DeepSeek V5 reportedly planned to be pre-trained on Huawei Ascend chips
hedgehog0 · reddit · 2026-11-03
According to STAR Market Daily sources across the supply chain, DeepSeek's next-gen model V5 is planned to be pre-trained on Huawei Ascend chips — until now Ascend was mainly used for inference. Key background from the post:
- Spring Festival 2025: Huawei leadership ordered rapid support; 3,000+ Ascend chips were deployed in days to host V3/R1 at Huawei Cloud's Guian datacenter.
- After V4 launched, Huawei's "super-node" lineup provided full support; DeepSeek notes V4-Pro throughput is constrained by high-end compute, with price cuts expected once Ascend 950 super-nodes ship at scale in H2.
- On Sept 30, DeepSeek open-sourced Ascend-tailored infra components (TileLang, DeepGEMM, FlashMLA, TileKernel, DeepSelect, plus the Ascend SuperPoD Flex architecture and UBL128 networking) — systematic co-development of the Ascend software stack.
- A Shenzhen/HIT team recently completed full-parameter post-training of the 1.6T-parameter DeepSeek-V4-Pro on Ascend 910C clusters (pre-training excluded).
If V5 ships pre-trained on Ascend, domestic chips would enter the pre-training stage for the first time; ecosystem migration cost for training remains the key variable.
More from Infra
- Underrated free compute: Oracle's 1,500 ARM hours/month, VS Code subs, Cloud Run and more — vincent_koc · 2026-11-03
- SGLang HiCache lifts agentic KV cache hit rate from 30% to 95% in production — AccBalanced · 2026-11-03
- 500M-row benchmark: hand-written BFS beats NebulaGraph in MySQL, 2.7ms vs 23.6ms — JeremyCMorgan · 2026-11-03
- Anthropic open-sources FlashProve: provable inference costs ~165x more than raw inference — arthurcolle · 2026-11-03
- PostHog re-rebuilds its data warehouse five months after switching from ClickHouse to DuckDB — teodorio · 2026-11-03
- Local MoE inference on a Tesla V100: 55 tok/s decode with 8,109 experts cached in 31.6GB VRAM — Traditional_Bell8153 · 2026-11-03