Fidenza Macro Analyst Geo Chen Exits AI Holdings Amid Leverage Crashes and Oversupply Risks

Author: Geo Chen (Fidenza Macro)

Compiled by Deep潮 TechFlow

Shenchao Summary: Geo Chen of Fidenza Macro liquidated all his positions in AI semiconductors and infrastructure in June this year, and now he presents his full bearish rationale. His analysis spans the collapse of South Korean leveraged ETFs, the impact of Chinese models on proprietary AI standards, the turning point where compute supply shifts from shortage to surplus, and the risk of the Federal Reserve losing credibility amid stagflation pressures—essential reading for any investor holding AI-related assets.

March 2000. March 2008. January 2020. December 2022. Certain months stand out vividly in my memory—they were all turning points just before major market upheavals. I believe we’ll look back on this month with the same sentiment.

In June this year, I liquidated all my positions in AI semiconductors and infrastructure and moved into cash. Afterward, I took a vacation, enjoying a summer away from the market, expecting a dull sideways trading period with few opportunities.

The outcome was completely unexpected.

Everything that has happened over the past month has strengthened my belief that the stock market bull run has already peaked. I’m typically an optimist, so I didn’t reach this conclusion lightly. Unfortunately, developments in AI, the war in Iran, and Federal Reserve policy are converging to create stagflationary conditions, with more volatility ahead. In this article, I’ll break down my bearish reasoning point by point.

AI is the leader of this bull market and the primary driver of GDP growth. Without AI, this bull market lacks support. Strong earnings from hyperscale cloud providers and semiconductor companies have fueled this rally, leading some to argue that as long as earnings grow and fundamentals remain solid, the bull market will continue. But in reality, stock prices reflect capital flows and market narratives—earnings are just one of many drivers. Most bull markets peak before earnings begin to decline, and often end even before analysts start revising down their expectations.

Many bull markets feature an overshoot phase: low-quality capital pushes the market to new highs, often resulting in a parabolic price pattern. Low-quality capital refers to participants who operate with information asymmetry, are insensitive to price, and whose buying pressure is unsustainable. Overshoot phases are typically only clear in hindsight, but if you can identify low-quality buyers driving the final leg of the rally, you may be able to detect the overshoot in real time. This was precisely the framework I used to exit the crypto market in August 2025—I concluded that MicroStrategy and crypto treasury companies were providing exit liquidity for the cycle’s peak.

In this round of AI market momentum, low-quality capital primarily consisted of Korean retail investors who heavily bought 2x or 3x leveraged ETFs tracking SK Hynix, the KOSPI Composite Index, and other memory-related assets. This capital flow spilled over, pushing up valuations of companies across the AI supply chain benefiting from supply bottlenecks. These leveraged ETFs generated hundreds of billions of dollars in additional buying pressure, but this force was fundamentally unsustainable. Hedging demands from these products caused market makers to accumulate massive negative gamma exposure, forcing them to buy on up days and sell aggressively on down days—triggering violent price swings that led to liquidations and margin calls.

Citibank estimates that the aftermath of leveraged product liquidations has erased $38.7 billion to date, with online data showing 1.2 million accounts liquidated—equivalent to one in every 30 Korean adults being wiped out. The extreme speculation by Korean retail investors has surpassed anything I’ve witnessed in any other bull market. A sense of financial nihilism has driven many Korean retail traders to gamble everything, using leverage to amplify their positions simply because they feel they’ve arrived too late and must catch up:

Recently, a post on the workplace community Blind shared a story of someone suffering heavy losses during a margin call. The poster wrote: "SK Hynix and Samsung Electronics kept rising, but I felt I entered too late and panicked." He added: "I went all-in with my entire assets of 170 million KRW plus an additional 200 million KRW in unsettled margin, totaling 370 million KRW, only to see the Korean stock market crash the next day, resulting in massive losses."

— From Chosun Business

Even after enduring all this pain, momentum stocks remain highly crowded.

