Long-term investment and revenue estimates
Cumulative global data center investment is expected to reach $31.6 trillion, but equipment replacement and securing power will require continued funding. An analysis estimating an annual revenue shortfall of $4.2 trillion in 2031 highlights the challenge of turning technological advances in AI into businesses that can recoup that investment.
After facing the question of how quickly it can build data centers, the global artificial intelligence (AI) industry now faces the question of where it will earn the money to operate those facilities and replace their equipment. Accounting and consulting firm PwC projected that cumulative global data center capital investment could reach $31.6 trillion by 2050. This is not a sum of investment contracts already signed, but a long-term estimate based on expanding AI adoption and computing demand.
Reporting on this investment expansion on October 3, Reuters highlighted the possibility of a gap between the pace of technological progress and companies' investment recovery schedules. The issue is that expectations that AI could transform the economy must be distinguished from judgments that investors can generate sufficient returns within a specified period.
The analysis commissioned by PwC from Oxford Economics covers 46 countries and territories. Its baseline investment period runs from 2026 to 2050, with amounts expressed in 2025 dollars. It should therefore not be read as funding required this year or a bill payable by any particular company. Applying the October 2 exchange rate of approximately 1,343 won per dollar gives a simple conversion of $31.6 trillion to approximately 42,439 trillion won. This conversion is intended to convey the scale, not to forecast spending that reflects future exchange rates or inflation.
A notable aspect of this projection concerns the equipment inside data centers rather than the buildings. PwC generally assumes replacement cycles of 4 to 6 years for servers and graphics processing units (GPUs). Although buildings and land can remain in use for a long time, computing equipment requires repeated investment amid competition over performance and energy efficiency. This produces a cash flow pattern different from businesses whose funding burden falls substantially once construction is complete. Operators must prepare for the next equipment replacement even as they run their facilities and generate revenue. This is why information and communications equipment accounts for a growing share of total investment in PwC's long-term projection. Actual replacement timing, however, can vary with workloads and how equipment is used; it does not mean that every GPU becomes unusable after a fixed period.
Another key figure came from a technology report released by Bain & Company on September 29. The analysis, written by David Crawford and others, assumed that annual AI infrastructure investment could reach $1.5 trillion in 2031. Taking capital expenditure to be approximately 25% of industry revenue would then require a market generating approximately $6 trillion a year to sustain that investment. Simple conversions at the current exchange rate put these amounts at approximately 2,014.5 trillion won and 8,058 trillion won, respectively.
Bain projected annual revenue of up to $1.8 trillion, approximately 2,417.4 trillion won, from existing consumer and enterprise AI markets. Its calculation leaves approximately $4.2 trillion, or approximately 5,640.6 trillion won, to be generated in new markets. This is neither a cumulative revenue shortfall over the next 5 years nor a confirmed debt owed by US cloud companies. It is an assessment of the annual revenue required from the global AI market in 2031.
The result changes when the assumptions change. Falling equipment prices or lower costs for the same computing workload could reduce the revenue required. Conversely, if price competition lowers service charges while facility costs remain unchanged, recovering investment could become difficult even as user numbers increase. Revenue also differs from profit. Customer fees must cover electricity, labor, equipment operating expenses and financing costs, among other expenses. Meeting revenue targets and delivering sufficient returns to investors require separate assessments.
Productivity gains and financing risks
It would also be inaccurate to conclude that there is no evidence of AI's productivity effects. Material from the US National Bureau of Economic Research (NBER) describing research by Erik Brynjolfsson, Danielle Li and Lindsey Raymond reported that introducing generative AI into customer support increased issues resolved per hour by an average of approximately 14%. The gains were greater for less experienced employees. However, findings from customer support at a particular company cannot be extrapolated into an economy-wide productivity growth rate. Adoption effects differ by task, as does the time needed to integrate AI with existing systems and review its output.
