Most of the world's compute does not run in new buildings. It runs in facilities designed for a different decade, and in industrial buildings now being converted to hold racks they were never drawn for. Making those sites sustainable is a different problem from designing a new one on a clean sheet, because every upgrade has to fit inside decisions someone else already made.
This is the playbook for that problem: what to measure, what to change first, where AI densities break a legacy envelope, and how to decide when the building is worth keeping. We have written separately about the self-powered AI data center, which covers the off-grid definition, why water is a siting liability and how two-phase cooling works. This piece does not repeat that. It is about the retrofit decision.
What Should You Measure Before Retrofitting a Data Center?
Three ratios, all defined by The Green Grid, and all normalized to the energy used by the IT equipment itself. That shared denominator is the point: each one asks how much overhead, water or carbon it takes to deliver one unit of useful compute energy.
| Metric | Definition | Units | A sourced reference point |
|---|---|---|---|
| PUE Power usage effectiveness | Total facility energy ÷ IT equipment energy. Introduced by The Green Grid in 2007. | Dimensionless (kWh/kWh) | 1.54 global weighted average, 2025, flat for six years (Uptime Institute). US average fell from 1.6 in 2014 to 1.4 in 2023 (LBNL). |
| WUE Water usage effectiveness | Site water consumption ÷ IT equipment energy. The Green Grid, 2011. | Liters per kWh | US data centers consumed 66 billion liters directly in 2023, against nearly 800 billion liters indirectly through electricity generation (LBNL). |
| CUE Carbon usage effectiveness | CO2-equivalent emissions caused by the facility's total energy use ÷ IT equipment energy. The Green Grid, 2010. | kg CO2e per kWh | US data-center electricity averaged 0.34 kg CO2e/kWh in 2023 (LBNL). At a PUE of 1.4 that implies a CUE near 0.48, on our arithmetic. |
Sources: The Green Grid's white papers on WUE (WP#35, March 2011) and CUE (WP#32, December 2010); the Uptime Institute Global Data Center Survey 2025; and Lawrence Berkeley National Laboratory's 2024 United States Data Center Energy Usage Report.
Two cautions before anyone uses these numbers in a board deck. First, boundaries change the answer. LBNL separates site WUE, the water the facility itself consumes, from source WUE, which adds the water consumed generating its electricity. In its 2023 US estimate the second is about twelve times the first.
Second, the metrics trade against each other. LBNL notes that evaporation-based cooling is generally more energy efficient than waterless alternatives, and that air-cooled chillers avoid on-site water use but draw more energy. A retrofit that improves WUE by worsening PUE has moved the problem, not solved it.
Uptime adds a third caution that matters for credibility: sustainability data collection has stalled. In 2025, 74% of respondents tracked PUE, but worldwide only 47% collected water consumption data and 42% renewable energy usage. A facility that measures all three ratios properly is already ahead of most of its peers.
Can an Existing Data Center Be Made Sustainable?
Yes, and history says it is where most of the industry's efficiency came from. Uptime's survey shows the global average PUE dropping from 2.5 in 2007 to 1.65 in 2014, as operators upgraded older electrical systems, improved airflow management and optimized cooling controls: changes that were straightforward and did not interrupt live workloads.
That era is over. The 2025 weighted average of 1.54 is, in Uptime's words, the sixth consecutive year the figure has “virtually stood still”, and the survey is blunt about why: in older legacy facilities, further upgrades “often present a much weaker business case.” Legacy is not a niche either. Over a quarter of respondents, 27.5%, work in facilities 16 years or older.
Meanwhile the newest buildings show what the envelope allows. Facilities commissioned in the last five years averaged a PUE of 1.48, those of 20 MW and above averaged 1.44, and many new high-latitude facilities report 1.3 or better. Uptime also notes that in several major regions new regulations are mandating lower PUEs, forcing some older facilities “to either close or undertake major retrofits.”
