Chip stocks take another big dive, AI "horror stories" keep coming—is the market starting to re-evaluate?

By Li Dan and Ye Zhen, Wall Street Insights

The AI hardware sector has experienced adjustments for two consecutive days, but what has truly caught the market's attention is not the chip companies themselves, but the latest moves by two major AI model companies.

On Wednesday, news broke that Meta is exploring the commercialization of its surplus AI computing power. A day later, media reports indicated that Anthropic is discussing a collaboration with Samsung Electronics to develop its own AI chips, potentially using Samsung's 2nm process for foundry.

These two pieces of news seem unrelated, but together they touch upon the most sensitive topic in the current AI industry chain — whether the two-year rapid expansion of AI capital expenditure is entering a new phase.

The market has chosen to reprice first. U.S. chip stocks have continued to plummet over the past two days, with the Philadelphia Semiconductor Index (SOX) falling 11% cumulatively on Wednesday and Thursday, marking the largest two-day drop in nearly a month.

The semiconductor equipment sector, most sensitive to the capex cycle, led the decline. Teradyne (TER), Entegris (ENTG), KLA (KLAC), Applied Materials (AMAT), and Lam Research (LRCX) all fell over 10% at one point during Thursday's session, while European chip leader ASML's U.S. stock (ASML) once dropped more than 5% on Thursday.

Marvell closed down 9.84%, Arm fell 6.58%, Micron dropped 5.49%, AMD declined 4.26%, Broadcom fell 2.41%, NVIDIA was relatively resilient but still closed down 1.39%, and TSMC ADR fell 2.27%.

Goldman Sachs' basket of AI semiconductor stocks suffered heavy losses, recording their worst two-day performance since Tariff Day.

Memory stocks took a heavy hit. Goldman Sachs' basket of memory stocks fell more than 18% over the past two days, the most severe two-day decline in 12 years.

Storage maker SanDisk plummeted over 14%, falling about 27% from its recent high, entering bear market territory.

In contrast to the brutal performance of capital recipients like chip companies, the stocks of hyperscale cloud service providers, as capital spenders, have stabilized somewhat.

However, many institutions believe that the two pieces of news are more like catalysts for the market to re-examine AI investment logic, rather than a fundamental reversal in the AI industry's prosperity. What the market is truly trading is not "whether AI demand has peaked," but that the AI industry is moving from a phase of "competing on capital expenditure" to a new phase of "competing on capital efficiency."

What the market truly fears is not Anthropic making chips, but the change in AI capital expenditure logic

Over the past two years, the AI hardware sector has been on a tear, with the core logic remaining almost unchanged: rapid iteration of AI models drives sustained explosion in computing power demand, GPUs have been in short supply for a long time, tech giants have been continuously raising capital expenditures, which in turn drives demand for GPUs, high-bandwidth memory (HBM), high-speed networks, advanced packaging, and semiconductor equipment, forming an unprecedented "AI capex super cycle."

This logic not only propelled NVIDIA to become the world's most valuable company, but also made equipment makers such as Applied Materials, Lam Research, ASML (Netherlands), KLA, and storage manufacturers like Micron Technology and SanDisk the biggest winners in the capital markets.

However, two pieces of news over two consecutive days this week have prompted the market to seriously discuss: If the AI industry begins to focus more on capital efficiency rather than simply expanding investment, will this capex super cycle enter a new phase?

On Wednesday, reports indicated that Meta is planning to build an AI cloud computing business, potentially opening up AI models deployed on Meta's infrastructure to external customers or directly leasing surplus AI computing power, thereby achieving commercial returns on tens of billions of dollars in AI infrastructure investments.

Then, on Thursday, news emerged that Anthropic is discussing the development of its own AI chips.

Viewed separately, the two companies are taking different paths, but together they point to a common shift — AI companies are starting to think about how to improve the return on existing infrastructure investments, rather than simply continuing to expand capital expenditure.

