Hyperscale Data Center: What It Is, How It Works, and What It Costs

What Is a Hyperscale Data Center?

A hyperscale data center meets three thresholds: at least 5,000 servers, at least 10,000 square feet of white floor space, and at least 40 MW of power capacity. Meeting all three is what separates hyperscale from large enterprise or colocation facilities.

The term was coined to describe the infrastructure built by companies like Google and Amazon to serve hundreds of millions of users simultaneously. A standard enterprise data center is designed for one organisation's known workload, sized to peak demand with a buffer. A hyperscale facility is designed around the assumption that demand will keep growing, and the building must accommodate that growth without major reconstruction.

Scalability is the defining characteristic, not just size. A hyperscale facility uses modular designs, often called pods or blocks, where each unit adds a fixed increment of compute, power, and cooling capacity. When a new pod is needed, it is added to the campus without disrupting existing operations. This approach makes hyperscale fundamentally different from a conventional data center, where adding 20% more capacity typically requires significant infrastructure work.

The largest hyperscale campuses today hold millions of servers across multiple buildings on a single site, with dedicated substations, cooling towers, and in some cases on-site power generation. The threshold numbers (5,000 servers, 40 MW) represent the entry point for the category, not the norm.

The Three Minimum Thresholds

Approximately 800 hyperscale data centers operate worldwide as of 2025, with the United States holding the largest national concentration. The US had 5,381 total data centers as of March 2024, encompassing hyperscale, colocation, and enterprise facilities combined.

How a Hyperscale Facility Is Built and Operated

Hyperscale data centers are built around four interconnected systems: power distribution, cooling, server infrastructure, and networking. Each system is designed for redundancy and rapid expansion.

Power Infrastructure

Power arrives from the utility grid at high voltage (typically 115 kV to 230 kV), is stepped down through on-site substations, and distributed to individual server racks. Hyperscale operators maintain N+1 or 2N redundancy on power systems, meaning if one power path fails, another takes the full load without interruption. Backup diesel generators provide protection against grid outages, with fuel stored on site for 24-72 hours of full-load operation.

Power Usage Effectiveness (PUE) measures how efficiently a facility converts incoming power into useful server compute. A PUE of 1.0 is theoretical perfection: every watt goes to the servers. Legacy data centers commonly run PUE of 1.5 to 2.0, wasting half their power on overhead. Hyperscale operators use economies of scale and custom cooling designs to achieve PUE of 1.1 to 1.2 across large fleets. Google reported a trailing 12-month average PUE of 1.10 across its global data center fleet (Google Sustainability Report, 2023), meaning only 10% of power is lost to overhead rather than the 50% common in older facilities.

Cooling Systems

Cooling is the single largest overhead cost in a data center. Hyperscale operators use several approaches depending on climate and workload density:

• Air cooling with precision computer room air conditioning (CRAC) units: standard for traditional server racks
• Evaporative cooling using outside air in cooler climates (Google Oregon, Microsoft Wyoming): reduces mechanical cooling cost substantially
• Direct liquid cooling (DLC): pipes cool water directly to server components, required for high-density GPU racks drawing 30-60 kW per rack
• Immersion cooling: servers submerged in non-conductive fluid, used for the densest AI workloads

AI GPU racks running NVIDIA H100 or H200 accelerators draw 10-20x the power of standard server racks. That density makes air cooling insufficient for AI-specific pods, which is why liquid cooling infrastructure accounts for a significant share of the cost premium in AI-optimized facilities.

Networking

Hyperscale facilities require ultra-low latency, high-bandwidth networking between racks, pods, and buildings on the same campus. During AI model training, thousands of GPUs must synchronise gradient updates continuously. Even microsecond-level latency differences cause training inefficiency. This requirement drives investment in 400 Gbps and 800 Gbps network fabrics within campuses, with 93% of hyperscale operators targeting 40 Gbps or faster connections for their core fabric (Uptime Institute, 2025).

The Major Hyperscale Operators and Their Scale

Five companies account for the majority of hyperscale capacity worldwide: Amazon Web Services, Microsoft Azure, Google Cloud, Meta, and Alibaba Cloud. Each operates hundreds of individual facilities across multiple continents.

Microsoft's commitment stands out for its scale and timeline. On January 3, 2025, Microsoft President Brad Smith announced the company was on track to invest approximately $80 billion in AI-enabled data centers in fiscal year 2025 (ending June 2025), with more than half directed to the United States.

"This infrastructure is the foundation on which AI innovation is being built." (Brad Smith, Microsoft President, January 2025)

Google matched with its own announcement of $75 billion in planned capital expenditure for 2025, a figure that includes both cloud data center expansion and its custom Tensor Processing Unit (TPU) hardware programs.

The US concentration of hyperscale investment follows available land, power infrastructure, and proximity to major internet exchange points. Northern Virginia (known as Data Center Alley) has historically hosted the densest concentration, though power constraints are pushing new builds to the Southeast, Midwest, and Pacific Northwest.

Construction Costs in 2025 and 2026

Building a hyperscale data center has become substantially more expensive over the past three years. The average cost per square foot surpassed $1,000 for the first time in 2025, reaching $1,033 by year-end (ConstructConnect, 2025), compared to $535 in 2023. That near-doubling reflects tight supply chains for electrical gear, high demand for specialised construction crews, and the premium required for AI-ready power and cooling infrastructure.

Benchmark Construction Costs

Source: Turner & Townsend Data Centre Construction Cost Index, 2025-2026.