The honest answer is that an existing site can be made sustainable, but not by tuning. The remaining gap is structural: how heat leaves the rack, how power reaches it, and where that power comes from. Those are architecture decisions, and they have to be made in order.
What Should You Retrofit First in a Legacy Data Center?
The order below is not a preference. Each step sets the bounds for the next, and the most expensive retrofit mistakes we see come from buying a later step before an earlier one was answered.
Audit and baseline
Meter PUE, site WUE and CUE across at least a full seasonal cycle, with the boundary written down. Survey the structure, the electrical room, ceiling heights and routes for liquid distribution. Without a baseline, no later improvement can be claimed, and without the physical survey, no later design can be trusted.
Power headroom
Establish how much power the building can actually take and distribute today, and what it would take to raise that. This is the binding constraint on everything after it. A cooling upgrade sized for racks the electrical system cannot feed is stranded capital.
Cooling: air, then liquid, then two-phase direct-to-chip
Cooling is the largest controllable overhead. The IEA puts it at about 7% of consumption in efficient hyperscale facilities and over 30% in less-efficient enterprise ones. Airflow containment is the first rung, single-phase liquid the second. LBNL notes that liquid-cooled IT can run at higher coolant temperatures, which extends economizer hours.
Two-phase direct-to-chip with a low-pressure dielectric refrigerant is the rung our designs use, and it takes evaporative cooling water out of the loop entirely.
Power distribution
Rebuild distribution for the density the cooling now supports. Our designs use Class 4 fault-managed power distribution. Whatever the topology, it has to be sized for the next hardware generation, not the one being installed.
On-site generation
Once load and overhead are known, size generation to them. This is the step that decouples growth from the utility service the building inherited.
Heat reuse
Rejected heat is an output with a use if the architecture captures it at a useful temperature. Liquid cooling helps because it moves heat at higher temperatures than air. In our energy system, generator exhaust heat is routed back to drive the on-site reactor: heat out, fuel in.
Monitoring
Instrument all three ratios continuously, on the same boundaries as the baseline. The retrofit is only as credible as the measurement that proves it, and continuous data is what lets an operator catch a drifting setpoint before it becomes a year of lost efficiency.
The IEA figures in step 03 are from Energy and AI (IEA, 2025). The cooling principles are covered in depth in our post on the self-powered data center, including when single-phase is the better choice.
Planning a retrofit, or deciding whether a building is worth keeping? We act as the design authority across the whole facility and coordinate the partners who build it.
See Our Solution Architecture Service →Can a Legacy Data Center Be Upgraded for AI Rack Densities?
Some can. Most cannot without work that touches the structure, the electrical room and the cooling plant at once. The Uptime Institute's 2025 survey of more than 800 owners and operators found “greater adoption of racks in the 10–30 kW range” and that “few facilities exceed 30 kW” (Uptime Institute, 30 July 2025). A current-generation AI rack, NVIDIA's GB200 NVL72, is rated at 132 kW (Uptime Intelligence, 25 June 2025). Our own design targets about 250 kW for next-generation NVL72-class racks.
That is not an increment on the installed base. It is roughly an order of magnitude, and it arrives through three systems at once: heat that air cannot carry, current that legacy distribution was not sized for, and weight the floor was not rated for. We work through the per-rack arithmetic in how much power an AI data center needs.
There is a legitimate alternative when the building cannot get there. Uptime documents operators deliberately de-escalating density: 64-GPU racks held within 80 to 90 kW, four 8-GPU frames per rack brought under 50 kW, and further configurations below 20 kW, using readily available off-the-shelf distribution equipment. That runs accelerated compute in a legacy envelope. It runs less of it per cabinet, and it is a decision about how much of the building's future you are willing to cap.
How Much Energy and Water Can a Data Center Retrofit Save?