It is this shift in expectations that has triggered a re-evaluation of the AI trading logic in the market.

Anthropic's self-developed chips: Does it mean AI companies are entering the "era of cost optimization"?

Compared to the initial market concern that "self-developed chips might reduce GPU purchases," the business logic behind Anthropic's move is more worth noting.

Reports say Anthropic is discussing with Samsung Electronics the development of custom chips for AI training and inference, though it is still in the early stages.

If it goes ahead, Anthropic will become another foundational model company to develop its own AI chips, following Google, Amazon, Microsoft, and Meta.

Behind this is not an abandonment of NVIDIA GPUs, but a natural evolution of the AI industry.

Over the past two years, the focus of competition among large model companies has been who can obtain more GPUs and build more data centers. However, as model scales continue to expand and training and inference costs rise rapidly, how to reduce per-token cost, improve computing power utilization, and decrease reliance on a single supplier has become the new competitive focus.

ASICs designed for specific models can achieve a better balance among performance, power consumption, and cost, which is also the key reason why Google's TPU, Amazon's Trainium, and Meta's MTIA have been advancing in recent years.

In this sense, Anthropic's exploration of self-developed chips is more like an important sign that the AI industry is shifting from "competing on input" to "competing on efficiency," rather than cutting AI investment.

Meta and Anthropic: Two different paths point to the same goal

Meta and Anthropic have adopted different strategies, but their goals are highly aligned.

Meta hopes to generate revenue from temporarily idle AI computing power, improving the return on its hundreds of billions of dollars in capital expenditure; Anthropic, on the other hand, aims to reduce long-term computing costs through custom chips and enhance its autonomy in infrastructure.

Whether it is selling surplus computing power or deploying ASICs, neither essentially reduces AI investment; rather, they are searching for a more sustainable AI business model.

However, for the capital market, these two pieces of news can easily trigger another association: if AI companies start to pay more attention to capital efficiency, will future GPU purchases, cloud computing leasing, and new data center investments maintain the high growth rate of the past two years?

The market has thus begun to re-examine whether AI capital expenditure can continue to maintain the previous expectation of "only increasing, never decreasing."

This is also why, in the two-day market adjustment, the biggest decliners were not the model companies, but the semiconductor equipment companies most closely tied to new capital expenditure. Compared to GPU and memory companies, equipment orders more directly reflect future investment plans of fabs and chip companies, making them the most sensitive to changes in capex expectations.

Institutions: The market is more like reassessing AI trades, not denying the AI super cycle

Although semiconductor stocks have been adjusting for consecutive days, most institutions have not interpreted the two pieces of news as a sign of cooling AI demand.

Regarding Meta, many analysts believe that selling surplus computing power is more like seeking a commercial outlet for massive AI capex, thereby improving the sustainability of future investment in GPUs, networking equipment, data centers, and energy infrastructure, rather than cutting capex.

Regarding Anthropic, institutions generally believe that self-developed chips align with the long-term development trend of large AI model companies. Even as more companies adopt ASICs, they will still need to rely on advanced process manufacturing, HBM, high-speed interconnects, advanced packaging, and data center construction. AI infrastructure demand will not disappear but may be redistributed to different segments.

More importantly, the penetration rate of AI applications is still relatively low. Industry insiders point out that as inference demand continues to grow, the token consumption and computing power requirements of large models are still far higher than previously expected, and AI infrastructure construction still has a considerable cycle before reaching true maturity.

Therefore, this week's market appears to be a phased repricing of AI trades after a historic rally.

If the AI competition over the past two years was about "who invests more," then the signals from Meta and Anthropic suggest that the AI industry is entering a new phase — competition is shifting toward who can generate higher returns on every dollar of capital expenditure.

For the market, this shift in expectations is enough to serve as a catalyst for adjustments in the AI hardware sector. But for the industry itself, it does not necessarily mean the end of the super cycle; rather, it may indicate that AI infrastructure investment is starting to move toward a more mature stage that emphasizes a closed business loop.