NVIDIA H20 returns to China: The industrial chain disruption and symbiosis under the AI chip special supply game
On July 15, 2025, Nvidia H20 AI chip was approved by the United States to resume exports to China, ending a nearly three-month ban on sales.This kind of computing power is castrated and bandwidth is reserved.
China special editionThe return of chips will alleviate the domestic computing power shortage in the short term, but in the long term it will reflect the deep game of global semiconductor technology blockade - commercial counterattack - domestic breakthrough. It will also bring triple changes to the electronic components industry, including intensified competition, ecological differentiation, and supply chain restructuring.
1. H20 core positioning: Compliance-castrated version, accurately balancing regulation and business
1. Technical parameters (compared to H100 / Shengteng 910B)
H20 is based on Hopper architecture and TSMC 4N process. The core isReduce computing power and preserve video memory/internet, meets U.S. export controls (performance<20% of H100) while retaining large model training/inference practicality:
| parameters | NVIDIA H20 | Flagship H100 | Huawei Ascend 910B |
|---|
| FP16 computing power | 148 TFLOPS | 1979 TFLOPS(H20≈7.5%) | 256 TFLOPS (Super H20) |
| Video memory | 96GB HBM3 | 80GB HBM3 | 32GB HBM2e |
| Video memory bandwidth | 4.0 TB/s | 3.35 TB/s (H20 higher) | About 1.6 TB/s |
| Internet | NVLink 900GB/s | 900GB/s | Self-developed Internet (approximately 600GB/s) |
| Power consumption | 400W | 700W | 350W |
| Positioning | Inference + small and medium model training | Very large training | Training / Reasoning (Government / Industry) |
2. Design logic: compliance + usable double bottom line
- Computing power cut in half: FP16/TF32 H100 only7.5%, remove RT core
- Bandwidth overtake:96GB HBM3+4TB/s bandwidth,8 cards poolable 768GB of video memory, supports full training of 70B parameter model
- ecological integrity: 100% compatible with CUDA, migration cost approaches zero
2. Three major impacts: quenching thirst for computing power, explosion of industrial chain, acceleration of domestic production
1. Market side: AI computing power quenches thirst, and costs drop significantly
- short term relief gap: During the lockdown period, domestic AI companies stocked up on parallel imports and switched to domestic production, after H20 was approvedIncreased volume through official channels, 8-card server rental prices fell,Single Token cost reduced by 30%+
- Requirements stratification:
- ✅Suitable for:Reasoning, large vertical models, finance/medical/industrial quality inspection(H20 has the best price/performance ratio)
- ❌ Restricted:Very large-scale pre-training(The computing power is only 1/6 of H100, and the cluster efficiency is insufficient)36 krypton
- capital response: On the day H20 was approved, the AI computing power/optical module index surged3%~6%, Zhongji InnoLight, Xinyisheng, etc. closed their daily limit
2. Industrial chain: upstream components benefit comprehensively (core benefits for the component industry)
H20 server isHigh computing power, high power consumption, high integrationThe carrier directly drives the demand for upstream components to explode:
- Optical module: The demand for 800G/1.6T high-speed optical modules (NVLink supporting) has surged.Zhongji InnoLight, Xinyisheng, Tianfu CommunicationsOrder doubled
- Advanced packaging:CoWoS, 2.5D/3D package (HBM+GPU),Changdian Technology, Tongfu MicroelectronicsUndertake H20 packaging and testing orders
- storage:HBM3/High Bandwidth DDR,Montage Technology, Changxin StorageSupporting domestic solutions
- Cooling/Power: Liquid cooling (H20 power consumption 400W, 3~5 times traditional), high-power power supply,Invic, Subedbenefit
- PCB/Substrate: 16-layer or above high-frequency high-speed PCB, IC carrier board,Shennan Circuit, Shanghai Electric Co., Ltd.Both volume and price rise
3. Domestic chips: squeeze and pushback coexist, substitution accelerates
- short term shock: CUDA has strong ecological stickiness,Tencent / Alibaba / BytePriority purchase H20,Cambrian, sea light informationShort-term share pressure
- Long term pressure:
- Performance catch-up: Ascend 910B computing powerUltra H20, 910C approaches H100
- Ecological breakdown:Huawei CANN Compatible80% CUDA API, 85% operators automatically migrated(Performance loss 15%~20%)
- Policy support: The third phase of the big fund is tilted towards GPU/EDA/advanced packaging.Government Affairs/State-owned Assets CloudDomestic share reaches67%
- Pattern changes: Nvidia’s share in China increased from95%→50%, Domestic (Shengteng/Cambrian/Kunlun core) quickly filled
