NVIDIA's Strong Earnings Drive Semiconductor Rally

NVIDIA's recent earnings were excellent, surpassing market expectations. Lee Hyeong-soo, CEO, commented, "NVIDIA's performance was largely without issue, and the content of the earnings itself was very good." Notably, the company provided high revenue growth guidance of 70% for the fiscal year 2028, significantly exceeding the previous consensus of 40-45%. This strong performance and guidance are contributing factors to increased positive sentiment across the semiconductor market.

NVIDIA's order backlog has increased by an astounding $160 billion from the previous $120 billion, now approaching a total of $280 billion. Lee Hyeong-soo stated, "It was about $120 billion before, but an additional $160 billion was added to that." Such a large order backlog represents significant pre-orders directly linked to memory manufacturers' profit generation, suggesting that demand for memory semiconductors will remain robust in the future.

NVIDIA announced plans to increase AI server prices by 15% starting next year. This is interpreted as an intention to successfully pass on increased memory prices to hyperscalers. Lee Hyeong-soo explained, "They said that memory prices went up, so they had no choice but to raise the prices. So, from the memory perspective, there was nothing bad at all." This is a very positive situation for memory manufacturers and is expected to contribute to improved profitability.

While traditional general servers required a payback period of over 5 years, the period for AI data centers is significantly shortening. Lee Hyeong-soo mentioned, "Recently, for AI data centers, Amazon officially stated that it's recoverable within 3 years, and companies like SpaceX say it's recoverable within 1 year." With Amazon stating 3 years and SpaceX 1 year, the shortened duration has substantially reduced the burden of large-scale AI infrastructure investments. This creates an environment where AI-related investments can become more active.

Former President Trump's review of additional tariffs could act as short-term market noise. Lee Hyeong-soo analyzed, "It's short-term noise. It's about imposing additional tariffs on finished products, but this is a self-defeating move for US companies." The current memory and foundry markets are seller-dominant, meaning buyers must bear the tariffs. It is also projected that this could, paradoxically, serve as a mid-to-long-term opportunity for Korean companies like Samsung and Hynix, who are investing in the US, to strengthen their competitiveness.

Demand for FCBGA (Flip Chip Ball Grid Array) substrates for LPU (Logic Processing Unit) packaging is rapidly increasing. Lee Hyeong-soo explained, "It seems like the FCBGA substrate companies, which are for packaging LPUs, also received a good amount of these orders." Accordingly, related substrate companies are expected to benefit significantly. Samsung Electronics is responding to this demand by expanding production facilities, such as converting its Sejong plant parking lot into a factory site.

Producing HBM4 base dies, a next-generation high-bandwidth memory, requires a 4nm logic process. Lee Hyeong-soo pointed out, "The problem is the HBM4 base die, right? The HBM4 base die needs to be made with a logic chip, and that's 4nm." This means Samsung's HBM strategy is directly linked to foundry capacity allocation, demonstrating that securing foundry competitiveness is crucial for gaining an advantage in the HBM market.

Both 4nm and 2nm processes have high added value. Lee Hyeong-soo explained, "Is 4nm a low-value-added process? No. The added value is very high. The important thing is 2nm, right? 2nm is the most advanced process, and its added value will also be high. For 2nm, they lowered the price." While Samsung implemented a 10-15% price increase for North American and Chinese customers, it is attempting to secure competitiveness by lowering prices for the most advanced 2nm process.

While TSMC has built a strong ecosystem through its Open Innovation Platform (OIP), Samsung Foundry is noted for lagging in ecosystem expansion. Lee Hyeong-soo analyzed, "An ecosystem cannot be built in a short period. Samsung, as a company, was originally focused on memory, so basically, memory has a lot of vertical cooperation. TSMC, from its inception, was accustomed to horizontal collaboration." This seems to stem from the difference between memory-centric vertical corporate culture and foundry's horizontal collaboration culture.

If Samsung Foundry's utilization improves, OSAT (Outsourced Semiconductor Assembly and Test) companies are expected to benefit significantly. Lee Hyeong-soo commented, "If Samsung Foundry's utilization improves, the beta is said to be large. If Samsung Foundry improves by 10, it improves by 230... These are OSAT, which are back-end specialized companies. Such companies have recently been raising their heads, and their market capitalization is large." The stock performance of related companies like Hana Micron is positive, and their weighting in ETFs is increasing, leading to growing interest in these companies.

