Data Centers · AI Infrastructure

How AI Is Driving Demand for Data Center Infrastructure

AI doesn't just need more data centers — it needs a fundamentally different kind of data center. Understanding exactly what AI workloads require, and how that differs from conventional cloud, explains why the entire sector is being rebuilt from the ground up.

The data center industry has been growing steadily for three decades — driven first by enterprise IT outsourcing, then by cloud computing. AI represents a different kind of demand shock: not a continuation of existing trends, but a step-change in the physical requirements of the facilities, the power they consume, and the capital required to build them.

Understanding what AI workloads actually require — technically and financially — is the foundation for understanding the investment thesis, the development challenges, and where the capital needs to go.

AI Workloads Are Not Cloud Workloads

Conventional cloud computing is CPU-intensive and relatively power-efficient. A standard cloud server draws 200–400 watts. An AI training cluster using NVIDIA H100 or Blackwell GPUs draws 700–1,000 watts per GPU — and a meaningful training run may use thousands of GPUs simultaneously. The power density implications are profound.

A conventional cloud data center is designed for 5–10 kilowatts per rack. An AI-optimised facility must support 30–100+ kilowatts per rack — sometimes approaching 200kW in the most advanced liquid-cooled configurations. This is not an incremental upgrade to existing infrastructure. It requires purpose-built facilities with fundamentally different power distribution, cooling systems, and structural design.

SpecificationConventional CloudAI/GPU Data Center
Power density per rack5–10 kW30–200 kW
Cooling methodAir cooling (CRAC units)Direct liquid cooling (DLC) or immersion
Network architectureStandard Ethernet, 10–100GbEInfiniBand or 400GbE+ for GPU-to-GPU bandwidth
Floor load requirementStandard (800–1,200 kg/m²)Reinforced (1,200–2,000+ kg/m²)
Build cost per MW$8–12M/MW$15–25M/MW
Power supply redundancyN+1 or 2N2N minimum; 2N+1 preferred

Training vs Inference: Two Distinct Demand Profiles

AI training is the most compute-intensive phase — building a model from data requires massive parallel GPU computation over days or weeks. Training clusters are typically consolidated in a small number of very large, purpose-built facilities where GPU-to-GPU communication bandwidth can be maximised. Power consumption during training runs is enormous and largely continuous.

AI inference — using a trained model to respond to queries — is distributed differently. As AI-enabled products scale to millions of users, inference workloads grow continuously and are latency-sensitive. Inference infrastructure needs to be geographically distributed — closer to end users — making it a driver of both hyperscale and edge computing investment. This is one reason the edge computing buildout is accelerating alongside the hyperscale buildout: they serve different but complementary parts of the AI compute stack.

"The hyperscale buildout funds the training. The edge buildout funds the inference at scale. Both are growing simultaneously — for the same underlying reason."

Power: The Physical Constraint

AI data centers require dramatically more power per square metre than any prior generation of computing infrastructure. The implications cascade through the supply chain: more power requires larger grid connections; larger grid connections require upgraded substations; upgraded substations require high-voltage transformers with 18–24 month lead times; and the land required to site all of this restricts viable locations to a handful of corridors in each major market.

This is why power strategy — not just power availability — has become the most important differentiator among data center developers. Companies like Google (through its $4.75B acquisition of Intersect Power) are building behind-the-meter generation capability to bypass grid interconnection queues. Others are securing long-term PPAs from renewable energy developers, nuclear operators, and natural gas plants before sites are even identified.

The Cooling Revolution

At 30kW+ per rack, air cooling — the standard approach for conventional data centers — becomes physically inadequate. Heat cannot be removed fast enough by air movement alone. The industry is transitioning rapidly to direct liquid cooling (DLC), where coolant is circulated directly to the chip package, and in some configurations, to full immersion cooling, where servers are submerged in dielectric fluid.

These cooling transitions are not retrofittable to existing air-cooled infrastructure — they require new mechanical plant, different rack designs, different floor configurations, and specialist maintenance capability. This is one reason existing data center stock cannot simply be upgraded for AI workloads: purpose-built new construction is required.

The Capital Intensity Implication

An AI-optimised data center costs $15–25M per megawatt to build — roughly double the cost of a conventional cloud facility. A 200MW AI campus costs $3–5 billion. At this scale, development finance, construction debt, and equity structuring become complex capital markets exercises, not simple property transactions. AI data center financing has evolved into a specialist discipline with its own lender base, underwriting standards, and structuring conventions.

Frequently Asked Questions
AI workloads — particularly GPU-based training and inference — require fundamentally different infrastructure: power densities of 30–200kW per rack versus 5–10kW for conventional cloud, direct liquid or immersion cooling rather than air cooling, reinforced floors, specialist high-bandwidth networking (InfiniBand), and build costs roughly double those of conventional facilities.
Training builds an AI model from data — requiring massive parallel GPU computation over days or weeks in large, centralised facilities. Inference uses a trained model to respond to queries — requiring geographically distributed infrastructure close to end users due to latency requirements. Both are growing rapidly, driving both hyperscale and edge computing investment.
At 30kW+ per rack, air cooling cannot remove heat fast enough. Direct liquid cooling (DLC) circulates coolant directly to chip packages, enabling much higher power densities. Immersion cooling submerges servers in dielectric fluid for maximum heat removal. These systems are not retrofittable to air-cooled facilities — AI workloads require purpose-built infrastructure.
InfiniBand is a high-bandwidth, low-latency networking technology used to connect GPUs within and between servers in AI training clusters. GPU-to-GPU communication bandwidth is critical for training performance — bottlenecks in networking directly reduce the efficiency of trillion-dollar training runs. Standard cloud Ethernet networks are inadequate for high-performance AI training.
AI-optimised data centers typically cost $15–25 million per megawatt of IT load capacity — roughly double the cost of conventional cloud facilities. A 200MW AI campus therefore requires $3–5 billion of construction capital. This scale requires sophisticated project finance, construction debt, and equity structuring.
Direct liquid cooling circulates a coolant (typically water or a water-glycol mixture) directly to the chip package or to a heat exchanger mounted on the server — removing heat far more efficiently than air. It enables power densities of 50–100kW per rack. More advanced implementations use two-phase cooling or immersion in dielectric fluid.
AI inference — running trained models to serve user queries — is latency-sensitive. As AI-enabled products scale to billions of interactions, inference infrastructure must be distributed geographically, close to end users, to maintain acceptable response times. This is driving investment in edge computing facilities (smaller, distributed data centers) alongside hyperscale AI training campuses.
Behind-the-meter generation means the data center produces its own electricity on-site — through solar, wind, gas, or nuclear — rather than drawing entirely from the grid. Power is generated and consumed within the same site boundary, significantly reducing or eliminating grid interconnection requirements and bypassing much of the interconnection queue.

Finance Your Data Center Project

OAKRG advises on data center project finance, construction debt, hyperscale equity raises, and energy-linked infrastructure capital across North America, Europe, and Asia-Pacific.

Speak with an Advisor