Debate Swirls on Wall Street Over AI 'Speed Control'

Debate is ongoing in the stock market regarding the "AI speed control" argument raised by the heads of global artificial intelligence (AI) companies such as Anthropic, OpenAI, SpaceX, and Google DeepMind.

Kim Il-hyuk, a researcher at KB Securities, diagnosed the current situation in a report published today (September 15), stating, "As the competitive strategies of AI model developers diverge, conflicting interests in the AI market have come to the surface."

In recent months, as competition to develop AI models has heated up, the release cycles of new models by developers have shortened significantly compared to before.

The problem is that this has made it difficult to secure sufficient time to recover development costs, and Kim pointed out that "they likely concluded that striving to widen the performance gap among AI models is no longer a sound strategy."

Therefore, instead of spending vast amounts of money to create slightly better-performing models, they likely reached the conclusion that it is better to focus on designing a "harness" that maximizes the performance of existing models, enhancing the efficacy of AI transformation (AX) and expanding token demand.

However, as seen in NVIDIA CEO Jensen Huang's opposition to slowing down AI model development, this issue also demonstrates that the interests of AI model developers and NVIDIA do not align, Kim said.

He evaluated, "As the burden of building AI data centers increases, the influx of cheap, sufficiently performant competing models has weakened profit growth expectations for AI model developers. Furthermore, with NVIDIA opposing the AI developers' calls to slow down the development of frontier AI models, it has become clear that the interests of the two camps are no longer aligned."

Regarding the specific background behind Anthropic CEO Dario Amodei's proposal to slow down AI development, Hwang Soo-wook, a researcher at Meritz Securities, said, "It is unlikely that righteousness such as humanity's safety is the motivation behind the speed control proposal."

He noted that some already view this as a ladder-pulling strategy ahead of initial public offerings (IPOs) this year to cement the market monopoly structures of companies like Anthropic and OpenAI, adding, "If AI development speed is slowed down, the biggest beneficiaries will be the AI frontiers that spend massive amounts on AI development."

OpenAI is similarly preparing for an IPO.

This means there is considerable incentive to save on new model development costs and improve operating profit margins.

In fact, OpenAI's "GPT-6 Astra," estimated to have been developed between April and August this year, was trained using more than 100,000 Blackwell-class graphics processing units (GPUs). Based solely on GPU rental rates, it is estimated to have incurred opportunity costs amounting to USD 1.21 billion (KRW 1.6 trillion).

Nevertheless, Hwang advised, "Looking at the virtuous cycle of AI infrastructure investment driven by AI monetization, there is no need to interpret this issue pessimistically regarding AI infrastructure stocks."

He stated, "This is because it is a matter of securing the business sustainability of U.S. AI frontiers, which are the center of the industry," adding, "U.S. model usage is already rising again, centered around OpenAI's new Astra model and Luna lightweight model. For AI infrastructure investors, they can welcome the counterattack of the U.S. AI frontiers."

On the other hand, some argue that memory demand momentum has passed its peak.

Lee Min-hee, a researcher at BNK Securities, said in a report published yesterday, "As memory costs approach half of the bill of materials (BOM) for PCs and smartphones, mid-to-low-end OEMs began reducing memory capacity starting in the second quarter, while high-end PC and smartphone OEMs are moving to raise selling prices in the second half, demonstrating that memory costs have reached their limits."

He added, "Even in AI server systems, memory capacity plans for new models are being scaled back. Whatever the motivation, it shows that demand elasticity has begun to slow down," and forecasted that Samsung Electronics and SK Hynix's operating profits in the second half of this year will fall short of market expectations.
※ Please note: This article was translated by AI and may contain errors.