๐The Next Web (TNW)โขStalecollected in 16m
China drafts $295bn plan to build AI data centers
๐กChina's massive move to decouple from Nvidia could reshape global AI supply chains and hardware standards.
โก 30-Second TL;DR
What Changed
2 trillion yuan ($295bn) investment over five years
Why It Matters
This policy shift could force AI developers in China to pivot toward domestic silicon alternatives, potentially creating a fragmented global AI hardware ecosystem.
What To Do Next
Monitor the performance benchmarks of domestic Chinese AI accelerators like Huawei Ascend to assess potential compatibility for cross-border deployments.
Who should care:Enterprise & Security Teams
Key Points
- โข2 trillion yuan ($295bn) investment over five years
- โขStrategic push for domestic AI hardware independence
- โขExplicit goal to phase out Nvidia chips in data centers
๐ง Deep Insight
Web-grounded analysis with 20 cited sources.
๐ Enhanced Key Takeaways
- โขThe $295 billion investment plan is part of China's broader 'Six Networks' national strategy, which encompasses essential infrastructure development in areas like water, power, and computing.
- โขThe initiative aims for at least 80% of AI chips and related technology equipment to be sourced from domestic suppliers, further reducing the market share of foreign chip giants like Nvidia and AMD in China.
- โขThe blueprint includes connecting China's scattered data infrastructure into a unified national computing network by 2028, with the goal of deploying AI across public sectors such as healthcare, transportation, and urban management.
- โขIf power grid integration is factored into the plan, the total projected investment could reach at least 5 trillion yuan (approximately $738 billion).
- โขNine domestically designed AI processors, including those from Huawei and Alibaba, were recently certified for state procurement under China's Anke security certification framework, creating a formal buying channel for domestic AI accelerators across government agencies and state-linked entities.
๐ Competitor Analysisโธ Show
| Feature / Chip | Huawei Ascend 910B/910C | Baidu Kunlun1-T | Alibaba Hanguang 800 | Nvidia A100 (for context) | Nvidia H100 (for context) |
|---|---|---|---|---|---|
| Process Node | SMIC 7nm (for 910B) | 14nm (Samsung) | 12nm (TSMC) | 7nm (TSMC) | 4nm (TSMC) |
| Peak Performance (FP16/INT8) | 400 TFLOPS (FP16 for 910B), ~800 TFLOPS (FP16 for 910C) | 260 TOPS (INT8) / 230-281 TOPS (INT8) | 78,563 IPS (inference) | 312 TFLOPS (TF32), 624 TFLOPS (FP16), 1248 TOPS (INT8) | 1979 TFLOPS (FP16), 3958 TOPS (INT8) |
| Memory Bandwidth | 1.2 TB/s (910B), 3.2 TB/s (910C) | 512 GB/s | (Not explicitly stated in GB/s, uses 16-phase power supply) | 1.5 TB/s | 3.35 TB/s |
| Memory Capacity | 32GB (910B), 128GB (910C) HBM | 16GB HBM2 | (Not explicitly stated) | 40GB/80GB HBM2e | 80GB HBM3 |
| Primary Use Case | Training & Inference | Cloud & Edge (Training & Inference) | Inference | Training & Inference | Training & Inference |
| Relative Performance Claim | 910B comparable to Nvidia A100 (2020). 910C comparable to H100 (2022). CloudMatrix 384 system outperforms Nvidia GB200 NVL72 in BF16 compute. | Kunlun1-T showed 1.5-3x better performance for search engine models vs. Nvidia T4. | Hanguang 800 15x more powerful than Nvidia T4 GPU, 46x more powerful than Nvidia P4 GPU for inference. | (Baseline) | (Baseline) |
๐ ๏ธ Technical Deep Dive
- Huawei Ascend Series (910B, 910C): The Ascend 910B, a second-generation chip, has a maximum theoretical performance of 400 TFLOPS (FP16) and 1.2 TB/s memory bandwidth with 32GB HBM. The Ascend 910C is a dual-die package, combining two 910B-class processors, achieving approximately 800 TFLOPS (FP16) and 3.2 TB/s memory bandwidth with 128GB HBM. These chips are integrated into advanced data center architectures like CloudMatrix 384, which clusters 384 Ascend 910C chips in an optical mesh network, delivering significant BF16 compute power and memory bandwidth at a system level.
- Baidu Kunlun Series (e.g., Kunlun1-T): China's first domestically produced AI accelerator IC, the Kunlun1-T, features thousands of cores and provides up to 260 trillion operations per second (TOPS) (INT8) and 512 GB/s memory bandwidth to two HBM2 packages totaling 16GB of in-package memory. It is manufactured using Samsung's 14nm process technology and utilizes Samsung's I-CubeS 2.5D packaging technology. The chip is designed for both cloud and edge computing, supporting various AI applications including search ranking and deep learning frameworks like PaddlePaddle.
- Alibaba Hanguang 800: This AI inference chip, unveiled in 2019, is produced using TSMC's 12nm process. It focuses on inference tasks, with a peak performance of 78,563 images per second (IPS) in the Resnet-50 industry test and an energy efficiency ratio of 500 IPS/W. The Hanguang 800 contains 17 billion transistors and is designed with an in-house developed hardware framework and optimized algorithms for convolutional neural networks (CNN) and computer vision. It has been applied in Alibaba's core business units for tasks like product search, personalized recommendations, and smart city applications.
๐ฎ Future ImplicationsAI analysis grounded in cited sources
China will significantly reduce its reliance on foreign AI chips in state-backed data centers within the next five years.
The explicit $295 billion investment plan, coupled with procurement mandates for at least 80% domestic hardware, creates a protected and incentivized market for homegrown AI chip manufacturers.
Chinese AI chip manufacturers will experience accelerated growth and increased market share within China's domestic market.
Government support, the 'secure and reliable' procurement list, and the large-scale national data center buildout guarantee substantial demand and a stable market for domestic AI chip producers.
The global AI chip market will become increasingly bifurcated, with distinct hardware and software ecosystems emerging in China and the West.
Ongoing US export controls on advanced semiconductors and China's strategic push for technological self-sufficiency are driving the development of separate, self-reliant AI infrastructure and supply chains.
โณ Timeline
2018-07
Baidu releases Kunlun AI chip, China's first cloud-to-edge AI chip.
2019-09
Alibaba unveils Hanguang 800, its first AI inference chip.
2022-10
US implements sweeping export controls on advanced computing chips and semiconductor manufacturing equipment to China.
2023-10
US tightens export curbs, specifically restricting Nvidia's A800 and H800 chips to China.
2026-05
Nine domestically designed AI processors, including Huawei's and Alibaba's, are certified for state procurement in China.
2026-06
China drafts $295 billion plan to build AI data centers over five years, targeting 80% domestic hardware reliance.
๐ Sources (20)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
๐ฐ
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: The Next Web (TNW) โ