The toughest days for AI assets seem to be behind us, at least for now.

It's time to unpack what exactly happened during this plunge.

The plunge was not without warning; it followed a clear conduction (Note: ' conduction ' is translated as 'transmission' in broader contexts, but here it refers to the 'path of contagion' or 'chain reaction' in financial markets) path. Initially, the market began to worry about a slowdown in capital expenditures by cloud vendors. As some of these concerns gradually materialized, selling spread from cloud vendors to popular sectors like storage and connectivity, eventually engulfing the entire semiconductor supply chain.

Interestingly, while the 'little guys' in the supply chain were battered, the cloud vendors, which were the first to fall, were the first to stabilize. Throughout July, Microsoft rose 24.58%, while Amazon and Google increased by 13.95% and 0.94%, respectively.

Although this downturn covered most of the AI supply chain, the divergence in declines was striking.

Silicon-based analysts tracked 35 core companies in the AI supply chain, finding that over the past two months, these companies averaged a maximum drawdown of about 42%, but individual stock declines ranged from -20% to -74%.

Companies with core bottlenecks and pricing power fell less; assets whose profits are easily squeezed by upstream and downstream players, those with stronger cyclical attributes, or those highly reliant on marginal capital, bore greater drawdowns.

In other words, this downturn was not a simple indiscriminate sell-off but a revaluation of the entire supply chain by capital, seizing on panic.

So, today, Silicon-based analysts will delve into this matter.

/ 01 / The Plunge Began with 'Stagnation' in Cloud Vendors

The tide does not recede overnight.

The first signs of weakness appeared at the starting point of AI capital expenditures: cloud vendors.

Starting in May, cloud vendors collectively entered a period of stagnation. Amazon's stock price peaked on May 5 and then fell nearly 12% over the following month. Google and Microsoft were not spared, with declines of 6% and 17%, respectively, throughout June.

At the time, the entire AI supply chain was still in celebration. Throughout May, storage sector players like SK Hynix, Micron, and Samsung surged by an average of 71%, continuing to rise in June.

The connectivity sector was equally euphoric. Optical communication leader Corning rose 10.47% in May, with gains expanding to 41% in June; optical interconnect leader Marvell Technology rose 24.13% in May and another 45.31% in June.

The turning point came from market doubts about cloud vendors' capital expenditures. Starting in May, Goldman Sachs and Morgan Stanley released multiple research reports pointing to the unsustainability of cloud vendors' capital expenditures.

They projected that capital expenditures by the four major cloud vendors would reach $770 billion by 2026, consuming all operating cash flow for that year, with $1.8 trillion in off-balance-sheet exposure doubling the gross leverage ratio in two quarters, surpassing traditional energy sectors.

What truly turned expectations into reality was Broadcom's 'not bad' earnings report.

On June 3, Broadcom released its earnings, reporting $10.8 billion in AI semiconductor revenue for the second quarter, a record high. However, its Q3 guidance of $16 billion fell short of Wall Street's expectation of $17.2 billion, and it did not raise its 2027 target.

As a core supplier to cloud giants, Broadcom's conservatism directly confirmed market fears: cloud vendors' long-term computing power demand was slowing. The day after the earnings release, Broadcom plummeted 12.59%, shedding about $280 billion in market value in a single day, with Microsoft falling 3.17% and Amazon 2.53%.

The chill began to spread downward, first affecting chips and servers. In June, Nvidia began to decline, while AI server leader Super Micro Computer plummeted more than 36%.

Across the AI supply chain, storage and connectivity, which had surged the most this year, peaked last and fell the hardest.

The storage sector peaked on June 25, triggered by Micron's Q3 earnings.

Although all of Micron's data was impressive, management stated on the earnings call that the pace of storage price increases would slow significantly in the fourth quarter.

That single statement plunged the entire storage sector into icy waters. Throughout July, SK Hynix fell 35%, Micron 28%, and Samsung 21%. Maximum drawdowns were even more dire, with SK Hynix more than halving in value.

Once storage collapsed, the logic for the connectivity sector crumbled. The reason was simple: if even the most AI-logic-driven storage couldn't hold up, could the 'buy, buy, buy' myth of cloud vendors still be trusted?

In July, the connectivity sector collapsed across the board. Corning plummeted 45.88%, falling from $271.78 to $138.25; Marvell Technology and Coherent also retreated by 37% and 33%, respectively.

Finally, panic spread to the entire AI supply chain.

TSMC's stock price peaked on June 30, with a maximum drawdown of 22.19% in July; ASML peaked on the same day, with a maximum drawdown of 22.38% in July; U.S.-based semiconductor equipment maker Applied Materials saw a maximum drawdown of 41.01% from its June 30 high to the end of July.

Throughout July, the Philadelphia Semiconductor Index fell more than 20.61%, entering a technical bear market, with global chip stock market values evaporating by a combined $1.7 trillion.

