Japan’s FSA Puts AI Data-Center Finance Under Closer Watch as Banks and Insurers Expand Exposure

The artificial-intelligence race is no longer only a contest over chips and models. It is also a contest over physical infrastructure—and over who is willing to finance it. As hyperscalers commit extraordinary sums to data centers, power connections, cooling systems and computing hardware, Japan’s Financial Services Agency has made data-center lending an explicit focus of financial supervision.^[1]

The FSA’s “Strategic Priorities: July 2026–June 2027,” published September 15, lists project finance including credit to data centers alongside domestic real-estate lending, large exposures and lending to overseas funds. The agency says it will examine the effectiveness of financial institutions’ investment and lending policies, deal screening, ongoing monitoring and concentration-risk management. It will also look across institutions for exposures shared by multiple lenders to sectors where financing has become concentrated.^[1]

This is not a policy to shut off funding to AI infrastructure. Bloomberg reported on September 25, citing an unnamed senior FSA official, that the agency is stepping up scrutiny of major banks and life insurers and that the current focus is mainly on data-center projects in the United States. The official said the regulator does not intend to discourage financing to a growth sector, while pointing to uncertainty over whether hyperscalers will earn adequate returns on massive spending and to broader concerns about the pace of development.^[2]

Why data-center finance has become a supervisory issue

AI infrastructure has two characteristics that matter to lenders: enormous upfront costs and long payback periods. Land, grid connections, substations, cooling equipment, buildings, servers and accelerators must often be financed before a facility produces revenue. In project finance, repayment depends principally on cash generated by the project, so underwriting extends far beyond the sponsor’s name. A lender has to ask whether the project will be completed, whether tenants will remain, whether power will be available and whether the economics will survive technological change.

The AI boom is scaling that familiar infrastructure model at unusual speed. The Bank for International Settlements said in January that AI investment was surging and that the scale of expected spending would push firms increasingly from internal cash flow toward debt, with private credit playing a growing role. In March, BIS researchers described the rise of off-balance-sheet structures in which joint ventures or special-purpose vehicles own data-center assets and borrow against leases, capacity commitments, collateral and sometimes guarantees from hyperscalers.^[3][4]

That evolution turns a data center from a simple real-estate asset into a web of linked credit exposures. A transaction may appear diversified on paper, yet multiple projects can still depend on the same hyperscaler, power market, sponsor, private-credit manager or refinancing channel. The FSA’s emphasis on concentration risk is therefore about more than the size of any single loan.

Japan’s megabanks already sit near the center of global project finance

Japan matters to this story because its financial institutions are not peripheral lenders to global infrastructure. MUFG Bank, Sumitomo Mitsui Banking Corporation and Mizuho Bank have spent decades building large international project-finance businesses.

MUFG’s published Project Finance International league table for 2025 places MUFG first globally by mandated lead-arranger volume, SMBC second and Mizuho sixth. MUFG reports $33.27 billion across 256 transactions. Those figures cover project finance broadly rather than data centers alone, but they show why Japanese balance sheets and arranging capabilities matter when a capital-intensive industry suddenly requires enormous amounts of financing.^[5]

AI data centers are already part of that business. In investor material, MUFG said it had arranged multiple gigawatt-class AI data-center projects in lead-left positions during fiscal 2025. In December 2025, MUFG also announced that it had become a limited-partner investor in a fund under the AI Infrastructure Partnership, a consortium focused on data-center and supporting power infrastructure.^[6][7]

Mizuho has similarly described data-center financings as an important source of activity in U.S. leveraged-finance markets. In a 2026 discussion featuring its leveraged-finance and project-finance executives, the bank noted that 2025 activity had been driven in significant part by data-center deals, including transactions in which Mizuho had taken lead roles. The bank itself posed the central market question: can financing markets absorb the volume of projects now seeking capital?^[8]

Life insurers bring another pool of long-duration capital

The FSA’s cross-sector focus also matters because banks are not the only institutions funding infrastructure. Life insurers collect long-duration liabilities and invest for long horizons, making infrastructure debt and private assets potentially attractive matches. Japanese insurers expanded into overseas project finance and alternative assets during the long era of ultra-low domestic yields.

