Ai Boom or Bubble: Are data centres actually making money?

Synopsis: The AI boom is driving an unprecedented data-centre surge, with USD 3-7 trillion in global investments expected by 2030. But with a single 1GW facility costing USD 80 billion and AI chips depreciating within 5 years, a critical question emerges: can this trillion-dollar race actually make money, or is a bubble forming?

Every big leap in tech today seems to be powered by data centres, yet almost no one asks if these buildings are actually making any money? Companies are pouring in billions every few months, power grids are being reshaped just to keep up, and the spending commitments from hyperscalers sound less like corporate plans and more like the budgets of entire nations.

And now, as costs skyrocket while actual returns struggle to keep pace, a quieter, more uncomfortable thought is starting to surface: are we all watching the AI boom turn into the most expensive infrastructure bubble in history?

The Data Centre Boom 

The increasing computational demands of AI workloads, coupled with rising global interest in AI technologies, have triggered a rapid surge in data center construction. While data centers themselves are not a new concept, the energy-intensive nature of AI, along with the specialized hardware and cooling infrastructure it requires, far exceeds the needs of traditional non-AI business applications.

Meeting this demand necessitates larger facilities, reliable energy sources, and swift scaling to keep pace with global AI and cloud adoption. Morgan Stanley estimates that nearly USD 3 trillion will be invested globally in data centers by 2029, with McKinsey projecting capital expenditures approaching USD 7 trillion by 2030.

According to Allianz Research, leading technology firms, often referred to as “hyperscalers,” are driving this expansion. In Q2 2025, Amazon, Microsoft, and Google Cloud accounted for 63 percent of global cloud revenue. Alongside China’s Alibaba, Tencent, and Baidu, these companies are channeling billions into new server farms to meet growing digital demand.

In 2024, hyperscalers worldwide spent roughly USD 210 billion on data center capital expenditures specifically for AI deployments. Their 2025 capital expenditure plans remain substantial, with Amazon targeting approximately USD 100 billion, Alphabet USD 75 billion, and Meta USD 65 billion, largely for AI infrastructure development. Innovative projects are already underway, including the world’s first commercial underwater data center is operational in China, while Amazon founder Jeff Bezos has even suggested that data centers could eventually be built in space over the next 10 to 20 years.

Although the United States continues to dominate the data center market, other regions are seeing rapid growth, Allianz Research notes. In 2024, the US invested over USD 74 billion in data center construction. China is also expanding rapidly, with Greater Beijing alone representing around 10 percent of global hyperscale capacity and projected to double its installed IT load to over 8GW by 2030. 

Across the Asia-Pacific region, 3.2GW of capacity was under construction as of early 2025, with an additional 13.3GW planned, indicating strong growth through 2026-27. Europe and the broader EMEA region, historically trailing the US and China in AI investments, are now experiencing a 43 percent annual increase in pipeline activity. London and Dublin lead the market with over 1GW each, followed by Amsterdam, Frankfurt, Paris, and Milan. Collectively, these six cities represent roughly 45 percent of both operating (4.6GW) and planned (6.3GW) capacity.

In the Middle East, Saudi Arabia’s Public Investment Fund is expanding its data center footprint with the launch of Humain, aiming to build AI factories equipped with hundreds of thousands of Nvidia GPUs over the next five years. This initiative faces regional competition from the UAE’s Stargate project, which is developing a large-scale campus in Abu Dhabi.

How Do Data-Centres Make Money? 

Retail Colocation: Retail colocation represents the smallest segment in the data center revenue hierarchy but is often the most profitable on a per-rack basis. In this model, customers rent individual cabinets, cages, or small suites, sometimes only a few kilowatts of capacity within shared facilities. 

Customers provide their own servers, while the operator delivers essential services such as conditioned power, redundant cooling, physical security, and network connectivity. Pricing is typically per rack or per kilowatt, with additional revenue generated through cross-connect fees, remote hands support, and managed services.

Retail colocation appeals to enterprises, financial institutions, government agencies, and IT service providers that require control over their hardware but cannot justify constructing entire facilities. While margins are higher, operational complexity increases because each client has unique compliance requirements, uptime expectations, and connectivity needs, resulting in higher churn and more hands-on management. 

Companies like Equinix have achieved global leadership through this approach. Its dense interconnection ecosystems, connecting thousands of networks, clouds, and content providers, allow it to command premium pricing. In this model, interconnection itself becomes a significant profit driver. For investors, retail colocation resembles a multi-tenant office building in the digital realm: short-term leases, diversified tenants, and high returns per unit of space. It is operationally demanding but financially rewarding.

Wholesale Colocation: Wholesale colocation targets larger customers with longer-term contracts and lower per-unit margins. Clients lease entire suites or halls, typically ranging from 1 to 20 megawatts of capacity, under agreements spanning 5-15 years. These tenants, cloud providers, AI startups, or SaaS companies, install their own hardware, while the operator supplies the powered shell, cooling systems, and connectivity infrastructure.

The pricing structure shifts from per-rack to per-megawatt leases, generating predictable cash flows and reducing churn. Operators focus on keeping facilities powered, cooled, and compliant, while operating costs decline, maintenance standardizes, and exposure to client volatility diminishes. Though margins per kilowatt are lower than in retail colocation, the scale compensates for this. 

