
What makes a data centre AI-ready?

Why adaptability matters more than capacity
A data centre is AI-ready when its power and cooling can keep up as rack power density rises. For an AI-ready data centre, the capacity available today matters less than actually how well the facility adapts. Four factors decide this:
- Power density: how many Kilowatts (kW) each rack can draw not only today, but also in three years time
- Cooling: capacity that scales with IT load, with a plan for when air cooling runs out
- Adaptability: a modular design that can be reconfigured without a lengthy rebuild
- Resource use: efficient energy usage and low water consumption
Today, if you ask a CIO what they need from a data centre, the answer usually starts with AI. Eduardo Espinel, Ilkari’s data centre lead in Colombia, has seen the change first-hand. Two years ago, customers in regulated sectors asked about uptime, certification and cost. Now they want to know whether a facility can support the capacity their AI workloads will need, and how soon.
Roughly a third of data centre operators already run some AI training or inference, according to Uptime Institute. Decisions that once had years of lead time now have to be made in months.
Key takeaways
- AI workloads change rack power density far more than floor space. In 2024, the average rack sat under 8 kW, while the largest AI racks exceed 100 kW.
- In-row cooling and modular data centre design let a high-density data centre add capacity in stages, so customers avoid paying for it too early.
- Ask providers how their design will handle the density you will need in three years.
Explore the article
- What is an AI-ready data centre?
- Data centre cooling for AI
- Modular data centre design and scalable colocation: building for change
- Sustainable data centre operations
- What AI readiness means for regulated organisations in Latin America
- Questions to ask an AI-ready data centre provider
- FAQs
What is an AI-ready data centre?
An AI-ready data centre is a facility whose power, cooling and network can support high-density AI workloads today. Equally important, it can also keep adapting as those workloads grow, without a lengthy rebuild. A facility that can only meet today’s needs will reach its limits sooner than its customers expect. A data centre built for both stays useful as AI demand rises.
AI hardware asks more of a site than standard enterprise servers. Each rack draws more power and gives off more heat, and GPU clusters depend on low-latency networking. Training is the most power-dense use. Inference, which runs trained models, is more distributed and focused on availability and latency.
Why rack power density matters more than floor space
Rack power density is the amount of power a single rack draws, measured in kW. It is a better guide to AI readiness than floor space, because AI workloads concentrate far more power and heat in each rack.
Given that nearly all facilities were designed for low-density racks, Eduardo argues that the real story behind the capacity headlines is power density, cooling and grid capacity.
Ilkari’s own figures show how fast this is moving. A rack that ran standard enterprise workloads at 5 to 10 kW two years ago is now expected to support 15 kW within the next 18 months. Ilkari is already planning for 25 to 30 kW within a few years.
“That is not an incremental change”, Eduardo says. Ilkari’s Tocancipá data centre already supports up to 30 kW per rack. The company is also doubling energy capacity per rack at the site.
Uptime Institute found average rack density was still below 8 kW in 2024, with growth since then mostly in the 10 to 30 kW range. At the far end, a single NVIDIA GB200 NVL72 rack draws 132 kW. Meaning that the bulk of the industry is still behind.
The majority organisations will land somewhere between those points, and nobody can say exactly when. A high-density, AI-ready data centre has to be ready for the whole range.
Given that extra generated heat previously mentioned, cooling methods are also a big consideration for AI readiness.
Data centre cooling for AI
Data centre cooling for AI has to remove far more heat per rack than cooling for standard enterprise IT. Air and in-row cooling handle racks up to about 20 to 30 kW. Above that, liquid cooling is usually needed.
Cooling technology keeps improving, which Eduardo notes makes high density possible without relying on cold climates.
While estimates vary, Schneider Electric’s 2023 design guidance put the practical ceiling for air at about 20 kW per rack. Other estimates sit closer to 30 kW.
Cooling approaches for an AI-ready data centre
| Approach | How it removes heat | Where it fits |
| Room air cooling | Cools the whole data hall | Lower-density racks |
| In-row cooling | Units placed between racks cool air close to the heat source | Up to the air-cooling threshold of about 20 to 30 kW, with units added as load grows |
| Rear-door heat exchangers | A liquid-cooled door removes heat as air leaves the rack | Denser racks that still use air-cooled servers |
| Direct-to-chip liquid | Coolant runs to cold plates on processors | Very high densities, such as a 132 kW GB200 NVL72 rack |
| Immersion | Servers sit in a tank of dielectric fluid | The highest-density deployments |
In every case, water usage depends mostly on how the heat is finally rejected from the building.
How in-row cooling scales with IT load
In-row cooling places cooling units between racks, close to the heat source, so capacity can be added unit by unit. Ilkari’s approach is to add cooling as customers grow. Most start with an initial setup. As their IT load increases, more in-row units are switched on, so cooling capacity stays matched to what is actually running.
The facility in Tocancipá runs N+1 cooling redundancy and has a design PUE below 1.4. While the lower the PUE, the better, Uptime Institute put the industry’s average PUE at about 1.56 in 2024, a figure that has barely moved for several years. A design figure is not directly comparable with a measured average, but it shows the efficiency the facility was built for.
When air cooling reaches its limit
As AI pushes density higher, air cooling will not cover every workload. The densest AI racks already need direct liquid cooling, and average densities keep climbing towards the air-cooling threshold.
Eduardo suggests buyers ask a direct question. As density increases, has the provider planned for the point where air cooling is no longer enough? And what cooling technology and infrastructure would that change require?
