Meritz Securities: Middle East Sovereign AI Capital Inflows Drive Up Memory Prices; Kimi K3 Is a Hardware Tailwind
Taylor Wilson
Middle East sovereign funds are negotiating directly with Samsung and SK Hynix for multi-year memory supply; spot prices have surged ~146% above contract. Meritz calls Kimi K3 a mid-term hardware positive and recommends buying semis over Big Tech.
Why are Middle East sovereign funds suddenly buying memory chips?
The driver is national security and data sovereignty — countries like Saudi Arabia want domestic data centers running AI in their own languages, on their own terms.
This means → these are not one-off purchases but multi-year supply negotiations that lock up capacity at Samsung and SK Hynix.
In plain terms = the queue for memory used to be American hyperscalers only. Now nation-state buyers are pushing in.
How can one buyer move prices this much?
AI data-center procurement is extremely concentrated. A single new buyer can shift the supply-demand gap by 3–5 percentage points.
64 GB DDR5 server DRAM spot prices have surged from a late-June contract price of roughly $1,380 to $3,100–$3,400 — a jump of about 146%.
This reflects a market where the real problem is not price but availability. As Meritz analysts put it: "There simply isn't enough — everyone is scrambling to secure supply."
Is the shortage limited to data centers?
No. Memory tightness has spilled into consumer electronics: Apple and Chinese smartphone makers cannot secure enough chips ahead of the second-half peak season, leaving Q4 production plans with gaps.
Meritz expects Q3 2026 server DRAM contract prices to rise beyond the ~15% quarter-on-quarter the market had priced in.
This means → suppliers who offered softer pricing in Q2 will hike the most aggressively in H2 — the margin they gave up earlier will be clawed back.
What makes Kimi K3 fundamentally different from DeepSeek?
DeepSeek's core narrative was training AI for just $6 million — which spooked the market into fearing lower hardware demand.
Kimi K3 is the opposite: no training cost was disclosed, and the company stated explicitly that running K3 requires a large cluster of at least 64 high-end chips.
On usage cost, K3 runs at $0.95 per task — on par with GPT-5.6 Sol ($1.04), and an order of magnitude pricier than DeepSeek V4 Pro ($0.04).
In plain terms = DeepSeek made people think "AI can run on very little hardware." K3 proves that frontier models are still hardware-hungry.
Why is the "K3 kills hardware demand" thesis wrong?
Meritz says the thesis rests on two misjudgments.
Misjudgment one: AI infrastructure spending is not just Big Tech. Nvidia's data-center chip sales show hyperscaler share declining while smaller AI cloud providers and enterprises keep growing.
Misjudgment two: Chinese AI models' inference workloads actually run heavily on American hardware — providers like Fireworks AI and Together AI serve enterprise clients on U.S. chips.
This means → open-source or closed, Chinese or American — the underlying hardware companies benefit either way.
What comes next?
Meritz recommends buying semiconductor and memory names, while trimming Google, Microsoft, and other Big Tech positions.
The key test: whether Middle East sovereign inflows and the Kimi K3 effect converge during Q3–Q4 contract negotiations.
This reflects a broader market pivot — from "which AI model is best" to a more basic question: is there enough hardware to go around?
Content is for reference only, not financial advice.