TLDRs;
- DeepSeek confirms China will soon unveil domestically made AI chips optimized for its new V3.1 AI model.
- The firm’s adoption of UE8M0 FP8 scaling could cut memory use by 75%, boosting efficiency on local processors.
- Huawei and other firms are racing to adapt chips for Chinese AI models amid tightening U.S. export controls.
- Despite progress, Nvidia chips remain dominant, though China’s $340B semiconductor push could accelerate domestic innovation.
China is preparing to unveil domestically developed artificial intelligence (AI) chips in a move that could reshape its fast-evolving tech landscape, according to a new announcement from AI startup DeepSeek.
The Hangzhou-based company disclosed the development through its official WeChat account while introducing its upgraded V3.1 model, sparking speculation that a major breakthrough in China’s chip ecosystem may be imminent.
While DeepSeek did not name the chip supplier, it confirmed that the new model has been designed to run on next-generation local processors using the UE8M0 FP8 data format, a method aimed at optimizing memory efficiency and computational performance.
DeepSeek’s V3.1 Signals Strategic Shift
DeepSeek’s latest model, the V3.1, combines reasoning and non-reasoning modes and is optimized for compatibility with the upcoming Chinese chips.
Introducing DeepSeek-V3.1: our first step toward the agent era! 🚀
🧠 Hybrid inference: Think & Non-Think — one model, two modes
⚡️ Faster thinking: DeepSeek-V3.1-Think reaches answers in less time vs. DeepSeek-R1-0528
🛠️ Stronger agent skills: Post-training boosts tool use and…— DeepSeek (@deepseek_ai) August 21, 2025
The company emphasized that its use of UE8M0 FP8 scaling could lower memory requirements by up to 75%, making training and inference far more efficient compared to conventional approaches.
The AI firm had previously relied on foreign hardware, including 2,048 Nvidia H800 GPUs to train its earlier V3 model. However, with access to advanced U.S. chips increasingly restricted, the pivot toward homegrown processors indicates a deliberate shift in strategy, integrating domestic hardware with software designed for maximum efficiency.
Domestic Competition Intensifies
DeepSeek’s announcement comes as Chinese technology giants including Huawei Technologies and Moore Threads accelerate efforts to make their self-developed chips compatible with local AI models.
Huawei’s Cloud Matrix architecture, combined with its Ascend chip series, has reportedly demonstrated the ability to run DeepSeek’s models more efficiently than Nvidia-based systems under certain workloads, according to joint research published earlier this year.
This reflects a growing trend of hardware-software co-design in China, where firms are not merely attempting to replicate Nvidia’s products but are instead tailoring chip designs for specific AI applications. Such specialization could allow Chinese firms to gain unique advantages in targeted areas, even if performance gaps remain in broader applications.
Export Controls Drive Urgency
The push for domestic chip development has been fueled in large part by U.S. export restrictions that limit Chinese access to advanced semiconductors. Analysts project that domestic AI chip production could rise to 55% of total demand by 2027, up from today’s relatively modest share.
Despite this progress, Nvidia chips remain highly sought after in China. In 2024 alone, Chinese firms purchased nearly one million Nvidia H20 chips compared to around 450,000 Huawei Ascend 910B chips. Industry leaders such as Tencent and Baidu continue to favor Nvidia’s hardware, citing superior memory capacity and bandwidth.
To bridge this gap, Beijing has committed more than $340 billion in funding through its semiconductor “Big Fund” and related initiatives, underscoring the state’s determination to build a resilient and competitive chip industry.
