DeepSeek unveils V4 model using Huawei chips, undercuts US labs on price.
The AMW Read
Novelty 2: updates DeepSeek's case study with V4 and Huawei chip integration, a meaningful advance but not a paradigm shift. Significance 3: pricing and chip independence have cross-segment impact on foundation model economics and geopolitics.
DeepSeek unveils V4 model using Huawei chips, undercuts US labs on price.
Chinese AI startup DeepSeek has released its V4 family of large language models, consisting of V4-Pro (1.6 trillion parameters) and V4-Flash, trained on Huawei's Ascend AI processors. The Pro version costs $3.48 per million output tokens, roughly one-tenth the price of OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6, and is claimed to trail these frontier models by just three to six months on key benchmarks. The cheaper Flash variant costs $0.28 per million tokens. Shares of Chinese chipmaker SMIC, which produces Huawei's Ascend processors, jumped 10% on the news.
Why it matters: DeepSeek's rock-bottom pricing—$0.28/M tokens for Flash—and its assertion that performance rivals closed-source leaders intensifies pressure on US labs already facing questions about their competitive moats. The announcement also showcases a viable training alternative to NVIDIA GPUs via Huawei chips, validating China's sovereign AI compute strategy. By bucking the industry trend of price increases, DeepSeek is reasserting the fast-follower dynamic that emerged in 2024–2025, where Chinese open-source labs compress margins for all frontier labs. This move updates the capital-compression arc in foundation models and may push OpenAI and Anthropic to justify premium pricing as the performance gap narrows.
Expert Take: DeepSeek's V4 release exemplifies the pattern of rising Chinese AI labs using open-source distribution and aggressive pricing to force structural changes in the market. The use of Huawei Ascend processors underscores China's ability to train frontier models despite US export controls, though the model's reliance on these chips may limit performance compared to NVIDIA-trained counterparts. Long-term, DeepSeek's price war could commoditize foundation models, accelerating the shift of value to applications. However, the company's reported $20 billion fundraising round and talent retention challenges suggest even DeepSeek may need capital to sustain its trajectory.
#DeepSeek #V4 #Huawei #OpenSource #ChinaAI #FrontierModels #PriceWar #SovereignAI