Chart: S&P 500 Momentum Leaders' Crowding (J.P. Morgan, currently at a high of 93.3%, nearing the peak before July 2026). Source: J.P. Morgan

Parabolic bull markets almost always end with prolonged bear markets. The more extreme the sentiment, prices, and leverage during the rally, the harsher the aftermath. Once margin calls are triggered, this capital is damaged and rarely recovers. For a bull market to return to new highs, these impaired funds must be replaced by new capital and entirely new narratives—this healing process takes a long time and may never occur.

Those waiting for a new narrative to reignite the AI bull market are likely to be greatly disappointed. If anything, the narrative over the past few weeks has only worsened. China’s AI lab, Moonshot AI, released Kimi K3, which outperformed Claude Fable on Code Arena.

Chart: Frontend Code Arena Rankings, with Moonshot's Kimi-K3 leading at 1,679 points, surpassing Claude Fable 5 (1,631 points). Source: Arena

Following in the footsteps of Moonshot, Alibaba has launched Qwen 3.8, a large open-weight model with 2.4 trillion parameters. Market sentiment is also shifting toward open-source models, with NVIDIA’s Jensen Huang being the latest prominent figure in AI to publicly support open source.

Global macro master Louis Gave once said, "When China enters, profits run away." This has played out in the electric vehicle and solar panel industries, and now the market fears that Chinese competition will commoditize intelligence. All signs point in the same direction: the cost of intelligence is converging with the cost of the compute required to power it. Users can achieve near-equivalent performance from open-source models at a fraction of the price of closed-source models, while also enjoying better data privacy and no vendor lock-in—making it increasingly hard to justify paying a premium for OpenAI and Anthropic.

Cheaper AI is good for end users, but it’s a nightmare for closed-source AI labs like OpenAI, Anthropic, and Google that have committed to massive expenditures on compute leasing or procurement. As their profit margins and market share erode, their ability to raise capital at high valuations diminishes, weakening their capacity to fulfill their compute commitments. OpenAI’s decision to delay its IPO until next year is likely due to a lack of confidence in achieving a $1 trillion valuation. SpaceX’s drop to $112 billion—27% below its IPO price—may further dampen their IPO prospects. Because OpenAI and Anthropic engage in circular transactions with other participants across the ecosystem, they have become single points of failure for the entire AI industry.

AI bulls argue that AI computing power, memory, optical networks, and other components are still in short supply, but this logic is flawed. Every commodities trader knows that supply shocks and bottlenecks feel most acute precisely at the peak of a bull market. By the time supply and demand rebalance, the bull market has often already reversed entirely. Often, shortages eventually turn into surpluses, leading to prolonged bear markets.

Given the computing power capacity already committed or under construction by emerging and hyperscale cloud providers, I wouldn’t be surprised at all to see an oversupply of computing power in one or two years.

Computing power has shifted from shortage to surplus, which can perfectly coexist with continued rapid growth in token consumption and AI model revenues. The rate of efficiency improvements in hardware and AI algorithms has outpaced the growth in user token consumption, causing total token spending to decline since June this year. Silicon Data’s token spending index shows that overall token spending has been steadily decreasing since peaking in June.

Chart: SDLLMTK Index (Token Spending Index), peaked in June and has been declining since. Source: Silicon Data

What would excess computing power look like? Abandoned data centers, breached commitments, and in some cases, debt defaults. The scene could get messy. The credit spreads on corporate bonds issued by data center and hyperscale cloud providers are sending a signal: massive investments in computing power are becoming an increasingly risky business decision.

Chart: AI data center bond spreads widen (Hut 8, QTS, Meta, etc.). Source: Bloomberg

Chart: AI Ecosystem Credit Risk Surges (5-Year CDS Spreads, SPCX Soars). Source: Bloomberg

The stock market no longer rewards hyperscale cloud providers that announce increased AI capital expenditures, but they are ignoring this signal and continuing to ramp up their investments.

