As of the morning of the 9th, Samsung Electronics and SK Hynix were trading at 272,500 won and 1.85 million won, respectively. The ripple effects of the global artificial intelligence (AI) race on Korea's memory semiconductor industry are growing once again. This focus is not about a simple new AI product launch.

"100,000 GPUs trained Astra"

It has been reported that more than 100,000 Nvidia Grace Blackwell NVLink72 units were deployed in the training of GPT-6 Astra. Nvidia CEO Jensen Huang said in a post on his X account that GPT-6 Astra was trained on approximately 100,000 or more NVIDIA Grace Blackwell NVLink72 units. He noted that it took four years from ChatGPT through o1 to Astra, assessed that "AGI has arrived," and sent congratulations to the OpenAI team.

With remarks adding that 400,000 more GPUs are expected to be brought online going forward, market attention has naturally turned to the Nvidia GPU supply chain. The structure in which the evolution of AI models directly translates into an explosive increase in semiconductor deployment has become even clearer.

HBM demand grows with GPU count

The key point is that as the number of GPUs increases, the scale of memory semiconductors required grows alongside it. AI accelerators must process massive data at high speed, which makes high-bandwidth memory (HBM) an essential pairing. The structure is such that as GPU performance improves and the number of interconnected GPUs grows, demands for memory bandwidth and capacity rise accordingly.

At this juncture, SK Hynix is mentioned first. SK Hynix has established itself as a major supplier in the HBM market for AI. As global big tech companies expand GPU investment for AI model training and inference, HBM demand is highly likely to increase in tandem, placing SK Hynix directly within the sphere of influence of AI infrastructure expansion. SK Hynix trading at 1.85 million won intraday on the 9th can also be read as a sign that the market is reacting sensitively to the possibility of expanding AI semiconductor demand.

Samsung Electronics cannot be left out of this trend either. Samsung Electronics holds not only HBM but also a broad portfolio of memory products needed to build AI data centers, including server DRAM, NAND flash, and enterprise SSDs. As AI investment grows, demand is unlikely to remain confined to HBM alone. Large-scale AI servers simultaneously require DRAM, storage, and network equipment, so expanding data center investment can ultimately broaden the demand base of the entire memory semiconductor market. With Samsung Electronics trading at 272,500 won intraday on the 9th, what the market is watching is its AI memory competitiveness and whether it can expand its position within the global supply chain.

From model competition to infrastructure competition

The biggest signal sent by GPT-6 Astra lies here as well. The AI industry is now moving beyond software competition over model performance and into infrastructure competition over who can secure more GPUs, memory, data centers, and power.

The fact that more than 100,000 GPUs were deployed to train a single ultra-large AI model, and that operating 400,000 GPUs is being discussed as the next step, shows that the capital intensity of the AI industry is rising to a level entirely different from the past. For the semiconductor industry, rather than how smart an AI model has become, how many semiconductors are needed to realize it may be a more important indicator. The structure is such that when GPU demand rises, HBM demand follows; and as data centers multiply, demand spreads to server DRAM, SSDs, power semiconductors, and network semiconductors.

Ultimately, the emergence of GPT-6 Astra can be seen as a symbolic event showing how far global AI investment can expand, going beyond the meaning of OpenAI's technological evolution. Samsung Electronics and SK Hynix are companies at the center of this massive investment cycle.

Points to note in interpretation

However, it cannot be concluded that AI infrastructure expansion will immediately translate into improved earnings for individual companies. The actual scope of benefits may vary depending on HBM supply contracts, customer composition, product-by-product supply volumes, prices, yields, and the competitive landscape. Moreover, CEO Jensen Huang's remark that "AGI has arrived" is his own assessment. Rather than concluding that GPT-6 Astra has achieved AGI in the sense agreed upon by academia and industry, it should be interpreted as a statement emphasizing the speed of AI technological advancement.

Nevertheless, one thing seems clear. As AI competition intensifies, the number of GPUs and memory units required is moving in the direction of increasing rather than decreasing. The trend shown by Samsung Electronics and SK Hynix on the 9th also needs to be viewed in the context of this global AI infrastructure expansion phase.

[This article was written with AI assistance based on publicly available materials and market information, and does not recommend buying or selling specific stocks or making investment decisions. Stock prices and market conditions mentioned in the article may change depending on the time of writing, and stock price volatility may increase depending on AI semiconductor demand, HBM supply contracts, semiconductor industry conditions, and corporate earnings. Final investment decisions and responsibility rest with the investor.]