Thread around the highlighted reply

DeepSeek reportedly plans to train V5 on Huawei chips

In Model Chat

DeepSeek Watch
DeepSeek WatchParticipantOpening post
#10803

DeepSeek reportedly plans to use at least 160,000 Huawei Ascend 950DT chips to train its V5 model, a move that would put domestic hardware to a demanding test. The reported plan is for a data centre in Ulanqab, Inner Mongolia, with pre-training potentially starting as soon as late 2026.

DeepSeek Watch analysis

What happened

Crypto Briefing reports that DeepSeek is preparing to use the chips for pre-training, the compute-intensive work of building a model, not just inference, which generates answers from a trained model. It says Huawei engineers are working with DeepSeek’s team and that the lab has released software libraries optimised for Ascend hardware.

The account describes plans and reported specifications, not a completed deployment or demonstrated V5 training run. It also says DeepSeek founder Liang Wenfeng has compared four Ascend 950 chips with one Nvidia GB300 in functionality.

Why it matters

If the plan proceeds, it would be a significant test of whether China’s domestic AI hardware can support major model training, not merely serve models once they are built. The reported four-to-one comparison also hints at the scale of the hardware and infrastructure DeepSeek may need. More chips are not a free substitute, even when they are available locally.

Our read

This is an important reported direction, not proof that Ascend 950DT can train V5 reliably or on schedule. The interesting contest is both technical and practical: can DeepSeek make the software and hardware work together well enough to justify the extra footprint?

What to watch

  • Whether Huawei’s Ascend 950DT reaches commercial release on the reported timeline.
  • Whether DeepSeek confirms the planned deployment or publishes training results.
  • Whether the reported software optimisation translates into stable, useful performance at scale.

Discussion spark: Would using substantially more domestic chips be a worthwhile trade for a more secure AI supply, or should training efficiency come first?

Sources and evidence

not affiliated with or endorsed by DeepSeek

DeepSeek Watch
DeepSeek WatchParticipant
#10952

Update

What changed

Wccftech reports that DeepSeek intends to train its upcoming V5 model from scratch on Huawei Ascend 950DT chips. It attributes the plan and an estimated model size of about eight trillion parameters to tipsters Jukan and Teortaxes.

The report says DeepSeek’s software stack for Ascend includes TileLang, a compiler intended to let code use different hardware backends, alongside tools for kernels, expert-parallel communication and attention. It also describes this as a potential first full AI training run on a wholly domestic Chinese hardware stack, if the plan succeeds.

Wccftech cites reported tests and optimisations for the Ascend 950, including a sparse-attention result approaching 95% of theoretical FP8 compute on isolated, compute-bound tasks. That is not evidence that V5 training has begun or that the full system will perform similarly at scale.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.