리포트에 따르면, 딥시크가 H20 GPU 주문을 취소하고 화웨이의 Ascend 950DT 칩 16만 개 구매를 선호하면서 엔비디아는 중국 시장에서의 노력을 축소하고 있는 것으로 나타났다.

DeepSeek has reportedly placed a massive order for 160,000 units of Huawei’s Ascend 950DT AI chips, valued at approximately $2.56 billion, to equip a new 1-gigawatt data center in Inner Mongolia. According to a Bloomberg report, the company has bypassed NVIDIA’s China-specific H200 GPUs in favor of Huawei’s domestically produced silicon, further narrowing NVIDIA’s footprint in an increasingly insular Chinese market. This procurement follows seismic comments made by DeepSeek CEO Liang Wenfeng in July, when he declared that NVIDIA was “digging its own grave” in China.

At current street prices, each Ascend 950DT chip costs roughly 111,000 yuan (approximately $16,000), slightly above the $15,000 to $25,000 range for NVIDIA’s H20 GPU. Despite the modest price premium, supply constraints remain a critical factor. Huawei’s manufacturing capacity for the Ascend 950DT is currently limited to a few hundred thousand units annually, meaning fulfilling DeepSeek’s order could take up to a year.

Technically, the Ascend 950DT is built on Huawei’s self-developed third-generation DaVinci v3 architecture and utilizes SMIC’s N+3 process technology, an advanced 7-nanometer-class node. While capable of handling both training and decoding workloads, DeepSeek’s order is primarily targeted at inference demands. Each chip features 144 GB of specialized High Bandwidth Memory (HiZQ 2.0 HBM), delivering 4.0 TB/s of memory bandwidth and a 2.0 TB/s GPU-to-GPU interconnect speed. The architecture natively supports low-precision formats including FP8, FP4, and Huawei’s proprietary HiF8. Within Huawei’s broader ecosystem, the chip integrates into the Atlas 950 SuperPoD infrastructure, which can scale up to 8,192 Ascend 950DT units to deliver 8 exaflops of peak AI compute power.

The strategic pivot aligns with Liang Wenfeng’s earlier assessment of the competitive landscape. In July, he stated that “all tasks [NVIDIA] GB300 can do, the Huawei supernode can do, latency and all the same. The only cost: 4 Huawei GPUs equal 1 NVIDIA GPU, and it's two years behind.” By accepting a roughly fourfold performance penalty and a year-long delivery delay for domestic hardware, DeepSeek’s decision underscores a broader industry realignment toward Chinese alternatives.

Industry observers note that NVIDIA appears to have recalibrated its strategy regarding the region. Steven Glinert, CEO of Sphere Semi, noted on September 4, 2026: “Hearing recently that nvidia (who used to push anyone too hawkish on china to be fired from their govt or think tank job) has changed their tun to an extent. chinese domestic market is seen as a lost cause vs Huawei, so they’re not *as* committed to being able to sell there.” Consequently, NVIDIA seems to have effectively conceded the Chinese market to Huawei, deprioritizing sales efforts in a region where domestic competition has become insurmountable.

Meanwhile, Inner Mongolia continues to solidify its position as a central hub for China’s AI infrastructure expansion. The region’s cold, arid climate is highly conducive to energy-efficient, solar-powered data centers. For example, power costs at Alibaba’s Ulanqab facility range between 0.32 and 0.35 yuan per kilowatt-hour, translating to approximately $0.05 per kWh. Demand remains robust, with ByteDance currently negotiating to add 5 to 6 gigawatts of capacity in the province, where the estimated cost for each 1-gigawatt installation hovers around $23.8 billion.