• The BIS says the financing under the AI infrastructure boom is opaque and debt-heavy, and Asia’s data centre hosts sit downstream of any pullback.
• Hyperscaler capex passed US$1 trillion, and private credit lending to AI hit US$40 billion, on contracts that are barely disclosed.

The financing holding up the global AI infrastructure boom is more fragile, and far harder to see, than the build-out itself, and the Bank for International Settlements has now put numbers to the problem. In its Annual Economic Report 2026, the institution that serves as a central bank for central banks named the AI capital expenditure cycle as one of four pressure points facing the global economy. 

It went further than most commentary by tracing how the money is actually being raised and where it could break. That detail matters for Asia, which caught much of the upside of the build-out and now sits directly downstream of any reversal.

The five largest hyperscalers are set to spend more than US$1 trillion on AI-related capital expenditure across 2025 and 2026, a pace that is now outrunning their earnings and free cash flow. To cover the gap, they have started borrowing. Hyperscaler corporate bond issuance topped US$100 billion in 2025, most of it carrying maturities beyond five years, according to the BIS Quarterly Review. 

Credit default swap spreads on that debt have widened, a signal that bond investors are less convinced than equity markets that the projects will pay off. The less visible money is where the BIS reserves its sharpest language. Companion research published as BIS Bulletin No. 120 found that private credit funds originated over US$40 billion in loans to AI-related companies in 2025, up from roughly US$3 billion in 2010. 

The report describes a web of what it calls circular financing: chipmakers and hyperscalers take equity stakes in AI labs and neocloud providers, who then commit to buying chips and computing power from the same firms that funded them. Data centre construction is increasingly handed to third parties that lease the facilities back on long-dated contracts with embedded exit clauses. 

The terms, the BIS notes, are typically poorly disclosed, raising the risk of the same asset being pledged more than once.

What Asia’s AI infrastructure actually carries

This is the part that should concern operators and enterprise buyers in the region rather than only investors. Asia absorbed the demand spillover from the build-out through semiconductors and digital infrastructure, and nowhere more visibly than Malaysia. Johor alone drew roughly US$39 billion in data centre investment by the second quarter of 2025, and built more than 900 megawatts of capacity in about three years, a scale that took Singapore over a decade to reach. 

DayOne committed US$3.5 billion in Johor, and AirTrunk is putting RM12 billion into two new campuses next to its existing sites. Much of that capacity is leased, sub-leased and cross-committed between operators. DayOne, for instance, partnered with Oracle to service ByteDance. 

The BIS flags the suppliers at the tail of this chain, the engineering and construction contractors, as carrying comparatively weak balance sheets with little cushion if hyperscalers slow their spending. A region that hosts the AI infrastructure without owning the demand behind it is, by definition, holding contract risk that it did not originate.

There is a physical version of the same exposure. The BIS warns of a supply-side roadblock in electricity and the grid equipment needed to deliver it, with computing demand already pushing up power prices. Malaysia knows this directly. Electricity can account for 60 to 70% of a data centre’s operating costs, and the government has already paused approvals for non-AI facilities over power and water constraints. 

CBRE has cautioned that a slowdown or correction could follow if infrastructure and local capacity are not scaled responsibly.

Why does the speed matter more than the question?

Whether the spending pays off is genuinely unsettled, and the BIS is careful about it. Task-level studies show AI delivering time savings of 20 to 50%, yet aggregate productivity estimates sit below 1% over a long horizon, reflecting how hard the technology is to scale inside real production. 

The report’s most striking idea is a demand bottleneck: as automation moves income away from workers and into AI investment, the consumer base that justifies the next round of build-out could erode, stalling returns for reasons that have nothing to do with the technology getting better.

What has changed since earlier booms is the speed at which a reassessment could travel. Because so much AI financing now runs through hedge funds and private credit rather than banks, Zhang Tao, the BIS chief representative for Asia and the Pacific, told the South China Morning Post that a correction could move “much faster than previous banking crisis episodes.” 

The core technologies will keep improving, whatever happens. The contracts, leases and loans wrapped around the AI infrastructure are the part that can seize.

The operators turning Johor’s palm estates into compute have a more practical task in front of them than debating whether this is a bubble. They need to read the fine print on who is funding the tenant, and on what terms.