The Number Most Guides Don't Show

At $10.7M to $11.3M per MW, a 100 MW standard hyperscale facility costs between $1.07B and $1.13B for the shell-and-core structure. Add full electrical fit-out, cooling systems, and fire suppression, and total project cost reaches $1.6B to $2.2B before a single server enters the building.

For AI-optimized facilities at $20M+ per MW, that same 100 MW footprint costs $2B or more in construction alone. The difference comes almost entirely from liquid cooling infrastructure and the higher-voltage power distribution required for GPU racks drawing 30-100 kW per rack, versus 5-10 kW for standard compute racks.

Microsoft's $80B commitment for fiscal year 2025 translates, at the AI-optimized benchmark of $20M per MW, to enough physical capacity for approximately 4,000 MW (4 GW) of new AI-ready data center space. That figure is consistent with the company's reported pipeline of 5+ GW total capacity.

The industry as a whole reflected this pace: US data center construction starts reached $77.7B in 2025, up 190% from the prior year (ConstructConnect, 2025). In October 2025 alone, US companies broke ground on 23 data center sites totalling nearly $11 billion, compared to $2.9 billion in October 2024.

Why AI Training Requires Hyperscale Infrastructure

Training a frontier AI model requires a scale of compute that only hyperscale infrastructure can provide. This is not about raw server count. It is about having thousands of high-performance GPUs or TPUs connected by a fast enough network that they can work together as a single computational unit.

GPT-4 was trained on an estimated 25,000 NVIDIA A100 GPUs running continuously for months. Gemini Ultra required a custom TPU pod architecture at Google that exists in no other context. Llama 3 (405B parameters) was trained on Meta's own hyperscale infrastructure using 16,000 H100 GPUs. None of these training runs were possible on anything smaller than a purpose-built hyperscale facility, because the inter-GPU communication during training requires network latency measured in microseconds across thousands of devices.

Inference, which is the process of answering a user query using a trained model, can run at smaller scale. A single H100 GPU can handle hundreds of simultaneous inference requests. But training a new frontier model is a different problem entirely. It requires:

• Thousands of co-located GPUs connected by high-bandwidth, low-latency networking (InfiniBand or RoCE at 400+ Gbps)
• Uninterrupted power delivery for weeks or months, with zero tolerance for outages mid-training
• Cooling systems that handle 30-100 kW per rack across thousands of racks simultaneously
• Software infrastructure to checkpoint, restart, and manage a training run across hardware failures

When you use ChatGPT or Claude, your request travels to one or a small number of GPUs in an inference cluster. That cluster could, in principle, sit in a colocation facility. The model those GPUs are running, however, was created in a hyperscale training facility that no colocation operator currently builds.

This distinction explains why Meta is constructing the Hyperion campus in Louisiana at a scale that analysts estimate at several gigawatts of AI compute capacity. It is not primarily to serve more users. It is to train the next generation of Llama models at a scale that requires an entire power plant's worth of electricity.

For a technical breakdown of the GPUs used inside these facilities, see our analysis of NVIDIA H100 specifications and pricing.

Hyperscale vs. Colocation vs. Enterprise: Key Differences

The three main data center types serve different purposes and different customers. The differences matter when evaluating where to run workloads.

A company that needs 100 servers has no reason to build a hyperscale facility. A startup running AI inference can use colocation or rent GPU time from a cloud provider. Hyperscale investment is what happens when AWS needs 100,000 additional servers in a new region, or when Meta needs a training cluster that does not yet exist anywhere in the world.

For a detailed comparison of colocation pricing, power options, and tier ratings, see our article on colocation data centers explained.

The practical implication for most organisations: they interact with hyperscale infrastructure every time they use AWS, Azure, Google Cloud, or any SaaS product built on those platforms. They do not build hyperscale themselves. Hyperscale is infrastructure at a cost and scale that only makes sense for companies serving hundreds of millions of users or training the world's most compute-intensive models.

AI Campuses and the Shift to Gigawatt-Scale Builds

The next phase of hyperscale development has a different name: the AI campus. Where a hyperscale data center is measured in megawatts (40 MW to 500 MW), an AI campus is planned in gigawatts (1 GW to 5 GW), with multiple buildings, a dedicated power substation, and in some cases on-site generation from solar, gas, or nuclear sources.

At the campus scale, the economics shift. Fully built-out AI campuses with power, land, connectivity, and equipment cost $45B to $55B per gigawatt (Turner & Townsend, 2025). A 5 GW campus, at the upper end of that range, represents a development value approaching $275B over its build-out period.

"The infrastructure investment cycle in AI is unlike anything we have seen before. Hyperscale operators are committing capital at a pace that would have seemed implausible three years ago." (Turner & Townsend, Data Centre Construction Cost Index, 2026)

The global investment trajectory supports this scale. Companies worldwide are expected to invest nearly $7 trillion cumulatively in data center infrastructure through 2030, with more than 40% directed to the United States. More than $4 trillion of that total is allocated to computing hardware rather than real estate and power.

Over 60 data center projects with a combined value exceeding $50 billion were expected to break ground in the first half of 2026, a pace that represents a multi-year acceleration driven almost entirely by demand for AI training and inference capacity. In October 2025 alone, US companies invested nearly $11 billion in 23 new data center construction sites (ConstructConnect, 2025).

The result is that hyperscale is no longer the ceiling for data center scale. It is, increasingly, the floor. The facilities being planned today are categorically larger than what the term described when it was coined a decade ago. For context on the hyperscaler companies making these investments and how their cloud businesses are structured, see our hyperscalers explained guide.

Frequently Asked Questions

Related Articles