AI Semiconductors: Investment Remains Valid Despite High Interest Rates

The Astra model is expected to greatly improve long-horizon task performance, and as new applications spread, token demand is expected to surge in the future.

Intensifying competition among model companies and the presentation of new paradigms are positive for the hardware industry, and the theory of slowing down is unlikely to materialize.

The U.S. government also lacks the will to regulate, and it is difficult to slow the pace of technological development amid the hegemony competition with China.

The Astra model, which applies a new architecture such as Loop Transformers, maximized the ability to complete instructed tasks from start to finish.

As usage expanded from coding-centric work to specific tasks of general office workers, inference performance was maximized through the application of new architectures such as Loop Transformers.

The launch race among model companies such as Anthropic ahead of its listing is leading to an expansion of the market size.

Loop Transformers adopted an architecture that improves performance by passing through layers repeatedly with the same parameters.

If parameter growth is limited, HBM capacity requirements may be somewhat lower, but instead the importance of bandwidth grows, raising the bar for hardware performance.

The reduction in HBM stack count is interpreted as a result of this architectural change combined with production realities.

Nvidia GPUs are expanding scale-up from 72 domains to 576 with Rubin Ultra and others.

As AI model demand spreads and inference computing increases, KV cache usage surges and the role of HBM is changing.

Starting with HBM4, the importance of bandwidth performance is growing due to bandwidth differentiation and intensifying competition among companies, and Samsung Electronics' recent improvement in HBM mass-production competitiveness and technological edge are mentioned.

Human-in-the-Loop refers to the concept of humans intervening in and controlling the decision-making process of AI models.

Concerns are being raised about areas of AI internal judgment that are difficult for humans to understand, such as the Hugging Face hacking incident.

The U.S. government prefers regulation that imposes strong legal liability when incidents occur rather than hindering innovation, and some argue that expanding context windows offsets the memory reduction effect.

Considering the order volumes secured by Big Tech and AI labs, demand for next year and the year after exceeds market expectations.

Contrary to ordinary investors' doubts, the order basis of industry leaders is clear and they have already secured specific order volumes.

As results actually prove out, stock prices are bound to rise, and strong demand is being confirmed in the earnings and data flows of major semiconductor companies.

Micron is scheduled to announce earnings on the morning of October 1 Korean time, drawing market attention.

Previously it provided only quarterly guidance, but there is a possibility it could mention next year's revenue growth rate (at least 70%), as in Nvidia's case.

If Micron presents high expectations like GPU companies when it announces earnings, market expectations are expected to heat up.

There is a conservative view that NAND has more suppliers than DRAM and faces stronger pressure from Chinese companies, creating higher price decline pressure.

The fact that bit growth from process generation transitions is larger than DRAM and that there is no core demand source like HBM are also cited as risk factors.

However, in specific areas such as data centers, entry by Chinese products is restricted, forming a dualized market.

Considering the structure in which DRAM shortages transfer to NAND, it appears unlikely that NAND will enter a down cycle, and some forecasts suggest the upcycle will last as long as DRAM's.

If competition restructuring eases infrastructure investment competition, spending capacity could plummet, and if the earnings growth of AI model companies such as OpenAI is not supported, external financing could become difficult and demand could plunge.

The current severe physical shortage appears likely to be resolved depending on whether model companies, the end customers, generate revenue.

There is a trend of focusing on inference model optimization to cut costs and improve profitability.

It is a process of efficiently allocating limited computing resources between training and inference services, and there have been cases in the past year where excessive investment in training compute worsened model competitiveness.

As maximizing annual recurring revenue (ARR) and maintaining frontier model capabilities become key goals, model companies are rushing to buy infrastructure after raising funds.

Due to this aggressive investment, memory and data center demand is expected to accelerate through 2028.

Despite rising costs, SK Hynix is making U.S. investments for customer proximity and supply chain stability.

This is interpreted as a strategic move to develop 3D stacking technology that directly places DRAM dies on top of logic dies.

Cooperation with Intel is also part of a memory-foundry technology convergence strategy, aiming to overcome bandwidth limits through logic-memory integration.

Semiconductor R&D is underway considering multiple directions, including HBM and single dies.

Before and after Anthropic's listing, supply-demand instability may arise in the tech sector, but if a correction occurs, it should be taken as a buying opportunity.

If Anthropic reveals large-scale infrastructure investment plans through an S-1 document or similar, it is expected to have a positive effect on the market.

Semiconductor stocks are more influenced by mid- to long-term infrastructure buildup plans than by short-term supply-demand burdens.

The structural trend of rising government bond yields as we enter a high-interest-rate era is expected to remain unchanged.

In a rising rate period, companies with earnings support and shareholder return capacity become more attractive, and AI infrastructure investment has the characteristic of having a lower macroeconomic impact than past B2C cycles.

In this environment, he projected that semiconductor stocks with strong earnings power will become more scarce than beneficiaries of abundant liquidity.

As demand exceeding supply growth continues through 2028, the positive outlook for the memory market is strengthening.

Samsung Electronics and SK Hynix are expected to see gradual stock price increases, and October earnings announcements and shareholder return policies are likely to support the downside.

In particular, he projected a high possibility of breaking through previous highs and setting new records within the first half of next year.