Recovering investment requires another step: determining how much of the benefit from employees completing work more quickly becomes revenue for AI suppliers. Companies using AI may retain most of their cost savings as profit, or competition may lead them to reduce product prices, passing the benefits to consumers. Even when a technology benefits society as a whole, infrastructure investors may not recover enough of that value. Assessments of AI's economic effects therefore need to examine improvements in work efficiency and suppliers' profitability separately. This is also why Bain judged that productivity improvements in existing services alone would struggle to support the scale of investment. The analysis calls for new products and markets to emerge. These markets, however, were presented in the report as possibilities and cannot be treated as secured customer orders or revenue.
Financing methods are also changing. In research submitted to the Brookings Institution's fall 2026 academic conference, Stijn Van Nieuwerburgh of Columbia Business School analyzed how the risks of AI facility investment are shifting from direct investment by large companies to joint ventures, private credit, special purpose vehicles and leases. Under these structures, the company using a data center, the company owning the building and the financial institution providing loans may be separate entities yet depend on the same customer's ability to pay. Investment that appears to be spread among multiple parties may therefore be tied to the business performance of a small number of AI companies. The researcher also explicitly stated that it was too early to characterize the situation as posing systemic risk on the same scale as past credit crises.
The ways these risks could materialize are relatively specific. If expected customers arrive late, facility utilization falls. Interest and rent must still be paid under contract even if revenue declines. If equipment resale values also fall, the amount recoverable by selling collateral declines as well. Investment scale alone is therefore insufficient to judge financial soundness. What matters is who the contractual counterparty is, whether minimum usage payments are guaranteed and who bears losses when a contract is canceled. Even funds spread across multiple financial products can face rising risks together if the ultimate source of repayment is the same customer.
Power supply and opportunities for Korean suppliers
According to the International Energy Agency's (IEA) 2026 report, global data center electricity demand increased by 17% in 2025, while electricity consumption at AI-focused data centers rose by 50%. This indicates that expanding computing facilities is translating into actual energy demand. The IEA noted that US developers are pursuing on-site gas-fired generation to avoid grid connection delays, but that gas turbine supply and regulatory and financing conditions mean this is not necessarily a faster solution. From a financing perspective, the power issue concerns when operations can begin. Even after buildings and equipment have been secured, a lack of electricity delays revenue generation. If facilities were financed with loans, financing costs can accumulate during that period.
This is also why cooling and power equipment need to be assessed together. Securing computing equipment alone does not allow operation at the planned performance level if power and cooling capacity cannot keep pace. Revenue depends on the operating capacity that can actually be offered to customers, rather than the maximum computing capacity stated in a business plan. Cost allocation also remains an issue for local communities. Assessments need to cover who pays for power infrastructure expanded for data centers and how remaining costs will be handled if expected demand declines. Contract structures need to be examined to prevent uncertainty in AI demand projections from being passed on to electricity rates and public infrastructure investment.
In material released on September 17, Samil PwC identified high-bandwidth memory (HBM), memory semiconductors, power equipment, energy storage systems and cooling equipment as opportunities for South Korea. Its analysis is that companies can supply the necessary equipment regardless of which country builds the data centers. However, projected cumulative global investment does not translate directly into revenue for Korean companies. Results will vary with competition in individual components, customers' purchasing schedules and the pace of suppliers' capacity expansion. Even in a rapidly growing market, prices and margins can come under pressure if supply grows faster than demand.
Korean companies also need to look beyond announced investment amounts when choosing indicators to monitor. They should examine whether customers have completed financing, whether power supply schedules have been finalized and whether conditions are in place to collect payment after equipment delivery. Long-term growth prospects and the security of individual contracts are different questions. Taken together, the published analyses suggest that the central issue in the AI investment race is timing rather than optimism about the technology itself. The timing of revenue from new services, power supply to data centers, equipment replacement and debt repayment must align. The possibility that AI will deliver substantial long-term economic benefits is not the same as the possibility that every data center now being built will generate sufficient returns. Alongside announced construction scale, the criteria for evaluating the investment race will be actual utilization, recurring paid demand and cash remaining after costs have been paid.
Original Korean article: efn.co.kr
English brief by Korea Wind Intelligence, koreawindintel.com
