THE ESSENTIALS IN 30 SECONDS

  • STAR Market Daily, citing unnamed industry sources, reported on November 3 that DeepSeek plans to pre-train its next model, V5, on Huawei Ascend chips.
  • The Information reported in October that DeepSeek aimed to begin pre-training on the Ascend 950DT as soon as the end of 2026, with Huawei engineers working in its offices.
  • Bloomberg had earlier reported a plan for at least 160,000 Ascend 950DT chips at a site in Ulanqab, Inner Mongolia, then described as mainly for running models.

DeepSeek intends to pre-train its next major model, DeepSeek V5, on Huawei's Ascend processors, according to a November 3 report by STAR Market Daily, a financial outlet in the Cailian Press group that cited several unnamed industry sources. If accurate, it would mark a shift: Ascend chips have so far been used mainly to serve DeepSeek's models rather than to build them from scratch. Neither DeepSeek nor Huawei has confirmed the plan, and V5 itself has not been announced.

What has been reported

The report fits a series of earlier leaks. In October, The Information wrote that DeepSeek wanted to start pre-training on Huawei's Ascend 950DT as early as the end of 2026, and that Huawei engineers were based in DeepSeek's offices, feeding its input into chip design. According to that report, the 950DT is due in the fourth quarter of 2026 with 144GB of Huawei's own HiZQ 2.0 high-bandwidth memory, and DeepSeek is to get priority on most of the following 960DT generation, expected in early 2027.

In September, Bloomberg reported that DeepSeek planned to deploy at least 160,000 Ascend 950DT chips at a data centre it is building in Ulanqab, Inner Mongolia, with roughly 1GW of capacity, part of it to be running by late 2027 or early 2028. At the time, that hardware was described as mainly for running models, not training them. TrendForce, citing The Information, also reported that founder Liang Wenfeng told investors DeepSeek would expand its use of domestic chips for training.

From inference to pre-training

The ties between the two companies are not new. Huawei reportedly deployed more than 3,000 Ascend chips to host DeepSeek V3 and R1 on its cloud in early 2025. On September 30, DeepSeek open-sourced compilers, compute libraries and communication software adapted to Ascend. A Shenzhen research group has also run full-parameter post-training of the 1.6-trillion-parameter V4-Pro on Ascend 910C chips, although that stopped short of pre-training.

What remains uncertain

Pre-training a frontier model is far more demanding than serving one: it needs thousands of chips running reliably together for weeks. Whether Ascend clusters can deliver that depends on training stability, cluster performance and software maturity, as Geopolitechs notes. Huawei's supply is another constraint. For now, the V5 plan rests on unnamed sources and should be treated as unconfirmed.