How did OpenAI’s September 3 release of GPT‑6 Astra—described as highly capable, trained on more than 100,000 Nvidia Grace Blackwell systems, and followed by Nvidia CEO Jensen Huang’s claim that “AGI has arrived”—drive sharp gains in SK Hynix, Samsung Electronics, SK Square, and

GPT‑6 Astra acted mainly as a demand signal, not as proof that AGI has been achieved. Its reported training scale—more than 100,000 Nvidia Grace Blackwell NVLink72 systems—and Huang’s “AGI has arrived” endorsement reinforced investors’ expectation that model developers and cloud providers will keep accelerating spending on AI compute and memory. 57

Why Korean chip shares rose

• SK Hynix and Samsung Electronics were the clearest beneficiaries because larger AI clusters require substantial high-bandwidth memory, server DRAM and enterprise SSD capacity; investors therefore priced in stronger volumes, firmer memory pricing and higher earnings. SK Hynix, Samsung and Micron had already been lifted by an AI-driven equity rally. 3
• SK Square rose as the parent company and major economic beneficiary of SK Hynix: a higher valuation and earnings outlook for SK Hynix directly improves the perceived value of its stake.
• DB HiTek gained through the broader semiconductor-capex and supply-chain trade: a sustained AI infrastructure buildout increases demand not only for memory leaders but also for chip manufacturing, packaging, testing and related component suppliers. Reported trading showed DB HiTek up as much as 12.46% in the Astra-related rally. 3
• The market response was amplified by a narrative shift: if increasingly capable models demonstrate commercially useful reasoning, cybersecurity and computer-use capabilities, hyperscalers have a stronger incentive to build more clusters rather than pause AI investment. OpenAI itself described Astra as its most capable broadly deployed model and said it reached the “Critical” cybersecurity-capability level in its preparedness framework. 4

Why the DRAM shortage thesis strengthened

• The core issue is demand rising faster than usable supply. KB Securities projected $1.3 trillion of hyperscaler AI-infrastructure investment next year, up 60% year on year, with memory’s share of that spending rising sharply. 7
• Samsung and SK Hynix reportedly held fewer than 10 days of memory inventory in the third quarter, an unusually thin buffer that leaves little ability to absorb an unexpected demand surge. KB Securities characterized next year’s conditions as potentially the tightest in history. 7
• AI demand affects more than HBM: AI servers also consume server DDR5 and enterprise SSDs, creating simultaneous pressure across DRAM and NAND categories. 7
• Supply cannot be redirected freely. HBM4 uses roughly three times the wafer capacity of conventional DRAM, so increasing HBM output can reduce capacity available for standard DRAM and worsen shortages in both markets. 7
• The concentration of supply magnifies this effect. With Samsung and SK Hynix together accounting for roughly 64.3% of DRAM market share under the figures in the question, production constraints or allocation decisions at either leader materially affect global availability and prices.
• This is not merely a short-term broker call: SK Hynix’s chief executive has forecast the industry’s worst memory shortage in 2027 and said demand could outstrip the company’s production capability well beyond that point. 1

Important caveat: AGI is not established

Huang’s statement is an executive assessment and a powerful market catalyst, but it is not a scientific consensus. Researchers still disagree over the definition and measurement of AGI, and critics have challenged the claim that Astra meets that threshold. 46

For the memory-investment case, however, the label is secondary: the relevant economic fact is that frontier-model development is already requiring enormous GPU clusters and memory-intensive AI infrastructure. That supports the demand outlook even if today’s systems are ultimately classified as advanced narrow AI rather than AGI.