It depends entirely on where the facility starts, which is why we will not quote a percentage without a baseline. What can be stated is the range the sources support, and the arithmetic that follows from it.
| Reference point | PUE | Facility load for 10 MW of IT | Overhead vs. 1.54 |
|---|---|---|---|
| 2025 global weighted average (Uptime) | 1.54 | 15.4 MW | – |
| Facilities commissioned since about 2020 (Uptime) | 1.48 | 14.8 MW | −0.6 MW |
| Facilities of 20 MW and above (Uptime) | 1.44 | 14.4 MW | −1.0 MW |
| Many new high-latitude facilities (Uptime) | 1.30 or better | 13.0 MW | −2.4 MW |
On our arithmetic, a 10 MW IT load moved from 1.54 to 1.30 sheds 2.4 MW of continuous overhead, about 21,000 MWh a year at constant load. That is the size of the prize a cooling and distribution retrofit is competing for. Whether a given building can reach 1.30 depends on its climate, its heat rejection options and how much of the envelope is replaced.
Water is harder to generalize and easier to mis-state. LBNL estimates US data centers consumed 66 billion liters directly in 2023, and much of that direct figure is evaporative heat rejection, which is the term a closed two-phase loop removes.
The peer-reviewed work by Li, Yang, Islam and Ren, Making AI Less “Thirsty”, makes the same distinction between withdrawal and consumption, and between on-site and off-site water, that a retrofit claim needs to respect. That is why the WUE baseline in step 01 has to state its boundary before any saving is claimed against it.
Should You Retrofit or Rebuild a Legacy Data Center?
On four things the existing building either has or does not. The table is our judgement, drawn from how these projects are scoped rather than from a published dataset, and each row is a question to answer at the audit stage.
| Decision factor | Favors retrofit when | Favors new build when |
|---|---|---|
| Structure | The floor and roof can take liquid-cooled racks and the plant that serves them, or can be reinforced locally. | Floor loading or clear height rules out AI racks across most of the building. |
| Power path | There is electrical headroom, or room on site for generation and distribution to grow. | The electrical room, the service and the distribution all need replacing at once. |
| Cooling routes | Liquid distribution can reach the racks without rebuilding the hall. | Getting liquid to the rack means gutting the space. |
| Schedule | The shell, the permits and the site access save time over starting fresh. | The retrofit scope is large enough that a purpose-built pod would be operating sooner. |
| Regulation | The retrofit can meet efficiency and reporting rules on the existing footprint. | Rules triggered by a substantial refurbishment make the old envelope uneconomic to bring up to standard. |
The regulation row is not hypothetical in every market. Under the EU's Energy Efficiency Directive, newly planned or substantially refurbished data centers with a total rated energy input above 1 MW must carry out a cost-benefit analysis of waste heat recovery, and operators with installed IT power of at least 500 kW must publish energy and water information (Covington & Burling, April 2025).
Many projects do not have to choose. Our 73 MW self-generating campus in North Texas, currently in development, is designed to combine an existing industrial building with new self-powered 10 MW pods: the building keeps the work it is suited to, and the pods carry the density it is not. In our model, the first pod takes 18 to 24 months and each subsequent pod 6 to 8.
Making that call well is a whole-facility question, which is why we act as solution architect and design authority across retrofits and new builds alike. We are not the general contractor. We set the architecture and coordinate a network of 165 partners and businesses that execute it, across construction, power generation, cooling, electrical distribution, cyber and NOC, network, compute hardware, water and environmental, and permitting and engineering. The method is on our data center solution architects page.
On-site generation is step 05 of the sequence, and at a legacy site it is the step that lets new load arrive on the building's schedule rather than the utility's. The off-grid, self-generated power model we design, how its fuel paths work and what it earns beyond compute, is covered in The Self-Powered AI Data Center, with the service detail on our on-site hydrogen power page.
