Societe Generale's top bear: The AI boom is replicating the Asian financial crisis, with the deadly risk being the "debt time bomb"

Societe Generale's Chief Strategist Albert Edwards has issued a warning: the current AI investment craze bears a striking resemblance to the 1997 Asian financial crisis—not because the technology itself is useless, but because the pace of capital inflows is vastly outpacing the rate of productivity improvement, and historically, it is creditors—not engineers—who fill this gap.

In his latest "Global Strategy Weekly," Edwards points out that Total Factor Productivity (TFP), which measures the actual benefits of technology, has shown no improvement so far, yet global AI capital expenditure is surging rapidly. Goldman Sachs predicts that in 2026 alone, global AI investment will exceed $1 trillion. At the same time, OpenAI's annualized revenue was revealed to be about $20 billion less than previously expected, sending the Nasdaq down more than 1% in a single day, directly challenging the narrative of "strong demand."

In Edwards' view, the real trigger of this crisis is whether the supply of cheap money fueling AI expansion can continue—not disappointing productivity data. The "vigilantes" of the bond market are testing the fragility of various national markets one by one—from Japan to France, bond sell-offs are spreading. As AI tech giants borrow heavily to finance expansion, they rely on the same group of duration buyers to absorb hundreds of billions in new bonds, a dynamic reminiscent of Asia's reliance on short-term dollar borrowing to sustain prosperity in the 1990s.

This round of warnings starts with a rather dull chart. Apollo's Chief Economist Torsten Slok published a report titled "No Sign of AI in Productivity Data." The data shows that, after adjusting for capacity utilization, TFP is currently slightly below zero, and since the AI capital expenditure cycle began, there have been no signs of acceleration, while the hourly output growth rate has remained steadily around 2.5%.

Slok points out that strong per-hour output growth combined with stagnant TFP is a typical feature of "capital deepening" rather than a signal of technological shock—giving every employee a new monitor (or a $40,000 GPU) will increase per-hour output, but it won't make the company truly more efficient. The AI craze "is clearly visible in investment data and stock valuations, but not yet reflected in productivity statistics," meaning that the so-called returns "are still forecasts rather than facts."

Bank of America's strategist Michael Hartnett has also noted this pattern and added an important detail: TFP and the Consumer Confidence Index have been highly correlated over the past half-century, and both are currently in decline. Chicago Fed President Austan Goolsbee has previously warned that sustained weak productivity will challenge the current narrative.

Edwards has not doubted "miracle narratives" for the first time. He recalls, "the first major, extremely risky, non-consensus call" of his career was to point out that the "East Asian economic miracle" was essentially a massive economic and financial bubble. His theoretical foundation came from economist Paul Krugman's 1994 Foreign Affairs article "The Myth of Asia's Miracle"—Krugman argued that growth in Asian economies was driven by massive inputs of capital and labor, not genuine efficiency gains, and that weak TFP growth fundamentally undermined the bullish narrative.

Mainstream views scoffed at this at the time. The World Bank published "The East Asian Miracle" in 1993, while the "pinnacle of arrogant optimism" occurred in August 1996, when another World Bank report hailed Thailand's macroeconomic miracle—less than a year before the baht collapsed.

Edwards jokingly refers to his bearish framework from those days as "Noddynomics," which was "universally ridiculed" by the market. When presenting in the U.S., he compared the late-1990s U.S. tech bubble to the Thai economic bubble and even needed his then-boss to protect him from angry clients. Both bubbles eventually burst.

Edwards' second indictment comes from his former colleague Rob Parenteau's research. Parenteau notes that, driven by AI, total corporate investment growth is impressive, but after deducting depreciation, net business investment is almost flat: in nominal terms, total investment accounts for about 14% of GDP, while net investment has stagnated around 3% for nearly a decade; in real terms, total investment as a percentage of GDP has reached a record high at about 15.5%, but the net real value is only about 3.5%, roughly on par with 2015 and 2019.

Edwards also points out that companies are extending the depreciation periods of assets such as GPUs, making net investment data look better on paper but masking weak actual economic returns. This echoes Michael Burry's previous criticism of the overextended estimated lifespan of GPUs.

Goldman Sachs' calculations further quantify this risk. According to their latest research, if the six major U.S. hyperscale cloud providers have a return on invested capital (ROIC) of zero for their AI capital spending, depreciation and operating costs alone would require about $920 billion in annual revenue to cover.

Goldman Sachs estimates AI capital expenditure in three phases: about $633 billion from 2023 to 2025, about $1.73 trillion from 2026 to 2027, and as much as $4.14 trillion from 2028 to 2030. To achieve just a 15% ROIC for the second phase, the six hyperscale vendors would have to generate a total of about $1.42 trillion in revenue from 2028 to 2030.

GDP-level data has also failed to offer much support for the bulls.

Edwards notes that U.S. corporate fixed investment's year-over-year contribution to GDP is less than one percentage point, just a fraction of the late-1990s peak (more than 2 percentage points).

Structurally, equipment investment has a nominal growth rate close to 14%, but the real growth rate is "barely in double digits"; nonresidential construction expenditure is contracting, down about 3% nominally and nearly 6% in real terms—even with data center construction included.

On this basis, Edwards concludes that much of the "AI boom" in GDP accounts is reflected in price increases rather than output expansion, with the fundamental reason being that all participants are simultaneously scrambling to buy the same chips, memory, and transformers, driving up prices. This also helps explain why the Federal Reserve's worry is about AI-induced inflation, not deflation.

Edwards points out in his report that the true root of the Asian crisis was not disappointing productivity data, but the sudden stop of cheap external capital that propped up resource misallocation. He warns that the bond market "vigilantes" are operating at the same pace: recently, Japanese fund repatriation led to sustained selling of U.S. Treasuries, which then spread to France, where French government bonds are now heading for their worst decade since 1803.

Against this backdrop, AI tech companies are still issuing large amounts of bonds—Broadcom, Oracle, SpaceX, and others have joined the AI chip financing trend, further increasing supply-side pressure in the bond market, directly competing with hyperscale vendors’ need for long-term capital buyers.

Edwards’ conclusion is direct: Thailand’s miracle did not die because of disappointing productivity, but ended abruptly the moment creditors took notice. The biggest risk facing the current AI boom is a repeat of the same scenario—only this time the financing tools have shifted from short-term dollar loans to investment-grade bonds, private credit, and special purpose vehicles.

Edwards cautiously notes that he is not declaring the death of the AI boom, but rather asking: when the only statistical data that could prove this isn’t a bubble refuse to cooperate, why is market consensus so unwavering? The answer to this question may only become clear after the tide of capital recedes.