NVIDIA × RISC-V, TSMC AI surge, domestic AI chip breakthrough: full analysis of six core semiconductor events (July 2025)

1. NVIDIA’s major push: CUDA fully transplanted to RISC-V, opening up the third largest ecosystem (2025.7)

1. Core progress (RISC-V China Summit)

• Nvidia announcedCUDA full stack supports RISC-V, ending the x86/ARM duopoly, RISC-V becomesThe third largest CUDA host architecture in the world
• Technical layout: NVCC compiler adds RISC-V backend, runtime driver reconstruction, and comprehensive adaptation of 900+ CUDA libraries
• Historical origin: Cooperating with Shanghai Jiao Tong University on RISC-V since 2017, embedded MCU hasCumulative shipments exceed 1 billion units

2. Technical architecture (CPU+GPU+DPU heterogeneous)

• RISC-V CPU: Host control, running OS and drivers, scheduling GPU tasks
• NVIDIA GPU: Responsible for AI/HPC parallel computing (main computing power)
• high speed internet: NVLink-CXI implements RISC-V↔GPU100GB/sUltra-high-speed data exchange

3. Strategic significance (core benefits of the Chinese market)

• Break the computing power blockade: Domestic RISC-V servers (Xuantie C950, etc.) canDirectly compatible with CUDA ecosystem, the model migration cost approaches zero
• supply chain security: Avoid x86/ARM licensing and supply interruption risks, buildAutonomy + Opennessdouble base
• cost revolution：RISC-VZero licensing fee, reducing server BOM cost by 30%~50%

4. Timeline

• 2025.7: Technology disclosure, development board online
• 2025.11: CUDA 13.0 officially released with RISC-V support
• 2026.3: Domestic RISC-V + NVIDIA GPU server will be commercially available on a large scale

2. TSMC Q2: Revenue increased by 38.6% under the AI craze, and advanced processes accounted for 74% (2025.7)

1. Core Finance (NTD)

• revenue: 933.79 billion (+38.6% year-on-year, +11.3% month-on-month)
• net profit: 398.27 billion (+60.7% year-on-year, +10.2% month-on-month)
• Gross profit margin: 58.6%, operating profit margin 49.6%, net profit margin 42.7%

2. Process structure (led by advanced processes)

• 3nm:24% (NVIDIA Blackwell, Apple M4, AMD MI350)
• 5nm: 36% (H100/A100, Ascend 910B, mobile phone SoC)
• 7nm：14%
• 7nm and below total：74%

3. Business structure (the absolute core of AI)

• HPC/AI: Proportion60%, month-on-month +14% (the first growth engine)
• Smartphone: 27%, +7% month-on-month (moderate recovery)
• Automotive/IoT/Consumer Electronics:Total 13%

4. Customers and Prospects

• North American customers: 68% (NVIDIA, AMD, Apple, Qualcomm)
• Q3 Guidance: Revenue of US$31.8~33 billion, gross profit margin 55.5%~57.5%
• core driver: Orders for NVIDIA Blackwell, AI inference chips, and domestic high-power computing chips exploded

3. Nationally produced AI chip milestones: Gangwon D10 mass production, Pinyuan AI all-in-one machine released (2025.7)

1. Gangwon D10 (China’s first full-process localized AI chip with large computing power)

• Craftsmanship：12nm domestic(Design/manufacture/packaging all local)
• Specifications：
  • Single card:128GB LPDDR5（6400Mbps）、PCIe 5.0（128GB/s）、72W
  • Accuracy: FP32/TF32/FP16/BF16/INT8 fully supported
  • Performance: Single machine can run 16 cards671B parametersLarge model (DeepSeek)
• mass production: Q2 film production and Q2 mass production in 2025 (China’s first 12nm nationally produced large-scale computing power will be launched)

2. Pinyuan AI all-in-one machine (PYD10 series)

• Configuration: Up to 16 D10 accelerator cards, domestic 8-core CPU
• Performance:Large model inference efficiency300% improvement, a single machine is a cluster
• price: MIN 24,900, PRO 330,000, MAX 450,000
• Application:Government cloud, operators, intelligent manufacturing, financial AI reasoning

3. Strategic value

• Zero external dependencies: Completely get rid of overseas EDA/IP/manufacturing blockade
• Xinchuang just needs: Filling the gap in high-density AI computing power across the country
• cost advantage: Only similar NVIDIA solutions1/3~1/2

4. NAND flash memory Q3 price increase is final: low capacity will increase by more than 15%, high capacity will increase by 5%~10% (2025.7)

1. Price increase forecast

• 512Gb and below: Increase **>15%** (SLC/MLC/low capacity TLC)
• 1Tb and above:5%~10% (QLC enterprise-class SSD)
• eMMC/UFS:3%~8% (mobile phone/industrial control)

2. Core reasons (supply contraction + demand recovery)

• Original factory production reduction: Micron, Kioxia, SanDisk From the second half of 2024Cut production capacity significantly, bit supply growth rate < demand growth rate
• structural adjustment: Low capacity, thin profit, production capacityFull shift to high-volume QLC
• AI driven: Data center SSD, AI server storage requirementsbreak out
• Inventory low: Low-capacity NAND inventory in the distribution market has bottomed out, and price elasticity is extremely high

3. Impact

• Terminal price increase: U disk, memory card, IoT/industrial control modulePrice goes up
• Forced to upgrade:32GB→64GB→128GB as standard,BOM cost increases
• Original factory profit: NAND business gross profit margin rebounded to35%+, reversing losses

5. Xidiwei: Automotive power supply explodes, LDO is shipped in batches, and DC-DC enters Qualcomm cockpit (2025.7)

1. Vehicle specification product matrix

• DC-DC (main force): EnteredQualcomm smart cockpitReference design
• LDO regulator: towardsMany leading domestic car companiesBulk shipment
• Driver/PMIC: High/low side driver, automotive camera PMIC

2. Customer coverage (globalization)

• Tier1：Germany Joynext, South Korea YuraTech
• OEM：
  • Overseas: Audi, Hyundai, Kia
  • Domestic: Xiaopeng, Hongqi, Wenjie, Changan

3. Industry signals

• Domestic substitution accelerates: Automotive PMIC from TI/ADI/InfineonTurn local
• Automotive electronics boom：NXP/ST/ Xidi Micro Financial Report confirmsThe strongest track for automotive chips

6. Industry panoramic summary: Five major trends are clear (second half of 2025)

1. Three pillars of AI computing power

• NVIDIA + x86: High-end training (short-term mainstream)
• NVIDIA + RISC-V: China’s main compliance force (outbreak period)
• Nationally produced: Gangwon D10, Shengteng, Cambrian (required by Xinchuang)

2. Hengqiang, the leader in advanced manufacturing processes

• TSMC 3/5nmMonopoly AI chip, accounting for over 60% of revenue
• Domestic 12nm (Gangwon D10)Break the blockade, to fill the middle and low-end computing power

3. Storage enters a price increase cycle

• Low capacity NAND: Supply shrinks → prices skyrocket → terminal upgrades
• High capacity SSD: AI driven → Steady rise → Original factory profit restoration

4. RISC-V ecological qualitative change

• CUDA support→Mature server-level ecosystem→Domestic substitution accelerates
• RISC-V in 2026AI reasoning market share exceeds 15%

5. Automotive electronics continues to grow rapidly

• Automotive PMIC, power devices, MCUBoth volume and price rise
• Domestic chipsQuick cutglobal supply chain