The global semiconductor market is undergoing a dramatic structural transformation. From the shift in foundry pricing power to the acceleration of self-developed chips by tech giants and the continuation of the memory super-cycle, multiple forces are intertwining to shape the industry’s trajectory. Ahead of TSMC’s (2330.TW) earnings call on July 16, the market’s focus has shifted from mere financial figures to the deep-seated rewriting of the entire AI supply chain’s logic.
Samsung Electronics recently announced its latest quarterly earnings, with profits surging approximately 18-fold year-over-year—a dazzling performance. However, its stock price plummeted consecutively after the release, surprising the market. Wu Jin-rong, General Manager of Microdrive Technology, analyzed that Samsung’s record profits followed by a stock decline is a classic case of “selling the news.” He pointed out that the memory market’s pricing mechanism is divided into contract prices and spot market prices. Contract prices are relatively stable, while spot prices are highly volatile, reflecting rapid changes in market demand. Looking back at the DRAM market during shortages, the contract price for 256K DRAM was about $3 (approximately NT$96), while the spot price once soared to $12 (approximately NT$400)—a fourfold difference.
With the rise of smartphones, the largest memory market once shifted to mobile devices, and recently it has further moved to the server sector, especially AI servers. Wu analyzed that AI servers and general servers combined now account for over 53% of total DRAM usage, and AI is still in its early stages. Regarding this year’s memory contract price trends, he estimates that in the first quarter, contract prices for Samsung, SK Hynix, and Micron (MU) rose by about 90% or more. The increase slowed to between 30% and 50% in the second quarter, and a 20% to 30% rise is still expected in the third quarter. Driven by AI, the overall memory market is transforming from a niche market into a massive one, and the silicon cycle caused by price fluctuations will be extended. This bull run “will last longer.”
Just as Samsung’s stock plunge triggered market concerns, the foundry industry signaled a major reshuffling of pricing logic. South Korean media, citing industry sources, reported that Samsung Electronics has raised foundry quotes for new customers, with prices for its 4nm, 5nm, and some automotive 8nm processes increasing by about 15%. Prior to this, TSMC had already notified core clients such as Nvidia, Apple, and AMD of planned price hikes for its 3nm, 5nm, and 7nm process nodes, with increases ranging from approximately 5% to 10%.
Market analysts suggest that Samsung’s move is not merely a reflection of rising production costs but a response to changing supply and demand structures, adjusting prices for specific popular processes to levels matching market conditions. Historically, as yields improved and capacity expanded, quotes for the same foundry process typically remained flat or gradually declined; reverse price hikes were rare. However, with cloud service providers (CSPs) aggressively building AI data centers and capital expenditures for next-generation 2nm process R&D and extreme ultraviolet (EUV) equipment continuing to soar, foundry pricing is increasingly determined by real-time supply-demand imbalances and massive investment costs rather than process maturity.
Analysts point out that the foundry market, long characterized by customer acquisition battles and fierce price competition, is shifting towards supplier dominance. The pricing power of foundries is rapidly transitioning from “customer-driven” to “supplier-driven.” The industry widely expects that with the AI investment cycle continuing, the upward pricing trend centered on advanced processes will be difficult to reverse in the short term. As foundry quotes rise, coupled with increasing High Bandwidth Memory (HBM) prices and persistently tight advanced packaging capacity, the overall manufacturing cost of AI chips will be further elevated. This cost pressure will not only raise the barrier for AI server deployment but is also highly likely to be gradually passed on, reflected in the retail prices of end-consumer electronics such as smartphones and personal computers.
Meanwhile, Meta Platforms’ AI strategy has also become a market focus. According to an internal Meta memo reviewed by Reuters, the company plans to begin mass production of an AI chip codenamed Iris in September this year, part of its “Meta Training and Inference Accelerator” four-generation chip program. The chip completed testing in just six weeks with no major issues found, representing significant progress for a self-developed chip project initiated over five years ago that had previously progressed slower than expected.
The Iris chip is designed by Meta according to its own needs, co-designed with Broadcom (AVGO), and manufactured by TSMC. This move is expected to reduce massive AI computing costs and lessen reliance on chip suppliers like Nvidia and AMD. Mike Gualtieri, Vice President and Principal Analyst at Forrester, stated: “If you have to rely on other companies for AI chips, you cannot become the AI hegemon. Hyperscale cloud service providers, and even SpaceX, are developing their own chips because that’s the only way to remain competitive on the cost of using AI models.”
The Meta memo shows the company plans to deploy 7GW of computing infrastructure this year and double its computing capacity next year, raising total capacity to 14GW. Meta estimates its AI infrastructure spending this year will reach up to $145 billion (approximately NT$4.7 trillion), accounting for a significant portion of the tech giants’ total estimated AI investment of over $700 billion (approximately NT$22.5 trillion) this year.
Morgan Stanley analysts noted that memory and other chip prices have risen rapidly and significantly recently, and “chipflation” is gradually becoming a macroeconomic issue worth monitoring.
Recently, Meta’s announcement that “computing power can be rented to external customers” led the market to interpret a potential oversupply of computing power, triggering a sharp sell-off in tech stocks. However, Wu Jin-rong believes that Meta still primarily relies on advertising revenue, and AI is for precise ad targeting. Its own chip development is still ongoing—somewhat slow but continuing—indicating that the overall market remains resilient and has not yet entered a cooling-off period.
Regarding the upcoming TSMC earnings call, Wu pointed out that according to pre-earnings call惯例, foreign institutional investors typically raise their target prices beforehand, but the stock price usually corrects slightly after the call. He emphasized that “any buying opportunity below 2,500 is excellent,” and Citi has even predicted TSMC’s stock price could reach 3,800. TSMC is a microcosm of the entire semiconductor industry, especially high-end semiconductors. The message TSMC delivers next week will reveal whether AI can continue to grow upward. Therefore, TSMC’s financial report and strategic layout this time will be a key observation point going forward.
Global chip stocks recently staged a broad rebound, recovering some of the losses from earlier this week. This followed a surge in China’s semiconductor stocks, which helped stabilize market sentiment. Additionally, Meta’s plan to double its AI computing capacity next year further alleviated market concerns about a slowdown in AI capital expenditure. AMD jumped 5.7%, Broadcom rose 3.2%, and Micron Technology gained 4.5%.
Chinese memory giant CXMT announced it will begin bookbuilding on July 15, preparing for an initial public offering (IPO) on the Shanghai Stock Exchange, expected to raise CNY 29.5 billion (approximately $4.4 billion). China’s CSI Semiconductor Index surged 8.8% at the close that day, with market investor sentiment clearly improving.
Overall, the semiconductor industry is at multiple inflection points. AI demand continues to drive the expansion of advanced processes and the memory market, foundries are gaining greater pricing power, tech giants are accelerating self-developed chips to reduce external dependence, and the memory super-cycle is expected to last longer amid the AI wave. The financial report and operational outlook released at TSMC’s earnings call next week will serve as a crucial litmus test for the market to validate these trends.