How to prepare a data center for AI
In the previous article, we explained why infrastructure is the true foundation of Artificial Intelligence. But understanding the theory is only the first step — the real challenge begins when a company needs to transform a traditional data center into an environment ready to train and run AI models.
Unlike conventional workloads, AI requires structural changes that go far beyond simply purchasing more servers. Below is a practical guide to the essential steps.
01 Assess the current capacity
Before making any investment, it is essential to map the limits of the existing environment:· How much power capacity is available per rack?· Can the cooling system handle the increased thermal density of GPUs?· Does the current network provide enough bandwidth for intensive data traffic?· Does the storage deliver the performance required by AI workloads?
This assessment prevents companies from investing in cutting-edge hardware without having the infrastructure needed to support it.
02 Scale power and cooling accordingly
High-performance GPUs consume and dissipate significantly more heat than traditional servers. An AI-ready data center typically requires:· High-density racks capable of supporting greater electrical loads per unit;· Liquid cooling or hot-/cold-aisle containment when thermal density exceeds the limits of forced-air cooling;· Power redundancy sized to handle consumption peaks during training.
In most projects, power and cooling — rather than the hardware itself — are the areas that generate the most rework when underestimated during the initial planning stage.
03 Choose the right hardware for the workload
Not every AI project requires the same type of GPU or server. Before defining the hardware, it is important to distinguish between:· Model training: requires high-capacity GPUs with substantial memory and intensive parallel processing;· Production inference: can run on more streamlined hardware, prioritizing low latency and cost-effectiveness;· Hybrid workloads: environments that alternate between training and inference require a flexible architecture.
04 Design the network for intensive data traffic
Training models involves moving massive volumes of data between GPUs, servers, and storage systems. An undersized network becomes a bottleneck before the hardware itself reaches its limits.· High-speed connections between processing nodes;· Low-latency communication between GPUs, essential for distributed training;· Network segmentation to isolate critical AI traffic from the rest of the corporate environment.
05 Adapt storage to AI performance requirements
AI models continuously consume data during training. Slow storage means GPUs sit idle waiting for data, and every minute of idle GPU time comes at a cost.· High-performance storage (flash/NVMe) for actively used data;· More cost-effective tiers for cold or historical data;· Scalability to grow alongside the project's data volume.
06 Strengthen security and governance from the start
Data used in AI projects often includes strategic and sometimes sensitive information. This should be treated as a project requirement, not as a later-stage consideration:· Granular access control for data and models;· Encryption in transit and at rest;· Governance policies aligned with Brazil's LGPD and industry regulations;· Continuous monitoring for unauthorized access.
07 Plan for scalability from the first project
AI projects rarely remain the same size. Ideally, infrastructure should be designed in modules that can scale processing, storage, and networking without interrupting ongoing operations. This avoids both over-sizing the initial investment and having to rebuild the environment months later.
Frequently Asked Questions
Does every data center need liquid cooling to run AI?Not necessarily. It depends on the thermal density of the GPUs being used. Smaller inference workloads can operate with properly sized air cooling; intensive, large-scale training generally requires liquid cooling.
What is the biggest bottleneck when adapting a data center for AI?In most projects, power and cooling are the most underestimated factors — not the hardware itself.
Can an existing data center be adapted, or is a new one required?In most cases, an existing data center can be adapted, provided that its electrical and cooling infrastructure can support the upgrade. Large-scale projects may sometimes require a dedicated environment.
During SIGE 2026, RISC Technology will showcase how its high-performance infrastructure solutions help companies take this step with security and efficiency.
