Taiwan Central Bank Warns AI Data Center Debt Could Turn Delays Into Financial Stress

A $3.2 Trillion Building Boom Faces a Revenue Test

Taiwan's central bank has identified uncertain returns on artificial intelligence investment and rising financing risks as threats to the global economy, pointing to an estimated $3.2 trillion in data center funding needs between 2026 and 2028.

In written materials submitted on October 27 ahead of Governor Yang Chin-long's scheduled October 28 appearance before the legislature's Finance Committee, the bank described a potential chain reaction: construction delays postpone income, disappointing profits weaken asset values, and losses make lenders less willing to finance further expansion.

The warning comes as Alphabet, Amazon, Microsoft and Meta are projected to spend $715.5 billion on capital investment in 2026, an increase of 90.3% from 2025. Their spending is projected to reach $991.6 billion in 2027, approaching $1 trillion in a single year.

Taiwan has a direct stake in whether those investments succeed. Demand for AI infrastructure is supporting its hardware exports, factory investment and economic growth. The central bank forecasts growth of 11.48% in 2026, even as it warns that the financing behind the global expansion is becoming more complicated.

The central problem is the gap between paying for infrastructure and earning money from it. Data centers need land, electricity, chips and buildings before they can serve customers. Borrowing helps cover those costs, but interest and repayment schedules can put pressure on developers when construction or revenue falls behind expectations.

Spending Is Outrunning Cash Generation

Alphabet, Amazon, Microsoft and Meta have substantial businesses generating cash. Even so, the central bank said operating cash flow has not kept pace with their expanding capital expenditure, reducing free cash flow and increasing the need for outside financing.

Operating cash flow is money generated through running a business. Free cash flow is what remains after capital spending, such as building facilities and buying equipment. A company can generate more cash from its customers while still having less left over if its construction bill rises faster.

The projected increase from $715.5 billion in 2026 to $991.6 billion in 2027 amounts to another $276.1 billion, or about 38.6%. That follows the much larger percentage increase projected for 2026.

Across the wider data center market, estimated financing needs of $3.2 trillion for 2026 through 2028 include $700 billion from private credit and $650 billion from investment grade bonds. Together, those two channels would provide $1.35 trillion, roughly 42% of the total.

Private credit generally means loans arranged outside public bond markets, often by investment funds. Investment grade bonds are debt securities carrying ratings that indicate comparatively lower credit risk. Both provide funding today in exchange for payments later, making future income central to the investment case.

Future Commitments Add to the Bill

The four large cloud providers had about $2.424 trillion in commitments and liabilities outside their balance sheets, compared with $604 billion in recorded long term borrowings and existing lease liabilities. The larger figure was roughly four times the smaller one.

Those future commitments included more than $1.52 trillion for chip purchases and about $904 billion for leases on data centers under construction. Payments on those future leases would begin after the facilities were completed.

For readers following the money, the timing matters. A chip purchase commitment promises spending on equipment. A future lease promises payments for access to a facility. Both can create substantial cash needs as projects move into operation.

Emerging cloud operators and data center developers are also increasingly using asset backed financing and special purpose vehicles, according to the central bank. Asset backed financing ties borrowing to assets or their expected income. A special purpose vehicle, or SPV, is a separate legal entity created to hold a project, assets or financing arrangements.

These structures can bring several investors and lenders into a single development. They also create more contractual connections, making the financial consequences of a delay depend on guarantees, repayment terms and the obligations of each participant.

Chips and Loans Run on Different Clocks

The central bank highlighted a mismatch between the pace of technological change and the duration of financial commitments. Graphics processing units, or GPUs, are key chips used to train and operate AI systems.

GPU accounting depreciation typically runs over four to six years. Depreciation spreads the recorded cost of equipment across its expected useful life. The bank said chip technology cycles have shortened to two to three years, while SPV leases and debt commonly run for five to 15 years.

A developer may therefore still be making payments under a lengthy financing agreement when newer chips have changed customer expectations. Older equipment can face pressure on rental prices or require replacement investment while earlier borrowing remains outstanding.

