DeepSeek Plans to Train V5 From Scratch on Huawei's Ascend 950DT SuperPoDs

On November 3, Wccftech reported that DeepSeek plans to train its upcoming V5 model from scratch on Huawei's Atlas SuperPoD clusters built around the Ascend 950DT accelerator. The outlet framed the move as part of China's push for domestic full-stack AI dominance, and the emphasis is less on a routine model refresh than on where the pretraining compute comes from.

The load-bearing phrase in the report is "from scratch." Running an already-finished model on non-Nvidia hardware is a serving problem; completing a full pretraining run on a domestic cluster is a different class of engineering, one that puts distributed training, parallelism strategy, interconnect scheduling and fault recovery on the critical path at the same time.

Ascend 950DT belongs to Huawei's Ascend 950 line of AI accelerators, while Atlas SuperPoD is Huawei's supernode architecture, which links many accelerators into what behaves like a single, much larger compute unit. The design logic is to close the gap with leading general-purpose GPUs through system-level scale and tight interconnect rather than by matching them chip for chip.

What the public record does not yet contain is equally important. The report offers a routing decision, not verifiable numbers: no cluster size, no training start date, no detail on what V5 is meant to be good at, and no accompanying statement from DeepSeek or Huawei. The accurate reading is a clear signal of technical direction, not evidence of a completed run.

That distinction matters because China's domestic compute story has so far been mostly about inference. Moving released models onto domestic accelerators is a bounded, lower-risk exercise that produces visible results quickly. Handing pretraining to a hardware and software stack that is still maturing is a much bigger bet, because it gives up the easiest place to look competent.

If V5 does complete a from-scratch run, the deliverable would be more than a model. It would be a reusable engineering reference: which parallelism schemes hold up on Ascend, whether supernode interconnect stays stable across weeks of training, and whether the domestic software stack can absorb the operator requirements of mainstream training frameworks. For the wider supply chain, those answers may be worth more than V5's benchmark scores.

The risks are just as concrete. Very large pretraining runs are brutally sensitive to cluster stability, where a single communication bottleneck or hardware fault can stretch the schedule and inflate cost, and ecosystem maturity cannot be bought with one project. Whether the route works will be decided by the quiet costs inside the training run, not by the numbers on a launch slide.

Three signals are worth watching. Whether DeepSeek discloses the hardware and cluster scale behind V5 when it ships; whether Huawei promotes the work as a public reference case for Atlas SuperPoD; and whether other Chinese model teams follow by moving pretraining wholesale onto domestic compute. If those appear in sequence, the report will have been confirmed rather than merely repeated.