DeepSeek V5 Targets Huawei Ascend 950DT for Pre-training: Domestic AI Chips Move into Core Training

DeepSeek V5 Pre-training on Ascend 950DT

According to The Information and confirmed by insiders, DeepSeek aims to begin pre-training its V5 model on Huawei Ascend 950DT chips as early as late 2026. Testing is ongoing and timelines may shift. This would extend domestic Chinese compute from inference into the core training phase of large-model development.

Hardware–Software Co-design Across Two Chip Generations

Huawei engineers have been stationed at DeepSeek for long-term collaboration. The partnership now covers two future Ascend chips:

Ascend 950DT – targeted for V5 pre-training in late 2026.

Ascend 960DT – slated for Q1 2027 release, with most supply reserved for DeepSeek training and inference.

The cooperation moves beyond simple deployment: model compute requirements, inter-chip data transport, and communication patterns are feeding directly into hardware architecture decisions. As one observer noted, this lets model needs "define hardware architecture in reverse."

DeepInfer: Unified Inference Framework for Ascend and NVIDIA

On October 20, DeepSeek open-sourced DeepInfer , a large-scale cloud inference framework that integrates high-performance operators and an inference engine. Key capabilities:

Distributed inference, speculative decoding, async scheduling, structured output.

Simultaneous support for Huawei Ascend 950-series NPU and NVIDIA SM100 (Blackwell) GPU , providing a single framework across domestic and overseas compute platforms.

Large-scale Deployment Practices Revealed

Accompanying DeepInfer, DeepSeek published "DeepSeek V4.1 Flash Large-scale Inference Deployment Practice," detailing:

Throughput: 155 trillion tokens per day per cluster.

Revenue estimate: ~$38 per GPU per hour (at list price).

KVCache retention: 72+ hours for history reuse.

Topology: 8-card machines for Prefill; 64–128 card supernodes for Decode with massive expert parallelism.

Rationale: Prefill handles input processing, Decode handles token-by-token generation; separate resource allocation improves overall cluster efficiency.

Software Toolchain Adaptation for Ascend

In September 2025, DeepSeek open-sourced Ascend-targeted components:

TileLang – higher-level operator programming reducing direct dependence on low-level APIs.

DeepGEMM and DeepEP – high-performance compute and distributed communication libraries for Ascend.

These pieces aim to bridge model, chip, and software toolchain into a cohesive stack.

Historical Context: From 3,000-chip Deployment to V5

2025 Spring Festival: Per Fang Xingdong's book "Ascend Rise," Huawei and DeepSeek deployed and validated 3,000+ Ascend chips, running DeepSeek-V3/R1 at Huawei Cloud Guian data center. Ren Zhengfei personally coordinated the response to traffic surge.

April 2026: After DeepSeek-V4 release, Huawei announced full Ascend super-node support for the V4 series.

Next step: V5 pre-training on Ascend would validate domestic chips for the most demanding training workloads, offering a reference for a CUDA-independent software-hardware ecosystem.