The AI hardware sector has seen two consecutive days of adjustment, yet the true market focus isn't on the chipmakers themselves, but on the latest moves from two major AI model companies.
On Wednesday, reports emerged that Meta Platforms, Inc. is exploring the commercialisation of its surplus AI computing power. A day later, further media coverage suggested Anthropic is in discussions with Samsung Electronics to co-develop proprietary AI chips, potentially using Samsung's 2-nanometer foundry process.
While these two developments may seem unrelated, they both touch on the most sensitive current topic in the AI supply chain: is the AI capital expenditure boom, which has expanded at a high speed for two years, now entering a new phase?
The market has been the first to react by repricing assets. US semiconductor stocks have broadly continued a sharp decline over the past two days, with the Philadelphia Semiconductor Index (SOX) falling over 10% cumulatively on Wednesday and Thursday, marking its largest two-day drop in nearly a month. The semiconductor equipment sector, most sensitive to capex cycles, led the losses. On Thursday, shares of Teradyne (TER), Entegris (ENTG), KLA Corp (KLAC), Applied Materials (AMAT), and Lam Research (LRCX) all fell by over 10% at one point, while the US-listed shares of European chip giant ASML (ASML) also dropped over 5%.
Catalysts for a Broader Review
In contrast, many institutions believe the two news items act more as catalysts for the market to re-examine AI investment logic, rather than signalling a fundamental reversal in the industry's momentum. What the market is truly trading on is not a question of whether "AI demand has peaked," but the idea that the AI industry is transitioning from a phase of "competing on capital expenditure" to a new stage focused on "capital efficiency."
The Core Concern: A Shift in Capex Logic
The market's real worry is not that Anthropic is making chips, but that the underlying logic for AI capital expenditure is beginning to change.
Over the past two years, the AI hardware sector's surge has been underpinned by a nearly unchanged core thesis: rapid AI model iteration has driven explosive demand for computing power, leading to a persistent shortage of GPUs. Tech giants have continuously raised their capital expenditure, in turn fuelling unprecedented demand for GPUs, High Bandwidth Memory (HBM), high-speed networking, advanced packaging, and semiconductor equipment—creating what has been termed an "AI capex super-cycle."
This logic not only propelled NVIDIA to become the world's most valuable company but also made equipment suppliers like Applied Materials, Lam Research, ASML, and KLA Corp, along with memory makers like Micron Technology and SanDisk, among the biggest winners in capital markets.
However, the two pieces of news emerging this week have prompted the market to seriously discuss a new question: if the AI industry starts focusing more on capital efficiency rather than simply expanding investment, will this capex super-cycle enter a new phase?
Wednesday's report suggested Meta is planning an AI cloud computing business, potentially opening up access to AI models deployed on its infrastructure to external clients or directly leasing out surplus AI compute to generate commercial returns on its tens of billions in AI infrastructure investment.
This was followed on Thursday by news of Anthropic's discussions on developing its own AI chips.
Individually, the companies are taking different paths, but together they point to a common shift—AI companies are starting to think about improving the return on their existing infrastructure investments, not just continuing to expand capital expenditure. It is precisely this change in expectation that has triggered a market reassessment of the AI trade.
Anthropic's Move Signals a Cost-Conscious Era
Beyond the initial market concern over whether "in-house chip development will reduce GPU purchases," the commercial logic behind Anthropic's move is more noteworthy.
Reports indicate Anthropic is in talks with Samsung Electronics to develop custom chips for AI training and inference, though the discussions are still in early stages.
If it proceeds, Anthropic would become another foundational model company, following Alphabet, Amazon, Microsoft, and Meta, in developing proprietary AI chips.
This does not signify an abandonment of NVIDIA GPUs but rather represents a natural evolution of the AI industry.
Over the past two years, competition among large model companies focused on who could secure more GPUs and build more data centers. As model scales continue to expand, driving up training and inference costs, the new competitive focus is shifting to reducing cost per token, improving compute utilization, and reducing reliance on single suppliers.
Application-Specific Integrated Circuits (ASICs) designed for particular models can achieve a better balance of performance, power efficiency, and cost. This is a key reason for the ongoing development of Alphabet's TPU, Amazon's Trainium, and Meta's MTIA in recent years.
In this sense, Anthropic's exploration of in-house chips is more a sign of the AI industry moving from "competing on investment" to "competing on efficiency," rather than a signal of cutting AI investment.
Different Paths, Same Objective for Meta and Anthropic
Meta and Anthropic are employing different strategies, but their goals are highly aligned.
Meta aims to generate revenue from temporarily idle AI compute, improving the return on its tens of billions in capital expenditure. Anthropic seeks to lower long-term computing costs through custom chips and enhance its autonomy over its infrastructure.
Whether it's selling surplus compute or developing ASICs, the essence is not about reducing AI investment but about finding a more sustainable AI business model.
However, for the capital markets, these two developments easily provoke another line of thought: if AI companies start paying more attention to capital efficiency, will the high-speed growth in GPU procurement, cloud computing rentals, and new data center investments seen over the past two years be sustained?
Consequently, the market has begun to reassess whether AI capital expenditure can continue the near "only-up" trajectory previously expected.
This explains why, during the recent two-day market adjustment, the steepest declines were not seen in model companies but in semiconductor equipment firms, which are most closely tied to new capital expenditure. Compared to GPU and memory vendors, equipment makers' orders more directly reflect future investment plans from foundries and chip companies, making them most sensitive to changes in capex expectations.
Institutional View: A Revaluation, Not a Rejection
Despite the recent sector-wide adjustment, most institutions are not interpreting the two news items as a sign of cooling AI demand.
Regarding Meta, many analysts believe selling surplus compute is more about finding a commercial outlet for massive AI capex, thereby improving the sustainability of future investments in GPUs, networking gear, data centers, and energy infrastructure, rather than about scaling back expenditure.
Regarding Anthropic, institutions generally view in-house chip development as aligning with the long-term trend for major AI model companies. Even as more firms adopt ASICs, they will still rely on advanced process manufacturing, HBM, high-speed interconnects, advanced packaging, and data center construction. Demand for AI infrastructure is unlikely to disappear but may be redistributed across different parts of the supply chain.
More importantly, AI application penetration remains at a relatively low level. Industry insiders point out that as inference demand continues to grow, token consumption and compute needs for large models are still far higher than previously anticipated. The build-out of AI infrastructure is still a considerable distance from true maturity.
Therefore, this week's market action resembles more of a periodic repricing of the AI trade following its historic rally.
If the AI competition of the past two years was about "who invests more," the signals from Meta and Anthropic suggest the industry is entering a new phase—where competition is shifting towards who can generate a higher return on every dollar of capital expenditure.
For the market, this shift in expectations is enough to act as a catalyst for an adjustment in the AI hardware sector. For the industry itself, however, it may not signify the end of a super-cycle. Instead, it could mark the beginning of a more mature development stage for AI infrastructure investment, one that places greater emphasis on a closed commercial loop.
Comments