Most interestingly, while the 'little guys' were battered, the 'big brother' who fell first—cloud vendors—stabilized and even rose. In July alone, Microsoft rose 24.58%, Amazon 13.95%, and Google 0.94%.

/ 02 / Behind the Divergence in Declines, the Market Completed a Revaluation

This plunge was not merely a simple valuation kill but a brutal revaluation of assets. In terms of declines, while all AI assets fell, the extent varied significantly.

Silicon-based analysts tracked 35 core companies in the AI supply chain, finding that over the past two months, these companies averaged a maximum drawdown of about 42%, but individual stock declines ranged from -20% to -74%.

Behind this divergence lie three entirely different types of assets.

The first category consists of core assets with pricing power, with drawdowns mostly around 20%.

Nvidia, TSMC, and ASML dominate the three hardest-to-replace segments: GPUs, advanced manufacturing processes, and lithography machines. While AI capital expenditures may slow, as long as the industry continues to invest, they remain unavoidable tollbooths.

Microsoft, Google, and Amazon follow a slightly different logic. While they don't control hardware bottlenecks, they possess the ability to convert AI investments into cloud revenue, software subscriptions, and long-term contracts.

In late July, Microsoft and Amazon released their earnings, with both stocks surging more than 10% in after-hours trading. The reason? The market began to reward companies that strategically exited model competition and returned to infrastructure fees.

The second category consists of assets close to AI but far from the profit pool, with drawdowns generally reaching 40%—60%.

New cloud vendors like CoreWeave and Nebius essentially purchase GPUs in bulk and then rent out computing power by the hour. The problem is that they can neither influence upstream GPU prices nor escape competition from giants like Microsoft, Amazon, and Meta, which continuously release computing power supplies.

With rigid upstream costs and intensifying downstream competition, profits are naturally squeezed on both ends.

Fiber optics, connectivity, and AI servers follow a similar logic. While these segments do benefit from AI investments, their technological barriers and industry concentration are far lower than those of GPUs. Customers can compare prices among multiple suppliers, and orders can quickly shift due to price, compliance, or delivery capabilities.

Take Corning as an example: while it manufactures fiber optic preforms, this is not an irreplaceable hard technology. To ensure supply chain safety, Nvidia invests in multiple vendors simultaneously, including Corning, Lumentum, and Coherent.

AI servers follow the same logic. It's a labor-intensive assembly business, with server manufacturers' gross margins year-round (Note: ' year-round ' is translated as 'consistently' or 'over the long term') pressed at around 10%. This is why Super Micro saw a 54.5% drawdown.

Notably, in this round of AI asset drawdowns, Hong Kong-listed AI assets generally fell more than their U.S.-listed counterparts. For example, YOFC, Kingboard Chemical, and Victory Giant Technology saw maximum drawdowns of 65%—74%.

The reason is that as an offshore market, Hong Kong is highly reliant on marginal capital. When the market rises, themes, scarce chip (Note: ' chip ' is translated as 'float' or 'shares available for trading') , and smaller float amplify gains; once capital withdraws, the same structure exacerbates declines.

Compared to U.S. leaders with pensions, passive funds, and long-term institutional holdings, some Hong Kong-listed AI hardware assets lack stable 'ballast.'

During rallies, limited float amplifies gains from capital inflows; once sentiment reverses, the same structure accelerates selling. The faster the initial rise, the less support there often is during drawdowns.

The third category consists of storage assets with the strongest fundamentals but also the greatest cyclical risks.

In this drawdown, storage, despite its strong fundamentals, became one of the hardest-hit sectors. SK Hynix, Samsung, and Micron saw maximum drawdowns of 40%—58%.

The reason is that the market never trades solely on current profits but on their sustainability.

The storage industry has followed the same cycle for decades: price hikes, expansion, oversupply, and price declines. HBM has raised technological barriers but has not eliminated cycles. Instead, higher profits strengthen manufacturers' incentives to expand.

Micron just reported a record quarterly profit, only to raise its 2026 fiscal year capital expenditures from $20 billion to $27 billion, with Samsung and SK Hynix also frantically building new factories.

Historically, high gross margins have not been proof of cycle disappearance but rather a signal of impending new supply releases. Amid continuous declines, such concerns have gradually amplified.

Reviewing this round of AI asset declines, it's less about the market abandoning AI and more about capital seizing on panic to complete a revaluation.

Over the past two years, almost any company on the AI capital expenditure chain could command a valuation premium. But this correction shattered the universal rally logic: companies with core bottlenecks and pricing power fell less; assets with easily squeezed profits, stronger cyclical attributes, or high reliance on marginal capital bore greater drawdowns.

The market no longer just looks at whether a company benefits from AI but begins to ask: How much profit can it retain from this industrial wave, and how sustainable is that profit?

The AI story is far from over, but the era where merely being associated with AI could command a premium is fading.

By Yuan Yuan