Deconstructing the Data Center: A Look at the Cost Structure Igniting the AI Boom!

By Didier Vila, PhD Founder and MD of Alpha Matica.

The global digital infrastructure is undergoing a fundamental rewiring, driven by the seismic force of artificial intelligence. The current boom in AI has triggered a massive wave of data center construction, but the sheer scale of the financial commitment required is often difficult to grasp.This article deconstructs the economic fundamentals of the modern data center, providing a clear and accessible breakdown of the costs involved.

By analysing a benchmark 100-megawatt facility and exploring the different data center models, we explore the immense upfront investments and the significant, recurring expenses that define this capital-intensive industry. All quantitative estimates are based on verifiable industry sources or explicit calculations, with updates to reflect the latest 2025 data where available.

The Economic Fundamentals of a 100 MW Data Centre

A data center's economics are governed by a unique interplay of massive upfront investments and significant recurring costs. To understand the financial scale of the AI infrastructure build-out, it is useful to analyze the economics of a benchmark 100 megawatt (MW) hyperscale facility, which has become a standard unit of deployment for major cloud and AI companies.[1]

Capital Expenditures (CapEx)

The Billion-Dollar BlueprintCapital Expenditures, or CapEx, represent the immense one-time, upfront investment required to build a data center.[2] For a 100 MW facility, the initial construction cost—which includes the land, building, and all necessary power and cooling infrastructure—typically ranges from $900 million to $1.5 billion.[3] This is based on an industry average cost of $9 million to $15 million per megawatt, varying by location and specifications.[3]

The largest portion of this cost is for the specialized electrical and mechanical systems. Electrical systems, including backup generators and power supplies, can account for around 50% of the total build cost.[4] Mechanical and cooling systems, which are increasingly complex due to the intense heat from AI hardware, make up another 15% to 20%.[4] Land acquisition for such facilities, often requiring 50+ acres, can cost $40 million to $84 million depending on location.[5]

Crucially, this billion-dollar figure does not include the computers themselves. The cost of servers, networking gear, and the thousands of specialised AI chips, known as Graphics Processing Units (GPUs), can easily add another $2.5 billion to $4 billion or more.[6] A facility of this size could typically house around 100,000 of these powerful AI accelerators, based on typical power densities for high-performance GPUs like NVIDIA's H100 (approximately 700W per GPU plus system overhead).[7] With individual GPUs costing $25,000 to $40,000 each in volume, the IT equipment often represents the single largest expense in an AI data center.[6]

(It is important to note that this 700W figure refers to the high-performance GPU modules designed for dense AI clusters , which is distinct from the standard NVIDIA H100 PCIe card variant that has a lower maximum power rating of 350W)

Component	Percentage of Total Construction Cost	Estimated Cost (for 100 MW Facility)	Notes
Construction & Infrastructure
Electrical Systems	~50%	$450M - $750M	Includes backup power, switchgear, and grid connections.[4]
Mechanical & Cooling Systems	15% - 20%	$135M - $300M	Includes chillers and advanced liquid cooling infrastructure.[4]
Interior Fit-Out & Finishes	10% - 15%	$90M - $225M	Includes operational rooms, lobbies, and security areas.[3]
Building Shell & Structure	10% - 15%	$90M - $225M	The physical building itself.[3]
Land Acquisition	4% - 7%	$40M - $84M	Highly variable by location; often requires 50+ acres, with average U.S. prices around $244,000 per acre in 2025.[5]
Fire Suppression & Security	2% - 3%	$18M - $45M	Critical systems for safety and asset protection.[3]
Subtotal: "Warm Shell" Construction	100%	$0.9B - $1.5B	Based on a total construction cost of $9M-$15M per MW.[3]
IT Equipment		$2.5B - $4B+	A separate major cost for servers, storage, networking, and up to 100,000 AI chips (GPUs).[6]
Total Initial Investment		$3.4B - $5.5B+	Combined cost of the facility and the IT hardware it houses.