Companies such as Digital Realty and CyrusOne have optimized this model, constructing multi-megawatt campuses with modular expansion capabilities. Facilities are often designed for flexibility, allowing halls to be subdivided, power blocks added, or fiber connections extended as tenants expand. 

Wholesale colocation aligns closely with infrastructure fund strategies, offering stable cash flow, long-term leases, and creditworthy tenants. For operators, it provides anchor stability within a mixed portfolio, balancing the higher-touch volatility inherent in retail colocation.

Hyperscale Leases and Build-to-Suit: Hyperscale leasing represents the top tier of the market, involving long-term, multi-megawatt agreements between data center developers and the largest global tech companies, including Microsoft, Google, Amazon, Meta, Oracle, and increasingly, sovereign cloud operators. 

Hyperscale arrangements typically take two forms: direct leases of completed facilities or portions of campuses, and build-to-suit partnerships, where developers finance and construct facilities according to the hyperscaler’s precise specifications.

Pricing is usually confidential and may include performance-based components linked to uptime or energy efficiency, with lease durations often ranging from 10 to 20 years or more. Hyperscale tenants require standardization across all regions, including electrical topology, cooling configurations, and sustainability metrics, aiming for uniformity at scale. 

These tenants rarely churn, and once their infrastructure is deployed, it becomes integral to their global cloud networks. For developers, a single hyperscale lease can anchor an entire campus, de-risking financing and attracting co-location tenants willing to pay premium rates for proximity.

While hyperscale deals compress margins, they provide long-term predictability and align the operator’s success with the tenant’s growth. The scale is immense: a single Microsoft or Meta campus can exceed 100 MW, sufficient to power a city of roughly 80,000 homes. Each campus also fosters an ecosystem of fiber carriers, cloud exchanges, and smaller tenants, generating additional premium revenue for adjacency.

Are Data Centres Actually Making Money?

According to IBM CEO Arvind Krishna, data center spending continues to surge as AI adoption accelerates. Large technology firms now emphasize the need for greater “capacity” and AI-focused “infrastructure” in their earnings discussions, and companies such as Google have even explored long-term possibilities like constructing data centers in space. Despite this momentum, Krishna believes a fundamental question remains unanswered: whether the revenue potential of data centers can ever offset the immense capital expenditure required to build and maintain them.

Krishna estimates that constructing a single 1 gigawatt data center currently costs around USD 80 billion. Based on this estimate, any company committing to 20-30 gigawatts of capacity could end up spending roughly USD 1.5 trillion. He also points out that AI chips inside these facilities depreciate extremely quickly, typically needing full replacement within about 5 years, adding another layer of recurring cost. 

Krishna further notes that global commitments tied to AI development, including attempts to pursue artificial general intelligence (AGI), collectively point toward a need for around 100 gigawatts of computing capacity worldwide. At the same cost benchmark, this implies nearly USD 8 trillion in capital investment. In his view, generating enough profit to justify such spending, potentially requiring around USD 800 billion annually just to cover interest obligations, appears highly improbable.

The scale of these infrastructure demands has driven AI firms to seek external solutions. For instance, Sam Altman has advocated that the United States add around 100 gigawatts of new energy capacity every year to support future AI needs. 

Altman also believes that companies like OpenAI could eventually generate returns on massive infrastructure commitments, with the firm itself tying up an estimated USD 1.4 trillion across various agreements. Krishna, however, differs from this outlook. He remains unconvinced that the current generation of technology is capable of reaching AGI without a major breakthrough, estimating the probability of achieving it with existing methods at effectively 0-1 percent.

As a result, the combination of unprecedented capital requirements, rapid hardware depreciation, energy constraints, and uncertainty surrounding the path to AGI has created significant doubt about whether the global data center boom will translate into sustainable long-term profitability.

Is the AI boom becoming a bubble?

In the end, the AI boom increasingly looks like a race where the infrastructure bill is rising faster than the revenue it is supposed to generate. Arvind Krishna’s numbers make that tension unavoidable: if global AI ambitions truly require something close to USD 8 trillion in data-centre capex, with chips that lose value in just five years, the industry is operating on economics that resemble long-shot bets rather than stable returns. The mismatch between capital deployed and cash actually generated from AI services is widening, not shrinking.

This is where the bubble conversation enters. The explosive growth in data-centre buildouts, power demand, and AI chip orders is being driven by expectations of extraordinary future profitability, profits that Krishna openly questions, given that today’s technology is far from delivering AGI and the monetisation of existing models remains limited. When the entire ecosystem starts depending on future breakthroughs rather than current revenue, it mirrors past cycles where enthusiasm ran ahead of economics.

So, is there an AI bubble? If you look purely at valuations, spending, and infrastructure commitments, the pattern is undeniably bubble-like. But unlike past bubbles, the technology is real, what’s uncertain is whether the returns will ever catch up to the costs.

-Manan Gangwar

Disclaimer.

The views and investment tips expressed by investment experts/broking houses/rating agencies on tradebrains.in are their own, and not that of the website or its management. Investing in equities poses a risk of financial losses. Investors must therefore exercise due caution while investing or trading in stocks. Trade Brains Technologies Private Limited or the author are not liable for any losses caused as a result of the decision based on this article. Please consult your investment advisor before investing.

The post Ai Boom or Bubble: Are data centres actually making money? appeared first on Trade Brains.