DeepSeek V5 to Pre-Train on 160K Huawei Ascend Chips

DeepSeek is preparing to train its next-generation V5 model on a fleet of at least 160,000 Huawei Ascend 950DT chips, marking a significant hardware shift for the Chinese AI lab. The deployment will take place at a new data center in Ulanqab, Inner Mongolia, with pre-training workloads rumored to begin in late 2026. The move represents DeepSeek's most ambitious bet yet on domestically produced AI accelerators as US export restrictions tighten access to advanced Nvidia hardware.

From Inference to Pre-Training

The Ulanqab chip cluster was originally earmarked for inference, the less compute-intensive task of serving model responses to user prompts. DeepSeek has now reportedly redirected the entire 160,000-chip fleet toward pre-training, the multi-month process where models ingest enormous datasets and learn underlying patterns. Pre-training demands far greater computational stability and throughput than inference, making the hardware pivot a higher-stakes engineering challenge.

DeepSeek has struggled in the past with earlier-generation Ascend chips, encountering stability and performance issues during training runs. To avoid a repeat, Huawei engineers have embedded inside DeepSeek's offices to co-design the chip architecture around the lab's upcoming model requirements. The collaboration aims to shape the hardware before models are built, rather than forcing models to adapt to fixed chip specs after the fact.

Hardware Economics and the Four-to-One Trade

DeepSeek founder Liang Wenfeng has publicly stated that it takes four Ascend 950 chips to match the functionality of a single Nvidia GB300 GPU. That ratio implies DeepSeek will need to deploy more silicon, consume more power, and occupy more data center space to achieve comparable training performance. The lab appears willing to absorb those costs in exchange for supply chain security, a calculation driven by US export controls that have cut off Chinese AI developers from the latest Nvidia chips.

The Ascend 950DT carries 144 GB of HBM-style memory and offers around 4 TB/s of bandwidth, positioning it as Huawei's flagship chip for both training and inference. Huawei is targeting a commercial release in Q4 2026, leaving DeepSeek minimal margin between chip availability and the start of V5 training. Any production delay on Huawei's side could directly push back DeepSeek's timeline.

Software Stack and Utilization Gains

DeepSeek has open-sourced several software libraries optimized for Ascend hardware, reportedly achieving over 95% of the Ascend 950's peak performance during internal tests. High utilization rates are critical when using chips that lag behind Nvidia on raw compute, as they help close the gap through software efficiency. The libraries are part of a broader strategy to build domestic hardware independence by pairing open-source tooling with tightly optimized inference and training pipelines.

The lab's software work extends beyond its own models. By releasing Ascend-optimized libraries publicly, DeepSeek is effectively betting that a broader ecosystem of Chinese AI developers will adopt the Huawei stack, creating network effects that improve tooling quality and reduce integration friction over time.

Rumored Timeline and Risks

DeepSeek V5 is rumored to release on an unspecified late 2026 timeline, contingent on both Huawei's Q4 2026 chip launch and the success of early pre-training runs. The tight coupling between hardware availability and model development leaves little room for setbacks. If Huawei's production schedule slips or early training runs reveal new stability issues, the V5 timeline could extend into 2027. The lab's roadmap reportedly points to mainstream adoption of the Ascend stack by 2027, with V5 serving as a proof-of-concept for large-scale training on domestic chips.

What This Means

DeepSeek's pivot to Huawei hardware at scale signals a broader shift among Chinese AI labs forced to navigate US export restrictions. The 160,000-chip deployment is one of the largest known commitments to non-Nvidia AI accelerators, and its success or failure will shape how aggressively other Chinese developers pursue domestic alternatives. If V5 trains smoothly on Ascend 950 chips, it strengthens the case for homegrown hardware independence. If the deployment hits technical roadblocks, it may reinforce the view that Nvidia's GPU lead remains difficult to replicate, even with custom software stacks and embedded co-design partnerships.