3. Deep game: U.S. technological hegemony vs. China’s supply chain independence
1. U.S. policy: let go of small things and block big ones, and use commerce to stop big ones.
- Reasons for lifting the ban:
- Business losses: Nvidia banned for 3 monthsProvision for impairment of US$5.5 billion, Loss of Chinese orders
- The rise of domestic products: Shengteng 910B hasTied H20, the blockade fails
- Strategy adjustment: useLow performance special chipMaintain the market and delay domestic substitution
- future risks: The United States intends toInterconnect bandwidthUnder control, H20/B30 may be further restricted in the future
2. China’s response: two-pronged approach, independent and controllable
- Supply chain resilience: Built by leading cloud vendors **NVIDIA H20 + Huawei Ascend + self-researched three lines in parallel** to avoid the risk of supply interruption
- technological breakthrough:
- Chiplet/integrated storage and calculation: Overtaking on curves, reducing dependence on advanced processes
- EDA/IP/Materials: Large funds focus on investing to make up for basic shortcomings.
- Yangtze River Delta Cluster: AI chip production capacity accounts for 10% of the country’s total58%, full chain collaboration
4. Enlightenment from the component industry: Break the special supply trap and move towards symbiosis and win-win
1. Be wary of special chip traps (three major risks)
- performance ceiling: H20 only H100 1/6, long-term dependence willWidening the technological generation gap
- Ecological lock-in: CUDA has deep barriers,High migration costs(The migration of a financial company to Shengteng required rewriting 30% of the code, which took 2 to 3 months)
- Repeated policies: U.S. controls will be tightened at any time,Unstable supply
2. Path to industry disruption (core strategy of component companies)
- High-end card slot: focusHBM, high-speed optical modules, advanced packaging, liquid coolingWaiting for H20’s urgent needs, and simultaneously adapting to domestic chips
- Domestic adaptation: Provided for Ascension/CambrianCustomized components(Wide temperature, high reliability, low power consumption), share replacement dividends
- Technology crossover: willConsumer electronics technology vehicle planning / AI(such as Type-C enhancement, UWB anti-interference), broadening application boundaries
- Ecological co-construction: ParticipateDomestic framework compatibility certification(Adapted by MindSpore and PyTorch) to reduce customer migration costs
3. Future pattern: hierarchical competition, dual-line symbiosis
- High-end training: Domestic (Ascend 910C / Cambrian S70) + a small amount of compliant overseas chips
- Reasoning/Edge: H20 (ecological) + domestic (cost-effective/safety) dual mainstream
- Components:Global supply chain + domestic substitutionParallel, head enterprise walks on two legs
5. Summary: Balancing openness and autonomy, crossing the cycle
NVIDIA H20 returns to China and is the global AI chipControl - Countermeasure - SymbiosisThe epitome of: In the short term, it will alleviate the computing power shortage and promote the explosion of the component industry chain; in the long term, it will force China toChip design, ecological construction, and supply chain independenceAccelerate the breakthrough.
For the electronic components industry, the real breakthrough is not that the United States or China willMaintain autonomy and controllability in open cooperation and build differentiated advantages in technological catch-up——It can not only cut into NVIDIA's global supply chain to make quick money, but also take root in the domestic ecology to win the long-term future. This is the core logic of transcending cycles and continuing to grow.