Chart: Revised consensus estimates for capital expenditures for fiscal years 2024–2026. Source: Bloomberg

Google announced an increase in its 2026 AI capital expenditure from $195 billion to $205 billion, and its stock fell 7% that day.

I know this pessimistic scenario is hard to imagine, but recent history offers plenty of reminders. When the Strait of Hormuz was blocked in April, almost no one anticipated oil prices would drop back to $70 so quickly. In January this year, when silver traded at $120, very few expected it to fall to $55 within a year. In 2021, almost no one believed that the high-flying growth stocks from that bull market would plunge 80% to 90% the following year. Shortages can rapidly turn into surpluses, and positions can shift just as quickly.

I previously believed that a war with Iran would have limited long-term effects on the stock market, but my view has changed. This conflict is evolving into a protracted, intermittent quagmire. Iran’s hardliners have no intention of relinquishing their two key cards: their stockpile of enriched weapons-grade uranium and control over the Strait of Hormuz. Seizing these two assets would require a prolonged ground war, and even then, the likelihood of success is far from certain. The war also has the potential to escalate into a U.S.-China proxy conflict.

The U.S. Department of Defense estimates that the war has cost American taxpayers $37.5 billion so far, but this is likely an underestimate, as it does not account for economic costs or future expenditures needed to replenish equipment and ammunition to pre-war levels. Trump’s decision to drag the nation into an expensive, endless war without congressional approval will go down in history as one of the most emblematic examples of the erosion of American democracy.

Over the past six years, the world has experienced four inflation shocks (the COVID-19 pandemic, the Russia-Ukraine war, Trump tariffs, and the blockade of the Strait of Hormuz). Each led to tighter monetary policy and significant market pullbacks. A war involving Iran could be the most persistent force of stagflation, as it would impact global energy and commodity supplies while raising government financing costs.

Kevin Warsh is attempting to completely overhaul the Fed’s methods for measuring inflation, responding to inflation, and communicating with the public—under the dual pressures of a supply shock and the bursting of an AI-driven bull market bubble. It’s like changing all the parts of an airplane while flying through a storm.

Amid soaring inflation and expanding fiscal deficits, Walsh faces two bad options. He can tighten prematurely to flatten the yield curve, but this risks triggering a recession. Alternatively, he can delay tightening and let the long end of the bond market handle it. He currently appears to have chosen the latter.

At yesterday’s FOMC meeting, the Fed had the opportunity to validate Wash’s hawkish remarks with a rate hike, but they did not. The bond market responded with a sharp bear steepening, a signal that the Fed’s credibility in controlling inflation is eroding. Joseph Wang noted that while Wash pledges price stability and a 2% inflation target, he is simultaneously altering the methodology for measuring inflation in ways he cannot publicly acknowledge. Without a clear framework, bond investors lack a foundation for anchoring expectations, causing volatility to rise in ways that equity markets struggle to absorb.

Long-term yields broke above 5.2%, reaching a two-year high. This technical breakout signals the start of a new phase in the U.S. Treasury bear market, bringing fresh headwinds to the stock market.

Chart: U.S. 30-Year Treasury Yield Weekly, Breaks Above 5.2% to Reach a Two-Year High. Source: Bloomberg

Disclaimer: The content of this communication is provided solely for informational and educational purposes and does not constitute professional financial advice, investment recommendations, or an offer to buy or sell any securities or instruments. This communication is not a trading signal service. The author strongly advises readers not to follow any trading actions without relevant market experience and research. The author is not a registered investment advisor or financial planner. The information presented in this communication is based on the author’s own research and experience and should not be construed as personalized investment advice. Any investment or trading decisions you make based on this communication are undertaken at your own risk. Past performance is not indicative of future results. All investments carry the risk of loss, and any trades or strategies discussed in this communication do not guarantee profits or suitability for your specific circumstances. The author assumes no liability for any actions taken or not taken based on the content of this communication. The author shall not be held responsible for any losses, damages, or liabilities arising from the use or misuse of the information provided.