21 Comment: How long can this round of AI investment boom last

Article body

Since October 2026, share prices of global AI hardware companies have suffered a significant correction, and a pessimistic voice has begun to emerge in the market, arguing that the AI boom built on massive capital expenditure is heading toward a "Minsky moment," and even speculating that the bubble may burst around 2028, triggering concerns about the outlook for tech stocks.

Objectively examining the current business model and financing structure of the AI industry, market concerns about an AI bubble are not without reason, but if one looks through the micro-structure of the balance sheets of the relevant entities, one will find that the future risk release of the AI industry will show a distinct structural stratification, and the endgame of the AI infrastructure wave may not be a 2008-style systemic financial risk, but rather closer to the repricing of tech infrastructure and industry reshuffling seen when the internet and telecom bubble burst in 2001.

At present, US tech giants spend up to a trillion dollars a year on AI construction, yet this huge investment is financially to a large extent a product of circulation within the ecosystem: the Bank for International Settlements (BIS) recently published its 137th bulletin, showing that between 2021 and 2025, as much as 55.2% of the investment received by AI companies came from other AI companies, and nearly half of that (46.4%) involved transactions where both parties simultaneously had upstream-downstream buyer-seller relationships.

True industrial sustainability should be based outside the ecosystem, but multiple estimates show that for current computing power investment to earn a reasonable return, global AI industry annual revenue would need to leap to 2.5 trillion to 3 trillion US dollars by 2030, about 14 times the current level. Against the backdrop of open-source (open-weight) models rapidly closing the technology gap, exponentially falling token call prices, and limited willingness of the public to pay, this commercialization gap is difficult to bridge with grand narratives alone.

More critical is the "second derivative problem" in the cycle evolution: the AI computing power chain does not need an absolute decline in tech giants' capital expenditure; merely a "slowdown in the growth rate" of capital expenditure is enough to detonate upstream pressure. The reason is that many upstream data center operators and emerging computing power providers do not have deep pockets themselves; they rely on forward computing power purchase commitments signed by downstream AI labs as collateral to borrow heavily from the bond market and private credit institutions to expand capacity. Once the growth rate of giants' capital expenditure peaks and slows, new contract cash flow cannot keep up with rigid maturing debt principal and interest and rapidly increasing depreciation costs, and highly leveraged entities will be the first to face liquidity exhaustion.

Another market concern focuses on the rapidly expanding debt and "off-balance-sheet hidden liabilities" of the AI ecosystem. As own cash flow is exhausted, debt has become the main financing channel for AI construction, and it is estimated that in 2027 alone global AI-related new bond issuance will reach 700 billion US dollars (of which hyperscale cloud vendors account for about 400 billion US dollars). At the same time, through off-balance-sheet joint venture companies (SPVs), long-term leases not yet commenced, and supplier financing guarantees, the off-balance-sheet hidden commitments of the entire AI ecosystem have reached as high as 1.5 trillion to 2 trillion US dollars.

However, simply analogizing this off-balance-sheetization to the "shadow banking black box" before the 2008 subprime crisis underestimates the pricing efficiency of modern capital markets and the risk-transfer capability of tech giants. On the one hand, unlike financial institutions in the subprime crisis that concealed risk purely to evade capital regulation, the approximately 1.1 trillion US dollars of lease commitments not yet commenced and 1.7 trillion US dollars of purchase commitments disclosed in the financial statement notes by North American hyperscale cloud vendors and chipmakers such as Nvidia and Broadcom are relatively transparent, and buy-side institutions and rating agencies have long added them back when assessing leverage; even under an extreme scenario where all guarantees and backstops are immediately and fully counted as liabilities, leading cloud giants still have strong debt-servicing capacity and borrowing room, and have not touched the red line of financial safety.

On the other hand, the real essence of the off-balance-sheet game is that the highest-credit-rated cloud giants (AA/A rated) use their bargaining power to transfer heavy-asset depreciation and forward overcapacity risk in a targeted way to weaker-credit emerging computing power lessors (such as CoreWeave and others) and upstream hardware suppliers. Once future computing power demand slows or token price cuts transmit to the bottom layer, cloud giants holding flexible renewal rights can simply not renew at maturity or push for price cuts in renegotiation, while those truly bearing the impact will be emerging computing power lessors that aggressively expanded capacity relying on high-interest debt (some rates as high as 9% to 15%), and chipmakers that provide huge computing power backstops and residual value guarantees to customers.

This explains why in the global market volatility in October, shares of traditional North American cloud giants such as Microsoft and Meta showed significant resilience, while high-yield bond spreads widened sharply, emerging computing power lessors such as CoreWeave plunged in a single day, and storage and hardware supply chains in Japan and South Korea led the declines.

Examining the development prospects of this AI cycle, its endgame is closer to the bursting of the internet bubble in 2001 than the subprime crisis in 2008. The most core firewall is that commercial banks have not deeply or directly engaged in high-risk AI computing power financing. The funds currently supporting AI infrastructure, apart from tech giants' own operating cash flow, mainly come at the margin from investment-grade corporate bonds (IG), 144A private placements, and private credit funds. According to calculations by the Federal Reserve Bank of Chicago, large commercial banks' drawn loans to AI-adjacent industries account for only 0.8% of their total assets (about 9% of Tier 1 capital), while long-term funds such as US life insurers, which account for a quarter of the investment-grade bond market, are also extremely restrained in their concentration of holdings in AI-related debt and single tech issuers. Even if a hundred-billion-dollar-scale credit default occurs in the non-giant AI ecosystem in the next two to three years, the cost will mainly be borne by equity investors as well as high-yield bonds and private credit funds, and it does not have the soil to pierce the banking system and brew a systemic financial crisis.

Overall, regardless of the outlook for the AI industry in the next two to three years, the period when a technological revolution truly reshapes total factor productivity often begins after the bubble bursts. The future decline in global computing hardware premiums and the fall in inference costs will instead help the vast number of real-economy enterprises bypass the heavy burden of early-stage depreciation. When highly leveraged computing power speculators are cleared out in the reshuffle and computing power returns to its nature as a public utility, the second half, in which artificial intelligence truly takes root in the real economy and empowers all industries, will officially begin.

(Source: 21st Century Business Herald) Middle of the body Embedded advertisement