Meritz Securities (Korean sell side): GPT-6 Astra launch and implications for the memory stock rebound On September 3, OpenAI unveiled GPT-6 Astra. As interpretations of this model became a hot topic, semiconductor stocks rebounded on September 4 even as the broader US market fell on rising rates following a jobs surprise, with the DRAM ETF up 6.6% versus the prior day. Astra's implication is unlikely to be simply whether AGI has been achieved. What drew the most attention was the score of 99.9% on ARC-AGI-3, a benchmark used to judge AGI, a huge improvement over the previous model (Sol at 7.8%). This benchmark is not a knowledge test; it evaluates a model's ability to learn on its own in an abstract environment it has never seen before, and this is where the improvement over the previous model was large. General intelligence, as measured by AAII or Humanity's Last Exam, did not improve much. The differentiated strength improved in Astra is "the ability to acquire skills that humans learn in an unfamiliar environment as efficiently as a human does." Where AI until now found the answer by pressing this and that 100 times, Astra has started to behave more like a human: observing the phenomenon, inferring the rules, and executing right away. The key point is "an expanded scope for replacing human intelligence and human work." If existing AI was an AI that told you what to do, Astra is closer to an AI that, given only a goal, uses the computer directly and produces the result all the way to the end. In other words, an easy to use OpenAI model has begun to handle on its own part of the agent orchestration layer that had been the domain of less accessible tools such as OpenClaw. For users, the barrier to entry for AI agents has been lowered, meaning more work can be handed over. Expansion of AI workloads Astra naturally also comes with efficiency gains that lower the token cost per task versus the previous model. This is a trend across the AI industry as a whole, and if AI workloads were fixed, demand for AI data centers would have to plunge. The reason Jevons paradox continues to operate even after token price declines became a trend following the rise of Chinese models is that AI technological progress also expands the workload. What Astra's technological progress means is that where humans used to hand five minute, ten minute, and twenty minute tasks to AI, as AI performance improves and token prices get cheaper there is more to hand over, such as one hour and 24 hour tasks. Just as news flow about rising GPU rental prices has spread since Astra's arrival, it must be understood that falling AI token prices do not necessarily shrink or slow the AI hardware TAM. Rather, one should recognize that the emergence of a model like Astra can create another inflection point for the AI industry and structurally grow AI demand. Our understanding is that since early July, as the pace of GPU rental price increases slowed and token prices fell, a long IGV (software) / short SOX (semiconductors) pair trade has persisted in the US. This is because falling token prices were interpreted as positive for software, where tokens are a cost, but negative for infrastructure. If progress in models like Astra structurally spreads AI workloads and GPU rental prices begin to respond again, the perception that falling token prices are bad news for AI infrastructure companies could weaken (on 9/4 the DRAM ETF rebounded while IGV fell). As we have argued consistently, the issues accumulating in the AI industry since June (the proliferation of open models, this GPT-6 Astra release, and so on) are, in our interpretation, positive catalysts that generate new demand for AI infrastructure and hardware that did not exist before. We think that in a phase where rates are rising overall and liquidity is becoming scarce, these accumulated positives are not being reflected. The stock market in September is still uncomfortable with high rates, and within the Korean market there remain hurdles to get through, including digesting a round of earnings estimate cuts driven by the sharp won appreciation before the 3Q26 preview season. There is still discomfort standing in the way of the accumulated positives being reflected in a sustained trend. Overall, we continue to view the market conservatively. Even so, as emphasized in our September strategy, we believe one should not substantially empty out positions in core AI infrastructure stocks centered on memory. Positive catalysts not reflected in share prices are accumulating. While our baseline is conservative through mid October, one should keep the upside risk open that the trend, led by AI and semiconductor leaders, could turn at any time, even before October.