The world’s central bank watchdog flagged that an AI investment bust could cascade through global credit markets. The smart play isn’t to bet on which way the bubble goes — it’s to build systems that pay for themselves regardless.
In its Annual Economic Report released on June 28, the Bank for International Settlements flagged four pressure points threatening global economic stability: persistent inflation risks, sustainability of AI-related investment, growing financial vulnerabilities, and weakening fiscal positions.[1][2] The BIS coordinates central bank policy across 60+ economies — it doesn’t issue warnings lightly.[6]
The specific concern is that disappointing AI returns could trigger a sudden financing pullback, turning the current capital expenditure boom into a prolonged investment downturn. BIS General Manager Pablo Hernández de Cos said: “The race to capture market share may have led to overinvestment.” He warned this could leave firms “vulnerable to disappointments in AI payoffs.”[3]
For context, the BIS estimates the five largest hyperscalers are set to spend over $1 trillion on AI-related capital expenditure from 2025 through 2026.[1] These commitments are outpacing earnings and free cash flow, leading some firms to issue debt to raise additional financing.[5] If even a fraction of that spending is written off as overinvestment, the ripple effects hit chipmakers, data centre operators, cloud providers, and every company sitting in the middle of the supply chain.
The BIS draws parallels with historical investment booms — canal mania of the 1830s, railway mania of the 1840s, electrification exuberance of the 1920s, and the dotcom boom of the late 1990s.[1] Each shared one trait: a genuine technological breakthrough that attracted capital in excess of what commercial returns could ultimately justify. Each ended with an eventual reversal in investment, inducing economy-wide recessions.
The current AI build-out is also hitting supply-side bottlenecks in electricity, advanced semiconductors, and grid equipment. Fast-growing demand for computing power is pressuring electricity prices and input costs, with potential spillovers to inflation. Firms are locking in long-dated capacity contracts that further expose them to disappointments in demand.[1]
As the BIS warns: “Policymakers must act now. Delay will only make the necessary adjustments more costly.”[2]
In the AI boom scenario, the spending continues, valuations hold, and businesses that integrated AI early capture productivity gains. In the AI bust scenario, capex dries up, credit tightens, and companies that over-invested in AI infrastructure face write-downs.
Most Singapore businesses are watching the headlines and waiting to see which future wins. That’s the wrong frame.
The question isn’t whether AI booms or busts. The question is whether your AI investment pays for itself within months — or whether you’re betting on a multi-year payoff that depends on the broader market staying rational.
There’s a meaningful difference between spending $500K on a custom AI platform that you hope will transform your business in three years, and spending $50K on automating document processing that saves 40 hours of manual work per quarter.[4] The second pays for itself in the first cycle. The first is a bet.
Source: Business Times — “AI investment race could turn debt-fuelled boom to bust: BIS” (15 Jul 2026). businesstimes.com.sg
The AI applications that survive a correction are the ones with a clear ROI timeline. Three categories consistently deliver:
Document processing and data extraction. OCR-powered systems that pull structured data from PDFs, scanned invoices, and fund statements. A typical finance team spends 4–8 hours per month on manual data entry. Automating that returns 50–100 hours per year at a fraction of a junior analyst’s cost. See the related memo on AI-powered document processing.
Automated reporting pipelines. Systems that aggregate data from multiple sources and generate draft narratives for quarterly reports. A SET-listed company reduced its reporting cycle from 2–3 weeks to overnight processing plus hours of review.[M] The architecture behind that is straightforward — three layers: ingestion, processing, review.
Local AI inference. Running AI models on-premise eliminates ongoing API costs and keeps data sovereign. For Singapore businesses dealing with MAS compliance, this isn’t a premium feature — it’s the baseline. Local LLM deployment turns what would be a recurring cloud expense into a one-time hardware investment.
These aren’t speculative bets. They’re cost savings that compound from day one.
MAS reached the same conclusion earlier, and in more specific terms. Its 2025 review found that some Big Tech firms, primarily hyperscalers, had turned to novel and potentially circular private financing arrangements, including special purpose vehicles, private credit structures and accounting treatment that could mask leverage[7]. MAS stated the consequence plainly: a retrenchment of optimism in AI’s ability to generate sufficient future returns may lead to sharp corrections in the broader equity market and further defaults in the private credit market[7].
For a Singapore portfolio, the transmission runs through bank credit rather than equity. MAS stress-tests the domestic banks against an adverse scenario driven by a downturn in the AI-led global growth cycle[8], and the banking system’s total provisioning coverage reached 147% in Q2 2026[8]. The buffers are real. The scenario is the same one the BIS is describing.
Hernandez de Cos didn’t mince words: “Policymakers must act now. Delay will only make the necessary adjustments more costly.”[2] He was speaking to central bankers, but the principle applies to business leaders too.
The companies that position themselves now — building practical AI automation that delivers measurable ROI — will be in a stronger position whether the AI trade continues or corrects. The alternative is waiting for clarity that may never come before the market moves.
The BIS gave its warning. The question is whether you treat it as noise or as a signal.
Claims marked [M] are Kapital measurements from client deployments, not third-party data, and carry no public URL.