[Core News] Apple/MediaTek postpone 2nm, Nvidia RTX 5090 released

🔥 A quick overview of cutting-edge news

• Flagship chip collective upgrade: Due to cost and yield issues, Apple and MediaTek have postponed the adoption of TSMC’s 2nm process and switched to the more reliable 3nm or basic version of 2nm; Samsung has accelerated 2nm trial production to grab diversified orders from the supply chain.
• AI computing power breaks through again: NVIDIA released RTX 5090 at CES 2025, with Blackwell architecture, 32GB GDDR7 video memory, 3352 TOPS AI computing power, priced at US$1,999, reshaping the creation and gaming experience.
• Storage race upgrade: Samsung HBM4 has entered 4nm trial production and plans to mass-produce it by the end of 2025. It will use the 1c DRAM process to achieve both excellent performance and yield.
• Industry trend differentiation: Demand for AI servers is strong, but the automotive market is recovering slowly; local companies are accelerating the deployment of SiC and silicon wafers to seize emerging tracks.

1. Process game: Apple/MediaTek delay 2nm, Samsung accelerates to grab the position

• core reasons: TSMC’s 2-nanometer wafers are quoted at about US$30,000 per piece. The first batch yield is only about 60%, and the stacking performance improvement is limited (about 5%), resulting in a 20%-30% increase in single-chip costs.
• corporate strategy
  • apple: The iPhone 17 series switches to 3nm, and the A20 chip (iPhone 18 series) uses the basic version of 2nm (N2) instead of the high-performance version (N2P), giving priority to ensuring a balance between production capacity and cost.
  • MediaTek: Dimensity 9500 has postponed its 2nm plan and switched to 3nm or mature processes; Qualcomm is also considering shifting some orders to Samsung's 2nm to promote supply chain diversification.
  • Samsung: It started 2nm trial production in Q1 and has won an order from Japanese AI startup Preferred Networks. It is jointly testing with Nvidia and Qualcomm to compete for TSMC customers.
• Industry impact: Leading manufacturers are becoming more rational about advanced processes, with yield and cost becoming key decision-making factors, accelerating the de-unipolarization of the supply chain.

2. NVIDIA RTX 5090 released: Blackwell architecture reshapes AI + graphics ecosystem

• Core specifications
  • Architecture: Blackwell (GB202 core), 92 billion transistors
  • Performance: 21,760 CUDA cores (the first GeForce GPU to break through 20,000 cores), 32GB GDDR7/512bit video memory, bandwidth 1792GB/s
  • AI computing power: 3352 TOPS (FP4), supports DLSS 4, Reflex 2, significantly improved ray tracing performance
  • Power consumption: 575W (TDP), available January 30, priced at $1,999
• Application scenarios: Empowering 4K/8K games, film and television rendering, AI model training and inference, providing a top platform for creators and enthusiastsNVIDIA。
• market significance: Further consolidate the dominant position in the AI computing ecosystem and promote the upgrade of GPU to an all-round computing core.

3. Samsung HBM4 rushes into mass production: 4nm + 1c DRAM targets AI servers

• Technical route: The logic chip uses a 4-nanometer process, the DRAM layer uses a 10-nanometer 1c DRAM process, and hybrid bonding technology improves connection efficiency and yield.
• schedule planning: 4nm trial production will be completed in Q1 of 2025, mass production readiness certification (PRA) will be completed by the end of the year, and supplies will be supplied to customers such as Nvidia in early 2026.
• competitive advantage: The only HBM supplier that has achieved self-research on the entire process of DRAM, logic manufacturing and 3D packaging, aiming to seize Nvidia’s GPU supply chain share.

4. Supply chain and industry trends

5. Summary and outlook

• Process rationalization: The advanced process has shifted from parameter-only to cost-performance-yield balance. The popularization of 2nm has been delayed. 3nm and mature processes will still be the main force in the short term.
• AI computing power competition: GPU and HBM4 are upgraded simultaneously, AI server becomes the core growth engine, and NVIDIA, Samsung and other acceleration cards are available.
• Domestic opportunities: SiC, power management chips, silicon wafers and other subdivisions are ushering in the introduction window, and the local supply chain is expected to break through the barriers to automotive regulations and AI applications.