example, the Federal Energy Regulatory Commission (FERC) has warned about the impact of growing demand from large consumers, with data centers among the factors putting pressure on certain power grids.
Power availability can even determine whether a retrofit project makes sense at all. Adding thousands of GPUs to a room doesn’t help much if the facility doesn’t have enough contracted power and distribution capacity.
Microsoft, Google and the market for existing facilities
Big tech companies are ramping up their investments in AI infrastructure, but that doesn’t mean all of that capacity will come from brand-new buildings.
Microsoft, for example, said during its fiscal 2026 results that its investments include both short-lived assets, such as GPUs and CPUs, and long-lived assets for data centers. The company also extended the estimated useful life of its data centers and office buildings from 15 to 25 years, starting with fiscal year 2027.
That accounting change isn’t itself a data center retrofit program, but it does reflect the economic importance of maintaining and using existing facilities for longer.
Google is also working directly on this problem. Its Brazos system addresses a specific need: introducing high-density liquid cooling without having to fully convert a facility designed for air cooling.
At the same time, data center operators and infrastructure specialists are developing solutions to modernize existing facilities. Equinix and Digital Realty are among the operators with large portfolios of facilities that may need power, cooling and density upgrades as their customers’ needs change.
The infographic accompanying this information places Microsoft, Google, AWS, Meta, IBM, Oracle, Digital Realty, Equinix, Dell, HPE, Schneider Electric, Vertiv, Cisco, Huawei and Johnson Controls among the leading companies tied to data center retrofits. However, the specific figures shown in the chart, such as the $135.2 billion attributed to Microsoft or the $128.5 billion attributed to Google, aren’t backed by the public sources consulted as specific investments in data center retrofitting. For that reason, it wouldn’t be accurate to present them as a confirmed financial ranking.
What is backed up is the trend: existing infrastructure is gaining value because it can deliver capacity faster than a brand-new project when the right conditions are in place.
From conventional data center to AI-ready facility
A data center retrofit doesn’t stop at the GPUs, either. A project like this can touch power supply, UPS systems, distribution, cooling, cabling, networking, monitoring and automation.
The shift toward higher-density racks also requires rethinking how power is distributed and how heat is removed. In certain facilities, a hybrid solution can make more sense than replacing the entire cooling system at once.
The concept of a digital twin also comes into play, used to model facilities and study changes before implementing them physically. In complex projects, it can help plan retrofit phases and reduce the risk of interfering with systems that remain in production.
The result is a new category of infrastructure that doesn’t quite fit the traditional divide between new and old data centers. A building constructed years ago can still be useful for AI if the right systems are upgraded and there’s enough electrical capacity.
The key is knowing how far that upgrade can go. When available power, structure or cooling can’t reach the densities needed, building a new facility can still be the more logical option.
Retrofitting is gaining ground because it makes use of buildings, connections and systems that already exist. In a race where securing electricity and compute capacity can take years, reusing infrastructure isn’t just about saving money: in some cases, it can be the fastest way to get AI compute up and running, a dynamic also visible in how Schneider Electric and AMD’s reference design for AI factories and the broader shift toward liquid cooling in AI data centers are reshaping what counts as AI-ready infrastructure.
Frequently asked questions
What is a data center retrofit?
It involves upgrading an existing facility to increase or adapt its capacity. It can include electrical systems, cooling, networking, racks, storage, monitoring and other infrastructure elements.
Why is AI driving data center retrofits?
The GPUs used for AI pack far more power and generate more heat than most traditional systems. This forces facilities not designed for those densities to adapt their power supply and cooling.
Can liquid cooling be installed in an older data center?
Yes, there are several approaches. Options include direct-to-chip cooling, coolant distribution units or rear door heat exchangers, among others, although feasibility depends on the existing infrastructure.
Is it better to retrofit a data center or build a new one?
It depends on available power, the facility’s condition, cooling, space, costs and the required timeline. A retrofit can be faster when the building already has good baseline infrastructure, but not every facility is suitable for high-density AI workloads.