Construction delays make that timing problem harder. A facility waiting for electricity or equipment earns less income than planned, while its owners may continue paying financing and development costs. When it finally opens, the technology market may already have moved forward.

The bank warned that weaker cloud company revenue could lead to spending reductions or defaults. Falling asset values and credit losses could then reinforce each other: lenders recover less from troubled projects, become more cautious and restrict funding for other developments.

Four Obstacles Stand Between Construction and Income

US data center projects face four major constraints identified by Taiwan's central bank: electricity availability, equipment supplies, resident opposition and local government reviews. Each can interrupt the journey from an announced investment to an operating facility.

Community opposition is already affecting projects. A July 2026 analysis by Carbon Direct identified at least 46 US AI data center projects delayed or canceled following local opposition between January 2024 and May 2026. They represented about $170 billion in announced investment.

Residents have raised concerns about noise, air pollution, water consumption and pressure on local infrastructure. Developers and their investors face longer approval processes, legal challenges and the possibility that planned projects will be abandoned.

In Virginia, QTS Data Centers and Compass Datacenters abandoned plans for the Prince William Digital Gateway after a prolonged legal challenge led to the development's approval being overturned. In Pacific, Missouri, Beltline Energy withdrew a proposed $16 billion development in August 2026, shortly after local officials adopted a moratorium lasting one year.

Diode Ventures also abandoned a proposed data center near Cedar Creek Reservoir in Texas following opposition over water consumption and local infrastructure. These examples put concrete projects behind the central bank's concern that securing funding is only one part of bringing new capacity into service.

Loudoun County, Virginia, illustrates how international capital can concentrate in a single place. Overseas money helped drive a housing boom there about two decades ago before the market collapsed. A new wave of investment is now building data center campuses with electricity needs comparable to those of a small city. The current investments depend on computing demand, power access and operating income.

State Rules Are Changing Project Costs

US state and local governments are taking a larger role in deciding where data centers can operate and who pays for the infrastructure they require. For investors, those decisions affect both opening dates and future expenses.

Massachusetts Governor Maura Healey issued an executive order in September requiring municipal approval before state agencies could issue data center permits. California enacted seven data center bills that month, covering electricity and water reporting, protections against shifting infrastructure costs onto utility customers, and environmental review.

Oklahoma enacted the Data Center Customer Ratepayer Protection Act of 2026 to protect existing utility customers from infrastructure costs associated with large electricity users. New Jersey established a separate electricity rate class for data centers and new energy requirements, alongside measures concerning disclosure of electricity and water use.

New York lawmakers passed the Responsible Data Center Development Act, proposing a moratorium lasting one year on new permits for hyperscale data centers, environmental assessments, public hearings and separate electricity and water rate classes. Lawmakers in 16 states discussed bans on data center development.

The sequence of investment concerns and government action has unfolded across several months:

• July 2026: Carbon Direct identified at least 46 projects delayed or canceled following community opposition during January 2024 through May 2026.
• August 2026: Beltline Energy withdrew its proposed $16 billion project in Pacific, Missouri.
• September 2026: California enacted seven data center bills, while Massachusetts required municipal approval before state permitting.
• October 27: Taiwan's central bank submitted its written warning ahead of Yang's scheduled October 28 legislative appearance.

Pension Funds Are Reconsidering Their Exposure

Public pension funds are among the investors weighing whether data centers offer enough potential return for the risks involved. Their decisions matter because pension assets help finance retirement benefits for public employees.

Monte Tarbox, chief investment officer for New York City's pension systems, which manage about $330 billion, described a decision to avoid a technology fund because of the systems' existing AI exposure. He said the systems:

passed on a technology-focused private equity fund because of AI exposure concerns, both in that strategy and across the NYC portfolios.

The $3.1 billion Jacksonville Police & Fire Pension Fund declined a new commitment to Blue Owl Capital's Digital Infrastructure Fund IV. Its concerns included political opposition, access to utilities, regulatory uncertainty and community resistance.

The $66.7 billion Arizona State Retirement System has been exploring ways to reduce its concentration in data centers. Such investments account for about 13% of its $9.3 billion real estate portfolio, with that share projected to exceed 20% by the end of 2026 as development assets are revalued.

Using the current portfolio figure, 13% represents roughly $1.21 billion. The projected increase in the share reflects how development valuations can change a fund's exposure even without a matching decision to allocate more money. AI investments can also appear across technology funds, real estate, infrastructure and credit, giving a pension fund several routes into the same economic risk.

Revenue Determines Who Gets Repaid

AI companies ultimately need paying customers to support the infrastructure being built around them. Rapid revenue growth can coexist with very large losses, keeping pressure on investors to supply more capital.

Anthropic reported $16 billion in revenue for the first half of 2026, compared with $1 billion a year earlier, alongside an $81 billion net loss. Those figures illustrate the distance that can remain between expanding sales and profitability.

Gartner estimated that data centers would need to generate $8.5 trillion in revenue, while projecting no more than $3.6 trillion. The figures imply a $4.9 trillion gap, although the estimate's time horizon was not specified.

If income falls short, investors may reduce the recorded value of assets, lenders may face missed payments and weaker operators may seek buyers or enter bankruptcy. The scale and timing would depend on customer demand, operating costs and the terms of individual financing agreements.

Insurance companies provide another connection to household finances. The National Association of Insurance Commissioners puts insurers' AI exposure at about $1.2 trillion. Private capital firms have acquired insurers and used their funding to support investments, linking parts of the insurance industry to private credit.

Banks can also be exposed when they lend to the funds or other nonbank institutions financing AI projects. The size of banking exposure remains uncertain. Taiwan's central bank identified this connection as a potential route through which stress could move from the AI supply chain into the banking system.

Taiwan Is Benefiting From the Same Expansion

Taiwan's economy grew 14.15% in the first half of 2026, its strongest performance for that period since 1977. The central bank linked the expansion to strong demand for AI infrastructure hardware, rising exports and private investment, as well as wealth gains associated with higher share prices.

Its September forecast put growth for the full year at 11.48%. Net external demand, which measures the contribution from exports after accounting for imports, was expected to contribute 4.28 percentage points. Capital formation, or investment in productive assets, was expected to contribute another 4.40 points.

Together, those contributions total 8.68 percentage points, or 75.6% of projected growth. Taiwan's exposure therefore extends beyond stock prices to export orders and investment decisions by businesses.

Loans to manufacturing businesses connected with AI were growing 46.87% from a year earlier in August. The central bank said rising funding demand had tightened liquidity, while liquidity across the banking system remained ample. In practical terms, businesses were seeking more money even as banks retained substantial funding capacity.

Financial markets have already reacted to concerns about investment returns. Technology shares fell sharply after July, and renewed worries from October put global semiconductor shares under pressure, affecting Taiwan and South Korea, where technology companies carry substantial market weight.

The central bank forecasts Taiwan's growth slowing to 5.82% in 2027, mainly because the comparison will be against an unusually strong 2026. It still expects demand for emerging technology applications to support exports and private investment. Yang's bank said it would monitor AI prospects and international financial conditions and adjust monetary policy as needed to maintain price and financial stability.

Key Points

• Taiwan's central bank warned that uncertain AI returns and more complex financing could transmit stress through markets and lenders.
• Alphabet, Amazon, Microsoft and Meta are projected to spend $715.5 billion in 2026 and $991.6 billion in 2027.
• Global data center financing needs are estimated at $3.2 trillion for 2026 through 2028.
• Power shortages, equipment constraints, community opposition and permit reviews can delay income while financing costs continue.
• Pension funds are reviewing exposure, while insurers and banks provide potential connections to the wider financial system.
• Taiwan forecasts growth of 11.48% in 2026 and 5.82% in 2027, with AI infrastructure demand remaining a major economic driver.