SocGen's Edwards: AI Debt Expansion Is Replicating the Asian Financial Crisis
nashnova research
SocGen chief strategist Albert Edwards warns the AI investment boom is structurally similar to the run-up to the 1997 Asian financial crisis — the danger is not the technology itself but the cheap debt underwriting capital misallocation, which will trigger a crisis the moment it recedes.
Why compare the AI boom to the Asian financial crisis?
Edwards argues in his latest *Global Strategy Weekly* that the 1990s "East Asian miracle" narrative and today's "AI changes everything" narrative share the same fatal structure: cheap capital flooding into an unverified high-return expectation.
This means → the trigger is never technology failing — it is creditors suddenly losing faith in the story. Thailand in 1997 followed this script; today's AI investment chain faces the same risk.
A direct catalyst has already surfaced: OpenAI's annualized revenue came in roughly $20 billion below prior expectations, sending the Nasdaq down over 1% that day and dealing a direct blow to the "robust demand" narrative.
Why don't productivity numbers support the AI boom?
Apollo chief economist Torsten Slök reports that capacity-utilization-adjusted total factor productivity (TFP) — the measure of how much technology genuinely lifts economic efficiency — is currently slightly below zero, with no acceleration since the AI capex cycle began.
In plain terms = hand every worker a $40,000 GPU and hourly output rises, but the firm is not actually more efficient. Slök calls this "capital deepening," not a technology shock.
BofA strategist Michael Hartnett adds: TFP and the consumer confidence index have moved in lockstep for half a century, and both are now declining. This reflects a hard reality — AI returns "remain forecasts, not facts."
Companies are spending heavily — why isn't net investment rising?
Research by Edwards's former colleague Rob Parenteau reveals that while AI-driven gross investment growth looks impressive, net business investment — after depreciation — is essentially flat. Nominal gross investment runs at about 14% of GDP; net investment has stalled near 3% for roughly a decade.
This means → most of what companies are spending simply replaces equipment that is wearing out. Genuinely new productive capacity is minimal.
A subtler issue: firms are stretching the depreciation life of GPUs and similar assets, making the books look better while masking weak real returns. Investor Michael Burry has separately criticized the same practice.
How much revenue is needed just to break even?
Goldman Sachs estimates that if the six major U.S. hyperscalers earn a zero return on AI capex, depreciation and operating costs alone would require roughly $920 billion in annual revenue to cover.
Goldman breaks AI capex into three phases: 2023–2025 at roughly $633 billion, 2026–2027 at roughly $1.73 trillion, and 2028–2030 at roughly $4.14 trillion.
In plain terms = just to achieve a 15% return on phase-two spending, the six hyperscalers would need to generate a cumulative $1.42 trillion in revenue between 2028 and 2030 — a figure no current revenue-side data can support.
Is the AI boom in GDP data real?
Edwards notes that U.S. business fixed investment contributes less than one percentage point to year-on-year GDP growth — a fraction of the late-1990s peak, when the contribution exceeded 2 percentage points.
Equipment investment is growing at nearly 14% in nominal terms but "barely reaches double digits" in real terms. Non-residential construction spending is contracting — down about 3% nominally, closer to 6% in real terms — even including data-center builds.
This means → much of the "AI boom" in GDP accounts reflects rising prices, not expanding output. Everyone is bidding on the same chips, memory, and transformers at the same time, pushing prices up.
How does the debt "time bomb" get detonated?
Edwards's core thesis: Thailand's miracle did not die because productivity disappointed — it stopped the moment creditors noticed. Bond-market "vigilantes" are now operating on the same rhythm.
Japanese capital repatriation has driven sustained selling of U.S. Treasuries, which then spread to France — French government bonds are heading for their worst decade since 1803. Meanwhile, Broadcom, Oracle, SpaceX, and other AI-linked firms continue to issue debt aggressively, adding to supply pressure.
In plain terms = the funding instruments have shifted from 1990s-era Asian short-term dollar loans to investment-grade bonds and private credit, but the script is the same. Edwards asks: when the only data that could prove this is not a bubble refuses to cooperate, why is the market so certain?
市场有风险，内容仅供研究参考，不构成投资建议。