The trickle-down effect of foundries within the domestic value chain remains minimal compared to DRAM. Lee Hyeong-soo explained, "The trickle-down effect in the domestic value chain, based on DRAM, approaches almost 50%. But for foundries, it's still only about 15-20%." Nevertheless, Samsung Foundry's current utilization rate is higher than its past five-year average, with a production capacity of 500,000 wafers per month.

From the perspective of a semiconductor company, new investment in legacy (older) processes is a difficult decision. Lee Hyeong-soo said, "It's not easy for a semiconductor company to invest in a new legacy factory right now. It takes two years to build and complete a factory." Since it takes about two years to construct and complete a factory, investment decisions become even more challenging when it's difficult to predict market conditions two years down the line. This uncertainty is a major factor causing hesitation in new investments.

The semiconductor market is at an inflection point, transitioning from a P (price) centric model to a Q (volume) centric one. Lee Hyeong-soo explained, "Going from P to Q doesn't mean P breaks and Q goes up. The important thing is that for P to Q, P should be maintained or rise slightly, which is good. Currently, P itself is not breaking but rather rising, but the important thing is that the growth rate itself is slowing down." While price (P) itself is not declining and is maintaining or rising, there is concern that the growth rate of volume (Q) may slow down. Market investment ideas regarding Q have not yet reached a consensus, leading to differing opinions.

Analysis suggests that revolutionary memory reduction solutions are difficult to commercialize in the current situation. Lee Hyeong-soo stated, "In the current situation, that seems even more impossible to me. Rather, the focus of innovation now is less on using less memory and more on accelerating the efficiency of the infrastructure itself, so solutions that try to use less memory immediately are difficult to commercialize." This is because replacing existing standards incurs enormous costs. Therefore, the direction of AI infrastructure standards led by NVIDIA is focused on accelerating the efficiency of the infrastructure itself rather than memory reduction.

Chinese-made memory produced in China can be applied to domestic market products such as smartphones and PCs. Lee Hyeong-soo explained, "Chinese-made memory for internal consumers in China, like in smartphones or PCs, can be applied, but it cannot enter the servers of hyperscalers." However, entering key AI infrastructure markets, such as hyperscaler servers, is currently difficult due to quality gaps. While the threat is low for now, caution may be needed if China's yield improves.

The biggest Achilles' heel in China's semiconductor process is its dependence on ASML's lithography equipment. DRAM, in particular, has an absolute dependence on lithography equipment. In contrast, for NAND, in advanced processes, that is, shrinking, the importance of lithography is relatively lower, and technologies like deposition and etching are more crucial. Lee Hyeong-soo stated, "China's Achilles' heel in its current semiconductor process is ASML's lithography equipment. For NAND, in advanced processes, that is, shrinking, the importance of lithography is somewhat lower, and relatively, technologies like deposition and etching are more important." It is known that China had already secured inventory of state-of-the-art deposition and etching equipment before the US regulations.

Analysis suggests that resolving the undervaluation and reevaluating the Korean stock market will take time. Lee Hyeong-soo commented, "Ultimately, a reevaluation of the Korean market is needed. That won't happen immediately. Korea is still included in the MSCI Emerging Index, and trust has not yet been built." Since Korea is still part of the MSCI Emerging Index, there is a strong tendency for conservative investment. For long-term capital inflow, the sustainability of shareholder return policies must be sufficiently verified. Furthermore, market recognition that the cyclical nature of the memory industry has changed must be a prerequisite.

The key interest rate affecting semiconductor stock prices is the US Federal Reserve's interest rate, not Korea's. Lee Hyeong-soo emphasized, "Actually, the important interest rate for us is the US Fed's rate, not Korea's. If the US Fed raises interest rates, then stock prices could be severely impacted. However, the Bank of Korea's interest rate is still lower than the US interest rate." If the US Fed raises interest rates, it could burden hyperscalers' fundraising, which could be fatal to semiconductor stock prices. There is also a perspective that a US interest rate hike could be a bad move for the US itself amidst the US-China AI hegemony struggle.