AI Workloads and the Future of Data Centers

A New Era of Data Center Growth

Artificial intelligence is no longer just another workload inside the data center. It has become the primary force reshaping how data centers are designed, powered, financed, and deployed. As AI adoption accelerates, hyperscalers are rethinking everything from site selection to power strategy to infrastructure architecture.

#1 AI is now the primary driver of global data center infrastructure investment

60–120d Time to build a data center, often shorter than securing power permits

2 types Training and inference, each demanding entirely different infrastructure models

AI workloads are pushing data center capacity growth at an unprecedented pace. Power demand is rising rapidly as organisations deploy compute-intensive AI systems at scale. What was once a supporting workload is now the central driver of infrastructure investment.

This surge is forcing hyperscalers to prioritise access to power, speed to deployment, and long-term scalability over traditional cost optimisation alone.

Training vs Inference: Two Very Different Workloads

AI workloads fall into two distinct categories. Understanding how each shapes infrastructure decisions is fundamental to grasping where the industry is heading, and why a single design no longer fits all.

Side-by-side comparison

Dimension	Training	Inference
Purpose	Build and refine AI models	Run trained models in real time
Power density	Extremely high	Moderate, but distributed
Cooling needs	Advanced liquid cooling required	Conventional or hybrid cooling
Latency sensitivity	Low, location is flexible	High, must be close to users
Hardware	Specialised AI accelerators (GPUs, TPUs)	Mixed, CPUs, GPUs, custom chips
Location preference	Power-rich, land-available regions	Near population centres and cloud nodes
Demand pattern	Burst-intensive, periodic	Continuous, growing over time

Trend to Watch Over time, inference is expected to become the dominant AI workload, driving continuous compute demand rather than one-time training bursts. Infrastructure built today must account for this shift.

Why Data Center Design Is Changing

The rise of AI is producing two distinct design patterns simultaneously, and hyperscalers are investing in both at the same time.

The two emerging models

Large, high-density training campuses +

Built specifically for training workloads, these facilities feature liquid cooling, resilient power systems, and fault-tolerant architectures. They are typically located in power-rich regions where land and energy are more accessible, prioritising compute density over proximity to end users.

Smaller, distributed inference facilities +

These facilities are optimised for low latency, modular expansion, and energy efficiency. They are often integrated directly into existing cloud campuses and positioned closer to users and applications that require real-time AI responses, from search and chat to recommendations and analytics.

Why not one design for both? +

Training and inference have fundamentally different requirements, power density, latency tolerance, cooling approach, and geographic placement. A single facility architecture cannot efficiently serve both. Hyperscalers are therefore building a two-tier infrastructure estate: large central campuses for training, distributed edge-adjacent nodes for inference.

What role does liquid cooling play? +

AI accelerators like GPUs and TPUs generate heat densities that exceed what traditional air cooling can handle. Liquid cooling, whether direct-to-chip, immersion, or rear-door heat exchangers, is becoming standard for high-density AI training infrastructure. Some newer facilities are designed with liquid cooling as the primary thermal management strategy from day one.

Are existing data centers being retrofitted? +

Yes. Rather than replacing legacy facilities, hyperscalers are retrofitting existing data centers to support higher-density AI workloads. This involves upgrading power delivery, adding supplemental cooling, and replacing general-purpose rack configurations with GPU-optimised layouts, extending the life of existing assets while accelerating AI capacity.

Power: The Biggest Constraint

Access to reliable power is now the single biggest bottleneck for AI infrastructure growth. In many regions, securing power and permits takes longer than building the data center itself. This has transformed power from an operational input into a strategic asset.

To overcome this, hyperscalers are expanding beyond traditional markets and moving into secondary regions where power can be delivered faster. They are also exploring alternative energy strategies to reduce dependence on constrained grids.

Alternative power strategies gaining traction

On-site generation +

Rather than relying solely on grid power, some hyperscalers are installing dedicated generation capacity, gas turbines, fuel cells, or small-scale nuclear, directly on or adjacent to their data center campuses to ensure uninterrupted supply independent of grid constraints.

Microgrids +

Microgrids combine local generation sources with energy storage and intelligent switching, allowing data centers to operate independently of the main grid during peak demand or outage events. They also enable better integration of renewable energy sources.

Direct energy partnerships +

Hyperscalers are entering long-term power purchase agreements (PPAs) and direct contracts with utilities and renewable generators to secure guaranteed capacity years in advance, effectively locking in power before competitors can access the same resources.

Power availability is no longer just an operational concern. It is a competitive advantage.

Five Major Shifts in Hyperscaler Strategy

To keep pace with AI demand, hyperscalers are adjusting their strategies in five interconnected ways. These are not incremental adjustments, they represent a fundamental rethinking of how infrastructure is planned, built, and operated.

Track the shifts, click each to mark it understood

✓

Shift 01 · Energy
Investing directly in energy infrastructure to secure long-term power access, moving upstream from data center operator to energy stakeholder.
✓

Shift 02 · Ownership
Trading full ownership for faster deployment through leasing and partnerships, prioritising speed to capacity over balance-sheet control.
✓

Shift 03 · Construction
Adopting modular and prefabricated construction to reduce build times, factory-built infrastructure components that can be deployed and scaled faster than traditional builds.
✓

Shift 04 · Consolidation
Consolidating workloads into large, multi-facility campuses instead of scattered single sites, enabling shared power infrastructure, redundancy, and operational efficiency at scale.
✓

Shift 05 · Retrofitting
Retrofitting existing data centers to support higher-density AI workloads instead of replacing them, extracting more value from existing assets while accelerating AI capacity online.

The Future of AI Infrastructure

AI has become the gravitational centre of digital infrastructure. The line between data centers and energy systems is beginning to blur as hyperscalers take a more active role in power generation, financing, and grid coordination.

As AI workloads continue to grow, the organisations that succeed will be those that understand how compute, power, location, and design intersect. The next phase of AI growth will not be defined by models alone, but by the infrastructure built to support them.