A credible plan shows the facility can follow your AI workloads as they grow, whether or not it uses liquid cooling today.
Data centre water usage and why cooling choices matter
Data centre water usage depends mainly on the cooling system. Evaporative cooling uses the most water. Air-cooled systems use less water but more energy, which is why water and energy figures should be read together.
Lawrence Berkeley National Laboratory estimates that US data centres used about 64 billion litres of water directly for cooling in 2023. Indirect use through power generation was far larger, at around 800 billion litres, which is leading to increasing scrutiny.
Design makes a large difference. Google’s air-cooled site in Nevada used 1.5 million gallons (around 5.7 million litres) of water in 2024. Its site in Council Bluffs, Iowa used about 1 billion (around 3.8 billion litres).
Ilkari’s refrigerant-based system keeps water consumption low. Eduardo describes water as a limited resource that needs to be used responsibly. In Colombia, the point is practical: Bogotá spent a year under water rationing that ended in April 2025.
Modular data centre design and scalable colocation: building for change
Modular data centre design lets a high-density data centre grow in steps. At Ilkari, that means Performance Optimised Data Centre (POD) architecture with in-row cooling. Power and cooling are added as demand rises instead of being built in advance.
The benefit shows when requirements change. Ilkari recently reconfigured a POD in 30 days to support a customer’s high-density needs. Forty racks at 5 kW became 20 racks at 10 kW.
In Eduardo’s words, that turnaround is “only possible when a facility is designed for change and adaptability, rather than attempting to bolt flexibility on afterwards.”
This is where scalable colocation earns its place. Customers start with an initial setup and grow from there. They avoid overbuying today, and they avoid a facility that cannot reach the density they will need later.
Supply is tightening too. CBRE reports that average vacancy across major markets fell from 8.3% in 2025 to 6.7%.
Sustainable data centre operations
Sustainable data centre operations build efficiency, community engagement and skills into how a facility runs every day. AI has put data centres’ resource use under closer watch, and Eduardo welcomes that scrutiny. His argument is that sustainability belongs in daily operations rather than in a separate initiative.
In practice, that means efficient cooling, open engagement with the community and investment in the engineers who keep higher-density infrastructure running safely.
It also means working with local governments so the right investment goes into power and energy infrastructure. That matters in Colombia, where the grid operator XM projects a firm-energy shortfall rising to 4,811 GWh by 2028 to 2029.
Ilkari’s Tocancipá site draws more than 68% of its energy from renewable sources, mainly hydropower. It holds SS 564 certification for sustainable data centres. Eduardo also points to the site’s wider footprint, including indirect jobs and investment in local connectivity.
What AI readiness means for regulated organisations in Latin America
AI readiness in Latin America is partly a question of jurisdiction. Data protection rules across the region limit where personal data can go. For regulated organisations, that affects where AI workloads can run.
Colombia permits international transfers of personal data only to countries with adequate protection, under Law 1581 of 2012. Under SFC External Circular 005 of 2019, financial entities must also check where cloud data is processed and it is recommended to use providers certified to that have at least ISO 27001. If they have ISO 27017 and ISO 27018, this is an even greater advantage.
Brazil’s ANPD set its international transfer rules in Resolution 19/2024. In Chile, Law 21.719 is scheduled to take effect on 1 December 2026 however, the Senate committee backed a bill to postpone it to December 2027, but this has not yet become law.
Ilkari’s Colombia data centre holds ISO 27001, TIA-942-C Rated-3, and ICREA Level IV certifications. Claudia Liñán, Ilkari’s Technical Operations Manager, discusses the day-to-day side in designing and operating AI-ready infrastructure in emerging markets.
Questions to ask an AI-ready data centre provider
These questions reveal more about AI readiness than a capacity spreadsheet.
- Density: How does the design scale with rack power density, beyond floor space? What kW per rack is available now, and what is planned?
- Proof: Can the provider show it meets requirements under real conditions? How long would a reconfiguration take?
- Cooling: What happens when air cooling is no longer enough?
- Efficiency: What are the PUE and WUE? Is the PUE a design or a measured figure?
- Sustainability: Is it part of daily operations, covering cooling design, community engagement and skilled engineering talent?
- Jurisdiction: Which laws govern the data, who operates the facility, and which certifications are held?
Conclusion
An AI-ready data centre is one that can change. Rack density will keep rising, and cooling plans will need to rise with it. For CIOs, that makes adaptability the first thing to test when choosing a provider.
Read the AI power challenge whitepaper
Plan for the density your workloads need
FAQs
What rack power density counts as high density?
High density usually means racks well above the industry average, which Uptime Institute put below 8 kW per rack in 2024. Much of the growth is now in the 10 to 30 kW range. The most demanding AI racks, such as the NVIDIA GB200 NVL72, draw more than 100 kW and need liquid cooling.
Is air cooling enough for AI workloads?
For much of today’s demand, yes. Air and in-row cooling can support racks up to about 20 to 30 kW, the thresholds commonly cited for moving to liquid cooling. As AI pushes density higher, an AI-ready data centre needs a plan for that change.
Is colocation suitable for AI workloads?
Yes, when the facility can scale density and cooling in stages. Scalable colocation lets organisations start with an initial setup and grow with their AI workloads. They avoid paying for capacity in advance and avoid being stuck in a facility that cannot reach higher densities.
Stay ahead of the curve with Ilkari
Sign up to the latest news, cutting-edge insight, product updates and exclusive announcements – delivered straight ot your inbox.