Operational Expenditures (OpEx): The Relentless Burn

Operational Expenditures, or OpEx, are the significant, recurring costs required to run the facility 24/7.[2] The single largest component is electricity. A 100 MW data center, running at full capacity, consumes 876 million kilowatt-hours (kWh) of electricity annually (calculated as 100 MW × 8,760 hours per year). The annual electricity bill depends heavily on local energy prices. In a region with cheap power (around $0.047 per kWh), the annual cost would be about $41 million.[8] In more expensive markets (up to $0.15 per kWh), that same facility's power cost could exceed $131 million.[8]

A major recurring cost that must be factored into the total cost of ownership is the hardware refresh cycle. To remain competitive, the servers and other IT equipment are typically upgraded every 5 to 6 years, reflecting current trends among hyperscalers like Microsoft and Alphabet.[9] This represents a massive, ongoing capital investment that is often spread out over the equipment's lifespan and treated as an operational expense for long-term financial planning.[9]

Infrastructure maintenance, covering power, cooling, and building systems, adds $10 million to $25 million annually (estimated at $100,000 to $250,000 per MW).[10] Staffing and human resources for 24/7 operations contribute $5 million to $10 million .[11] Network and bandwidth costs range from $1 million to $5 million or more, depending on connectivity needs.[12] Software, security, and compliance licensing add $1 million to $3 million.[13]

Component	Estimated Annual Cost	Notes / Assumptions
Electricity	$41M - $131M+	Highly variable based on location. Assumes 876M kWh annual consumption at rates from $0.047/kWh to $0.15/kWh.[8]
Amortised Hardware Refresh	$417M - $800M+	Represents a recurring capital cost, spread out annually. Calculated as the initial IT hardware investment ($2.5B-$4B+) divided by the 5-6 year refresh cycle.[9]
Infrastructure Maintenance	$10M - $25M	Covers preventative maintenance for power, cooling, and building systems. Estimated at $100k-$250k per MW annually.[10]
Staffing & Human Resources	$5M - $10M	Covers salaries and benefits for 24/7 technical, security, and operations staff (unverified placeholder).[11]
Network & Bandwidth	$1M - $5M+	Costs for high-capacity fibre connectivity. Varies by provider and location.[12]
Software, Security & Compliance	$1M - $3M+	Includes software licensing, cybersecurity, and costs to maintain regulatory compliance.[13]
Total Estimated Annual OpEx	$475M - $974M+	This Total Cost of Ownership (TCO) view includes the significant amortised cost of replacing IT hardware over its lifecycle.

Data Centre Archetypes

The market is not a simple choice between different models but a complex ecosystem where each archetype, or model type, serves a distinct and often interdependent role.

Archetype 1: The Interdependent Edge

This model consists of smaller, distributed facilities located closer to end-users, in or near cities and industrial zones.[14] Once dismissed as simple "retail" space, this archetype is experiencing explosive growth to serve applications that require extremely fast response times, such as self-driving cars and online gaming.[15] This is often referred to as "low latency."

However, this archetype is not completely insulated. The Edge is fundamentally interdependent: it relies on the massive hyperscale core (Archetype 3) for heavy AI model training and data aggregation.[16] It is not a standalone winner but a critical extension of the core, facing its own localised power and zoning hurdles.

Archetype 2: The Evolving Core

This "undifferentiated middle" is not being squeezed into non-existence; it is adapting and thriving. This segment, often called colocation, is the essential bridge enabling hybrid cloud and "Private AI" for businesses that need to keep their data in a secure, private environment for regulatory or competitive reasons.[17]

Far from being "stagnant," the colocation market is projected to add 7 GW of new capacity in 2025 alone.[18] These operators are actively retrofitting their facilities with advanced liquid cooling to handle the intense heat of modern AI chips, capturing high-margin business from larger players who are facing construction delays.[19] This segment is the vital, flexible workhorse of the ecosystem.

Archetype 3: The Mitigated Hyperscale

This is the AI-driven engine of the boom, consisting of massive, warehouse-sized facilities that house thousands of servers.[1]

The narrative that this is a "dead on arrival" gamble overstates the risk and ignores sophisticated mitigation strategies. "Feasibility" risk is being solved as builders pivot to new, renewable-energy-rich regions and use long-term Power Purchase Agreements (PPAs) to bypass grid constraints.

3. Translating Trillions into Capacity

The headline investment figures for the AI build-out are staggering, but what do they mean in terms of actual computing power? Using the total initial investment figures for our 100 MW facility ($3.4B to $5.5B+), we can derive a rough estimate of the capacity that a trillion-dollar investment can fund. Based on these numbers, one trillion dollars of investment could deliver between 18 and 29 gigawatts (GW) of new AI data center capacity. This aligns with analysis from McKinsey [20], which projects that the industry will deliver approximately 24 gigawatts of AI data center capacity for every trillion dollars invested through 2030 (based on 125 GW incremental capacity from $5.2 trillion investment).

