Investments in AI data centers (DC) by global artificial intelligence (AI) companies must justify an expected annual revenue of $6 trillion by 2031, according to a recent analysis. This figure has tripled from the $2 trillion projected in last year's report, while existing AI services are estimated to generate a maximum of $1.8 trillion, leaving a gap of at least $4.2 trillion that must be created in new markets.
The challenge lies in the practical impossibility of bridging this gap. Approximately two-thirds of the shortfall must come from areas that currently lack commercial products, while the scale and costs of data centers double every 12 to 16 months, compounded by shortages in power, chips, and permits.
◆ Surge in Infrastructure Investment Requires $4.2 Trillion from New Markets
According to Bain & Company’s 'Global Technology Report' released on September 29, AI infrastructure spending, which includes data centers, computing capacity, accelerators, and memory chip upgrades, is projected to reach up to $150 billion annually by 2031. Bain applied trends from the cloud industry, where infrastructure investment accounts for about 25% of industry revenue, to derive the $6 trillion revenue requirement.
Bain estimates that consumer AI (subscription and advertising) will generate between $200 billion and $400 billion, while existing AI services, including software development, sales, marketing, and customer support, will yield between $1.2 trillion and $1.8 trillion. Even at the upper estimate of $1.8 trillion, this accounts for only 30% of the required revenue, and at the lower estimate of $1.2 trillion, the gap is fivefold.
However, the remaining $4.2 trillion is not entirely uncharted territory. Bain anticipates that AI could generate between $100 billion and $200 billion from advertising replacing search, about $400 billion from industrial automation such as autonomous vehicles and drones, and up to $900 billion from physical AI encompassing simulations, digital twins, and robotics. Yet, even if all these projections are realized, the total would only reach a maximum of $1.5 trillion, meaning $2.7 trillion must come from areas like drug development, mental health support, and battery and semiconductor material development, where no commercial products currently exist.
David Crawford, Bain's global head of technology, media, and telecommunications, stated, "The current debate is focused on employee productivity, but the economics of AI infrastructure require new revenues in the trillions, surpassing productivity gains. A wave of innovation is needed that overwhelms the innovations brought by mobile and cloud technologies." He added that AI infrastructure is being built ahead of demand curves, and to sustain this, global GDP growth must be increased by about 1 percentage point annually.
The tripling of the revenue requirement in just one year is attributed to a surge in infrastructure investment. Last year's report indicated a need for $2 trillion by 2030, with an $800 billion shortfall even considering AI cost-saving effects. Bain's latest report estimates that capital expenditures (capex) from Microsoft, Google, Amazon, Meta, and Oracle will reach up to $780 billion in 2026, nearly five times the amount from three years ago. Global data center investments are expected to total between $5 trillion and $6.5 trillion by 2030, with additional power demand reaching at least 150 GW, nearly tripling global data center capacity in five years. Major AI data centers are doubling in size and cost every 12 to 16 months.
Rising chip prices, including those from Nvidia, are also driving up costs. According to Bloomberg, a lack of power, transformers, and water, along with local opposition, has led to the blocking or delay of data center projects worth $68 billion in the U.S. during the second quarter alone.
◆ Investment Competition Continues Amid Cash Shortages
The investment competition among big tech companies continued during second-quarter earnings reports. Amazon raised its capital expenditure plan from $200 billion to $220 billion on July 30, while Alphabet increased its forecast to between $195 billion and $205 billion, and Meta raised its lower estimate to between $130 billion and $145 billion. Microsoft maintained its guidance of $190 billion set in April. The total investment from these four companies is expected to exceed $700 billion this year.
The issue lies in the availability of cash. U.S. media Semafor reported on September 15 that an analysis of S&P Capital IQ estimates shows that Amazon, Alphabet, Meta, and Microsoft will spend about $66 billion more than their operating cash flow over the next six quarters, which will be covered through borrowing, stock issuance, and off-balance-sheet transactions. The same analysis indicated that reducing capital expenditures by 10% would result in a $74 billion profit during the same period.
Individual companies are also showing signs of lowering their targets. OpenAI reportedly proposed to investors a reduction in its computing expenditure goal from $1.4 trillion to about $600 billion by 2030. OpenAI has also shelved plans for an approximately 0.8 GW expansion at the Abilene Stargate site in Texas in collaboration with Oracle, and Microsoft has taken over the Stargate project in Norway.
◆ Building AI Infrastructure on Debt Raises Bubble Concerns in Finance Sector
Concerns about a bubble in the finance sector funding AI infrastructure investments are growing. Goldman Sachs reported that as of August 5, bond issuances related to AI approached $500 billion, with AI accounting for 18% of U.S. investment-grade corporate bond supply in 2024, up from 1% last year. The debt portion of hyper-scalers' capital expenditures has risen from 27% in 2025 to about one-third this year, and is projected to peak at 35% in 2027.
J.P. Morgan Asset Management estimates that cumulative AI investments will reach $5.5 trillion by 2030, with only about a quarter of that being covered by operating cash flow and equity. Of the hyper-scalers' $1.4 trillion in data center leasing obligations, approximately $1.1 trillion is not reflected on their balance sheets.
Credit rating agency S&P Global Ratings has noted that "the credit quality of hyper-scalers is gradually weakening." It pointed out that hyper-scalers are increasingly engaging in 'shadow lending' by lending their creditworthiness to neo-cloud and AI startups through lease agreements, loan guarantees, and chip purchases.
Central banks and international organizations are highlighting the opacity of these structures. Pablo Hernández de Cos, president of the Bank for International Settlements (BIS), stated in a recent speech that AI investments are being financed through "debt rather than profits, particularly through private loans." He explained that the circular financing, where chip manufacturers and hyper-scalers buy stakes in AI companies that in turn purchase chips and computing resources, makes it difficult to identify connections.
The U.K. central bank's Financial Policy Committee (FPC) recently warned that AI valuations could face more severe corrections than the drop in AI and semiconductor stocks in July, and Tobias Adrian, director of the International Monetary Fund's (IMF) Monetary and Capital Markets Department, expressed concerns about mismatches between asset lifespans and debt maturities.
Construction progress is also lagging behind plans. Investment bank Jefferies reported in June that only 12 GW of the 24 GW of U.S. data centers scheduled to be operational this year have begun construction, and up to 80% of the projects planned for 2027-2028 have not yet started.
The domestic market is not immune to this cycle. The rise in corporate value for Samsung Electronics and SK Hynix hinges on whether AI service revenues can keep pace with infrastructure investments. As AI demand increases, revenues must grow, and AI infrastructure investments must maintain the current situation. SK Hynix stated during its second-quarter earnings report that "these investments are supported by revenues generated from AI services, so the momentum for memory demand will continue."
Investments in AIDC by the telecommunications sector are also exposed to the same variables. SK Telecom has announced plans to build AIDC with a capacity of up to 15 GW by 2035, while KT plans to secure 1 GW of capacity over five years with an investment of 5 trillion won, and LG Uplus is constructing a 200 MW AIDC in Paju, Gyeonggi Province.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.