Frequently Asked Questions
Yes, and most of the industry's efficiency gains so far came from exactly that. The Uptime Institute reports that global average PUE fell from 2.5 in 2007 to 1.65 in 2014 as operators upgraded older electrical systems, improved airflow management and optimized cooling controls. The easy gains are now mostly taken: the 2025 weighted average is 1.54, flat for the sixth consecutive year, and Uptime notes that further upgrades in older legacy facilities often present a much weaker business case. So the answer is yes, but the next increment comes from structural changes to cooling, power distribution and energy supply, not from tuning, and it has to be sequenced so each step does not strand the next.
Three ratios defined by The Green Grid. Power usage effectiveness, PUE, is total facility energy divided by IT equipment energy, so 1.0 would mean no overhead at all. Water usage effectiveness, WUE, is site water use in liters divided by IT equipment energy in kWh. Carbon usage effectiveness, CUE, is the carbon dioxide equivalent emissions caused by the facility's total energy use divided by IT equipment energy. Each is only meaningful with its boundary stated: site versus source water, for example, differs by roughly an order of magnitude in LBNL's 2023 US estimates. A retrofit claim that does not name its metric, its boundary and its baseline period is not a measurable claim.
Measurement, then power headroom, then cooling. You cannot claim an improvement you did not baseline, so the first step is metered PUE, site WUE and CUE over at least a full seasonal cycle. The second is an honest count of electrical headroom, because every later decision, including cooling, is bounded by how much power the building can actually take and distribute. Cooling comes third because it is the largest controllable overhead: the IEA puts cooling at about 7 percent of consumption in efficient hyperscale facilities and over 30 percent in less-efficient enterprise ones. Distribution, on-site generation, heat reuse and continuous monitoring follow, in that order.
Some can, most cannot without major work. The Uptime Institute's 2025 survey found greater adoption of racks in the 10 to 30 kW range and few facilities exceeding 30 kW, while a current AI rack such as NVIDIA's GB200 NVL72 is rated at 132 kW and our own design targets about 250 kW for next-generation racks. That gap is not closed by adding cabinets. It needs liquid cooling at the rack, distribution sized for the load, and floor capacity for the weight. Where a building cannot get there, operators can deliberately run lower-density configurations, which Uptime documents at under 50 kW and even below 20 kW, at the cost of less compute per cabinet.
It depends on whether liquid can reach the rack without rebuilding the hall. Cooling upgrades in a legacy building climb in rungs: airflow containment first, then single-phase liquid, then two-phase direct-to-chip with a low-pressure dielectric refrigerant, the rung our designs use, which takes evaporative cooling water out of the loop entirely. Lawrence Berkeley National Laboratory notes that liquid-cooled IT can run at higher coolant temperatures, which extends economizer hours. The audit decides the rest: whether the floor can take liquid-cooled racks and the plant that serves them, and whether the building has the power headroom to feed the density the cooling supports. If getting liquid to the rack means gutting the space, that favors a new build.
It depends on where the facility starts, so any single percentage is a guess. The sourced benchmarks are these: the 2025 global average PUE is 1.54, facilities commissioned in the last five years average 1.48, facilities of 20 MW and above average 1.44, and many new high-latitude facilities achieve 1.3 or better. On our arithmetic, taking a 10 MW IT load from 1.54 to 1.30 removes 2.4 MW of continuous overhead, about 21,000 MWh a year at constant load. For water, Lawrence Berkeley National Laboratory estimates US data centers consumed 66 billion liters directly in 2023, mostly through evaporative heat rejection, which is the term a closed liquid-cooling loop removes. Note the trade-off LBNL flags: waterless heat rejection can use more energy, so PUE and WUE have to be optimized together.
Decide on four things the existing building either has or does not: structural capacity for liquid-cooled racks, electrical headroom or a credible path to it, room to route liquid distribution, and a schedule advantage over building new. If the shell is sound and the power path exists, a retrofit usually reaches revenue sooner. If the floor, the electrical room and the utility service all need replacing, you are paying new-build costs inside an old envelope. Many projects land in between: keep the building for what it does well and add new, purpose-built capacity beside it, which is how the 73 MW North Texas campus we have in development is designed.