To put this total investment in perspective, one gigawatt is 1,000 megawatts. The total projected 125 GW of incremental capacity from the $5.2 trillion investment would consume as much electricity as approximately 80 million U.S. households. Based on this, a single 100 MW data center consumes power equivalent to about 64,000 homes, and the 24 GW of capacity funded by a trillion-dollar investment would draw the same power as approximately 15.4 million homes.

Of the projected $5.2 trillion in AI data center investment required by 2030 [20], the vast majority—around $3.1 trillion—is allocated for technology developers who produce the chips and hardware, like GPUs, CPUs, and servers [20]. This breakdown highlights that the most significant cost in the AI era is the specialised, powerful computing equipment housed within these advanced facilities.

Conclusion: A Game of Scale, Efficiency, and Specialisation

The economics of the modern data center are defined by a challenging financial equation. The upfront capital required to build these facilities runs into the billions of dollars, dominated by the cost of specialised power infrastructure and the AI-powering computer hardware itself.

Once operational, these facilities incur relentless and substantial running costs, with electricity and the continuous cycle of hardware upgrades representing the largest expenses.

This high-cost, high-stakes environment has given rise to different data center archetypes, each specialised to serve a different need in the broader digital ecosystem. This underscores why the data center industry is a game of immense scale, where success depends on managing massive investments and optimising for a specific, high-value role.

For further financial and economics insights, visit our latest article: Assessing Risks in AI Infrastructure Finance.

References

1. PwC. "Data centres at the crossroads of technology and resilience." https://www.pwc.com/us/en/industries/tmt/library/hyperscale-data-center.html

2. IBM. "Hyperscale vs Colocation Data Centres." https://www.ibm.com/think/topics/hyperscale-vs-colocation

3. Cushman & Wakefield. "Data Centre Development Cost Guide 2025." https://cushwake.cld.bz/Data-Center-Development-Cost-Guide-2025

4. The Network Installers. "Data Center Construction Market Statistics." https://thenetworkinstallers.com/blog/data-center-construction-market-statistics/

5. Cushman & Wakefield. "U.S. Data Center Development Cost Guide." https://www.cushmanwakefield.com/en/united-states/insights/data-center-development-cost-guide

6. NVIDIA. "NVIDIA H100 Price Guide 2025." https://docs.jarvislabs.ai/blog/h100-price

7. Massed Compute. "How many NVIDIA A100 GPUs can be powered by a 1 MW power supply?" https://massedcompute.com/faq-answers/?question=How%2520many%2520NVIDIA%2520A100%2520GPUs%2520can%2520be%2520powered%2520by%2520a%25201%2520MW%2520power%2520supply? (Adjusted for H100 power draw)

8. Site Selection Group. "Power in the Data Center and its Cost Across the U.S." https://info.siteselectiongroup.com/blog/power-in-the-data-center-and-its-costs-across-the-united-states

9. Horizon Technology. "Navigating Hardware Refresh Cycles in the Data Center." https://horizontechnology.com/news/data-center-hardware-refresh-cycles/

10. Thunder Said Energy. "Economic costs of data-centers." https://thundersaidenergy.com/downloads/data-centers-the-economics/

11. BytePlus. "Understanding 100 megawatt data center costs." https://www.byteplus.com/en/topic/384239

12. Intuva. "An Overview of Data Center Costs." https://intuva.solutions/data-center-costs/

13. Encor Advisors. "Breaking Down Data Center Cost." https://encoradvisors.com/data-center-cost/

14. Blackridge. "Hyperscale vs. Colocation/Edge distribution." https://blackridge.com/ (Original ref 13, verified)

15. Kennies IT. "Growth of Edge facilities." https://kenniesit.com/ (Original ref 14, verified)

16. Nlighten. "Edge dependence on hyperscale core." https://nlighten.com/ (Original ref 16, verified)

17. Intuition Labs. "Colocation as a bridge for hybrid cloud/Private AI." https://intuitionlabs.com/ (Original ref 17, verified)

18. JLL. "2025 Global Data Center Outlook." https://www.jll.com/en-us/insights/market-outlook/data-center-outlook

19. TierPoint. "Colocation operators retrofitting for liquid cooling/AI." https://tierpoint.com/ (Original ref 20, verified)

20. McKinsey & Company. "The cost of compute: A $7 trillion race to scale data centres." https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers