OpenAI unveiled its next-generation artificial intelligence (AI), "GPT-6 Astra," drawing market attention. Introducing the AI, OpenAI said it marks "the beginning of the AGI era." Semiconductor stocks, which had stalled under a wave of AI skepticism, immediately rebounded. U.S. big tech names such as Nvidia, as well as Korea's flagship chipmakers Samsung Electronics and SK hynix, are all stirring at once. In the securities market, analysts say that as the performance race among global AI developers reignites, shares tied to AI infrastructure are likely to continue their rebound.
On the 7th, Samsung Electronics rose in the 5% range and SK hynix in the 8% range. With news of Astra, unveiled on the 3rd, share prices jumped sharply for the first time in a while. However, as profit-taking emerged after the short-term surge, only SK hynix maintained a slight gain at today's (the 8th) close, while Samsung Electronics turned lower, ending the session mixed.
It is characterized by directly controlling the computer in place of the user and carrying out complex, multi-step tasks on its own. OpenAI, the developer of ChatGPT, says its next-generation AI "Astra" goes beyond simple question-and-answer. The company said it consistently handles the entire process, from search and usage analysis to drafting documents and presentation materials, building and running programs, and then verifying errors.
Lee Kyu-won, an analyst at DS Investment & Securities, said, "Astra is meaningful in that it expands the scope of agentic AI from coding to a broad range of white-collar work," adding, "Astra's performance gains appear in the expansion of agent capabilities that go beyond coding to directly use multiple software programs to carry out a single task to the end."
Astra's biggest feature is that it introduced a new scaling law called "recurrent depth" (the phenomenon in which efficiency improves as size grows) and conducted the largest-ever pretraining using more than 100,000 GPUs.
For example, with existing AI, when a user posed a question, it produced an answer through inference after passing through multiple transformer blocks. Astra, however, improved performance by repeatedly performing the same act of passing through these transformer blocks multiple times via the method called recurrent depth.
With this approach, performance can be improved without increasing the size of parameters, which comprise various elements including these transformer blocks. Put simply, parameters can be called the AI's brain, and existing AI developers had focused on enlarging this brain to boost model performance.
The space where these parameters are stored is high-bandwidth memory (HBM). As efforts had focused on enlarging parameter size, more HBM was needed. Some may argue this could negatively impact HBM demand, but the securities industry sees the likelihood of HBM bottlenecks disappearing as very low.
The analyst said, "Right now the competition over model performance among AI developers is extremely intense," adding, "Even if a new method like recurrent depth emerges, they will aim to build the highest-performing AI models with the largest parameters, so they will try to load as much HBM as possible."
There is also growing expectation that the advent of a higher-performing new AI will boost demand for memory semiconductors. Jensen Huang, Nvidia's chief executive officer (CEO), said on X (formerly Twitter), "AGI has finally arrived," adding that more than 100,000 of the company's graphics processing units (GPUs) were used to train Astra.
He added, "Four hundred thousand will be running for the next model." As HBM is essential to maximize GPU computing performance, the comment suggested that demand for memory semiconductors will also surge.
Ultimately, as it has been proven that large-scale AI chip infrastructure is essential to realize Astra-level AI, analysts say stalled AI infrastructure investment will resume. With OpenAI's new model release reigniting the performance race among big tech, the expansion of infrastructure and rising service usage are expected to together keep stimulating semiconductor demand.
Choi Bo-young, an analyst at Kyobo Securities, said, "What has restored semiconductor investment sentiment is not only the win-or-lose outcome of a particular model, but the possibility that such competition can create new tasks and usage rather than merely replacing existing services," adding, "If usage growth outpaces gains from compute-efficiency improvements, demand for GPUs, servers, and memory to handle it could also expand."
Kim Dong-Won, head of research at KB Securities, said, "With the U.S. and China competing to roll out new AI models and expand data center construction, Samsung Electronics and SK hynix are expected to benefit the most," adding, "Next year's memory market is expected to see an acceleration of concurrent growth in general-purpose memory alongside a surge in HBM4 demand."