TL;DR
- After GPUs, a new bottleneck appears: memory. AI is absorbing a growing share of global DRAM production.
- Apple seeks to secure its supplies with the Chinese manufacturer CXMT, not to reduce its costs, but to guarantee its production volumes.
- DRAM prices expected to rise sharply : +40 to +50% in the third quarter of 2026, then another +30 to +40% in the fourth quarter.
- The risk becomes geopolitical. A decision by Washington targeting CXMT could directly affect iPhone production in the next 18 months.
- AI is reshaping the memory market. By 2027, 15 to 20% of capacity intended for consumer electronics could be redirected to data centers.
- Each generation of AI servers consumes more memorywith more HBM for GPUs and more DDR5 for host processors.
- Shortages move up the value chain. After GPUs, advanced packaging, electricity and data centers, memory is becoming the new limiting factor.
- The next critical resources could be optical interconnections, electrical equipment and cooling systems.
- This change is already benefiting several French manufacturersnotably Schneider Electric, Legrand, Soitec and STMicroelectronics.
For two years, the technology industry lived to the rhythm of the same obsession: GPUs. NVIDIA’s accelerators have become the symbol of the rush to artificial intelligence, bringing in its wake a shortage of advanced packaging, an explosion of investment in data centers and a global race for electrical capacity. A new critical link now appears, and this time, it is no longer about computing power, but about memory.
Apple is taking steps to preserve the possibility of obtaining supplies from Chinese DRAM memory manufacturer CXMT, even as Washington plans to strengthen restrictions targeting certain Chinese technology companies.
Market projections predict lasting tension. Jefferies anticipates a DRAM price increase of between 40 and 50% in the third quarter of 2026, followed by a further increase of 30 to 40% in the fourth quarter. The year 2027 would remain on an upward trend before a possible rebalancing from 2028, provided that the new industrial capacities actually come into production. Behind this inflation lies a more structural development, which sees a growing share of the memory produced in the world now absorbed by artificial intelligence infrastructures.
The market is thus entering a new phase, after GPUs, electricity and data centers, memory is emerging as the next limiting factor in the development of AI. For electronics manufacturers, the question is no longer just about negotiating the best prices, but about ensuring that components will actually be available.
Apple reveals a paradigm shift: in AI, availability is now worth more than price
Apple’s attitude is probably the best indicator of this market change. The Cupertino company has distinguished itself by extremely sophisticated management of its supply chain. Diversification of suppliers, permanent negotiation of costs, upstream investments in certain strategic partners, this industrial discipline has become one of the company’s main competitive advantages.
Unlike the project carried out in 2022 with YMTC for NAND memory, essentially motivated by a search for price competitiveness and diversification, the merger with CXMT does not pursue an objective of cost reduction, but seeks to purchase an additional option on a resource which risks becoming rare.
Suffice to say that the CXMT file is critical for Apple, the decisions taken in Washington concerning the American Entity List can now directly influence the iPhone production volumes available in the next eighteen months.
Artificial intelligence is gradually sucking up the world’s memory
The current tension is fueled by a much deeper transformation: the rise of artificial intelligence infrastructures.
15 to 20% of memory capacities currently intended for consumer electronics could be redirected to data centers by 2027.
Each generation of AI servers simultaneously increases the number of installed GPUs, the amount of HBM memory directly integrated into the accelerators and the volumes of DDR5 required by the host processors. As models become larger and autonomous agents perform more simultaneous processing, memory grows at the same rate as computing power.
This dynamic reproduces the scenario already observed with GPUs. In 2023, scarcity was in accelerators, in 2024, it shifted to advanced packaging capabilities, in 2025, power infrastructure and data centers became the new limiting factors, and in 2026, memory joins this list.
The AI economy is thus entering a phase where bottlenecks are gradually moving towards less visible, but essential, components. After GPUs, memory, optical interconnects, electrical equipment or cooling could become the next critical resources. A global demand from which French manufacturers such as Schneider Electric, Legrand, Soitec Or STMicroelectronicspresent on several of the essential building blocks of the AI economy.
