Meritz Securities: Middle Eastern Sovereign AI Capital Enters Korean Memory Procurement; Kimi K3 Is a Tailwind, Not a Shock

Middle Eastern sovereign AI capital has officially entered the memory procurement market, reshaping the supply-demand landscape and driving prices to accelerate upward. Meanwhile, the recently much-discussed release of the Kimi K3 model in the AI circle is also interpreted by Korean brokerages as a medium-term tailwind for AI hardware, rather than the demand shock feared by the market.

According to a research report released by Meritz Securities on July 19, Middle Eastern sovereign AI investors, such as those from Saudi Arabia, have recently begun negotiations with Korean memory manufacturers like Samsung Electronics and SK Hynix regarding medium-to-long-term procurement plans, inquiring about supply capacity and volumes for the coming years. The report expects that the quarter-on-quarter increase in server DRAM contract prices in the third quarter will exceed the market's previous expectation of approximately 15%, with suppliers who priced more loosely in the second quarter seeing particularly prominent price hikes in the second half of the year.

Meanwhile, the spot market has already shown clear signs of price increases. The spot price for 64GB DDR5 server DRAM has surged sharply since mid-July, with recent quotes ranging from $3,100 to $3,400, approximately 146% higher than the contract price of around $1,380 at the end of June. High-end products with bus speeds reaching 6400Mbps have seen even more significant gains.

Amidst the dramatic changes in the supply-demand structure of the memory market, Meritz Securities' investment strategy team, in another report released on the same day, characterized the release of Moonshot AI's Kimi K3 model as a medium-term positive for AI hardware demand. It advised investors to buy shares of AI component companies such as semiconductors and memory makers, while reducing holdings in large-cap tech stocks like Google and Microsoft. Meritz believes that the market's judgment that "Kimi K3 will reduce AI hardware demand" is based on two misjudgments, and that K3 will instead bring incremental demand to hardware manufacturers from a medium-term perspective.

Middle Eastern Sovereign AI Capital Enters the Market; A Single Buyer Can Tip the Supply-Demand Balance

Analysts at Meritz Securities pointed out in the report that the core logic behind Middle Eastern sovereign AI investment lies in national security and data sovereignty—countries wish to build data centers within their own territories and operate AI systems autonomously using their own language and data. As this investment plan becomes more concrete, supply tightness expected in 2027 is appearing ahead of schedule.

The structural changes in the current memory market mean that the entry of a single large buyer has a strong impact on prices. The entities building AI data centers are highly concentrated; the entry of a single purchaser can instantly change the market demand satisfaction rate by 3 to 5 percentage points. The demand satisfaction rate is the ratio of actual supply to market demand; slight fluctuations in this indicator can trigger substantial price jumps. Currently, large customers account for over 70% of memory manufacturers' revenue, with pricing power highly concentrated among a few buyers.

This is vastly different from the dispersed procurement structure of the smartphone and PC eras. Back then, numerous manufacturers purchased in small batches, and order changes from a single customer had limited impact on the overall market; today, the centralized procurement model of AI data centers means that every new sovereign-level buyer entering the market can become a decisive variable in pricing.

Supply Shortages Have Spread to End Products; Volume Allocation, Not Price, Is the Core Conflict

Meritz pointed out that the impact of memory shortages has spread from the data center side to consumer electronics endpoints. Apple and Chinese smartphone manufacturers have faced gaps in their fourth-quarter production plans because they were unable to procure sufficient memory before the peak season in the second half of the year.

He emphasized that the essential problem facing the current market is "volume allocation" rather than just "price"—"In short, there isn't enough stock, and all parties are scrambling to buy." Addressing the cyclical concerns circulating in the market, he believed this judgment was somewhat off, as the current tension is rooted in structural supply insufficiency rather than short-term demand fluctuations.

The report also noted that some suppliers adopted relatively flexible low-price strategies in the second quarter to meet customer needs, which means these manufacturers have greater room for contract price increases in the third and fourth quarters, with more pronounced pricing elasticity.

Kimi K3 Release Interpreted as a Tailwind for AI Hardware Demand, Not a Shock

Analysts from Meritz Securities' investment strategy team released a report on the same day, characterizing the release of Moonshot AI's Kimi K3 model as a medium-term positive for AI hardware demand. They advised investors to buy shares of AI component companies such as semiconductors and memory makers, while reducing holdings in large-cap tech stocks like Google and Microsoft.

Meritz clearly distinguished this event from last year's DeepSeek shock. The core narrative of DeepSeek was completing AI training at an extremely low cost of $6 million, whereas Kimi K3 did not disclose training costs or GPU usage, and officials explicitly stated that running K3 requires a large cluster composed of at least 64 high-performance chips, a model that cannot be solved by low-cost hardware. In terms of usage costs, according to data from the AI performance evaluation firm Artificial Analysis, the cost per task for K3 is $0.95, which is in the same order of magnitude as GPT-5.6 Sol ($1.04) and Claude Fable 5 ($2.75), but differs by an order of magnitude from DeepSeek V4 Pro ($0.04).

Further analysis pointed out that the logic that "K3 will reduce AI hardware demand" is based on two misjudgments: First, AI infrastructure investment is not solely led by large tech companies—taking Nvidia's data center chip sales data as an example, the share of large cloud providers is declining, while the share of small and medium-sized AI cloud service providers and ordinary enterprises is continuing to rise; Second, inference services for Chinese AI models actually run largely on US hardware, with inference service providers like Fireworks AI and Together AI offering services to enterprise clients based on US chips.

He concluded that, "Whether closed-source or open-source, hardware companies that all AI business participants ultimately need to use will gain medium-term incremental demand from the release of K3."