One recent report illustrates the scale being contemplated. Nikkei reported that Nippon Life Insurance planned to raise its infrastructure-finance balance—including U.S. data-center projects—to ¥2 trillion by fiscal 2035. Reuters relayed the report on September 20, while stating that it could not independently verify the plan and that Nippon Life was not immediately available for comment. Japan.co.jp therefore treats the ¥2 trillion figure as a reported plan, not as a company-confirmed data-center commitment.^[9]

Nippon Life has separately confirmed a broader expansion into alternatives. In June it announced a strategic partnership with Blackstone aimed at investment-management services in private credit and real estate, saying it was expanding alternative investments as part of its effort to improve portfolio returns and stability. The agreement is not specific to AI data centers, but it illustrates why the FSA increasingly has to follow risk across banks, insurers and private-market structures rather than supervise each sector in isolation.^[10]

What lenders actually have to underwrite

The FSA’s priorities do not impose a new numerical rule specific to data centers. Instead, the language—lending policy, deal review, ongoing management and concentration risk—points to the quality of underwriting and monitoring. For a data-center project, that means examining several layers of risk at once.

In the United States, community and permitting risk has also become a financing variable. Reuters reported in August that lenders were paying closer attention to local approval and regulatory readiness as opposition to data-center projects grew over electricity use, water, noise and other concerns. A delayed grid connection or permit is not only a political or construction issue; it can postpone the start of cash flow that services project debt.^[11]

A famous tenant does not make the structure risk-free

The credit quality of a major technology company matters, but it does not eliminate project risk. BIS research highlights the growth of structures in which the hyperscaler does not borrow all the money directly. Instead, a separate vehicle may own the data center, funded with equity and debt, while a hyperscaler signs a long lease or capacity commitment and may provide guarantees. Most of the debt can therefore sit outside the technology company’s own balance sheet even though the economic relationship remains close.^[4]

For supervisors and lenders, the relevant borrower is therefore not just the legal entity named on a loan agreement. The full risk map can include the sponsor, tenant, utility, private-credit fund, banks providing financing lines, bond investors and contractual guarantors. If many projects ultimately depend on the same small group of technology companies, diversification can be less robust than it appears.

What happens if expected AI returns disappoint?

The BIS has framed the sustainability of the AI investment boom around a straightforward question: can AI firms meet the earnings expectations embedded in the buildout? In its 2026 Annual Economic Report, the BIS said the five largest hyperscalers were on course to spend more than $1 trillion on AI-related capital expenditure across 2025 and 2026, with investment commitments outpacing earnings and free cash flow and leading some firms toward more debt financing.^[12]

The FSA does not need to decide whether AI is a bubble. A prudential supervisor has a different task: ask whether lenders remain resilient if expectations are wrong. Demand for new compute could slow. Prices could fall more rapidly than expected. New chips could change the amount or type of physical capacity required. Power costs could rise. A lender’s job is to model those outcomes before the loan is made and continue testing them while the loan is outstanding.

A new supervisory structure for cross-sector risk

The data-center issue arrives as the FSA itself is changing how it supervises financial institutions. A reorganization in summer 2026 replaced the former single supervision bureau with the Asset Management and Insurance Supervision Bureau and the Banking and Securities Supervision Bureau. At the same time, cross-cutting monitoring under the Deputy Commissioner for Supervision is intended to link institution-level, or microprudential, supervision with a systemwide, or macroprudential, view.^[1]

AI data-center finance is almost a textbook case for that approach. A construction loan may sit on a bank balance sheet, an insurer may buy project debt, a private-credit vehicle may fund another layer, and the ultimate revenue may depend on a hyperscaler’s lease. The risk does not stop at the organizational boundary of any one regulator’s traditional silo.

Financing growth and policing risk are not opposites

The same FSA policy document that calls for closer monitoring also says Japan needs stronger financial mechanisms to supply capital for corporate growth investment. That is an important balance. The policy is not “do not finance data centers.” It is closer to “finance them with underwriting and risk controls that remain credible even when enthusiasm is high.”^[1]

Japanese institutions have real advantages in this market: deep project-finance expertise, large pools of long-term insurance capital and longstanding relationships with global corporate borrowers. Those strengths can make Japan an important supplier of capital to the AI buildout. They can also make common exposures large enough to matter at home if the same assumptions are repeated across many institutions.

The FSA’s September priorities mark a useful threshold. AI infrastructure has become large enough that it is no longer only a technology story, an energy story or a real-estate story. It is now a prudential-finance story. Behind every rack of accelerators sits a chain of land, power, leases, loans, bonds and guarantees. Whether the AI boom ultimately produces attractive returns is uncertain. Whether banks and insurers should know exactly what they are exposed to is not.