DeepSeek reportedly plans to train its V5 AI model on Huawei chips

DeepSeek plans to pre-train its next-generation model, DeepSeek V5, on Huawei’s Ascend AI chips, according to a STAR Market Daily report citing multiple industry sources. If accurate, it would mark a shift for the Chinese lab, because Ascend chips have so far been used mainly for inference, the work of running a finished model, rather than for building DeepSeek’s models from scratch.

A second Chinese outlet, Caiwen News, also citing industry sources, reports that DeepSeek intends to train V5 primarily on Huawei’s Ascend 950 chips, with the project expected to materialise in 2027. Neither DeepSeek nor Huawei has confirmed the plan, and the report has not been independently verified.

What the reports say about DeepSeek V5 and Ascend

The two reports overlap but are not worded identically. STAR Market Daily, as summarised by Geopolitechs, describes a plan to use Ascend chips to pre-train V5. Pre-training is the first and most compute-hungry phase of building a large language model, where it learns from a huge body of data before any fine-tuning. Caiwen’s sources, as summarised by TrendForce, say V5 could become DeepSeek’s first large language model trained natively on Ascend 950 chips.

Caiwen’s account adds several details:

• Ascend remains DeepSeek’s primary training platform, with no reported adoption of chips from other Chinese makers, including Moore Threads.
• Huawei holds no stake in DeepSeek. Cooperation is limited to products, computing resources and technical optimisation.
• That point comes amid market chatter about DeepSeek securing substantial new funding and speculation over a possible IPO.

One tension is worth flagging. Caiwen’s sources call Ascend DeepSeek’s “primary training platform”, while STAR Market Daily says Ascend has so far been used primarily for inference. The sources do not reconcile the two descriptions.

Item	What the sources say
Model	DeepSeek V5, successor to V4
Reported training chip	Huawei Ascend 950
Training phase reported	Pre-training
Development status	Selection phase; specs, capabilities and timeline undisclosed
Expected release	Sometime in 2027, depending on training progress
Huawei equity stake	None, per Caiwen’s sources

Why DeepSeek is reportedly turning to Huawei for training

According to Caiwen, the shift appears to be driven more by practical computing limits than by outside pressure. With US export restrictions limiting access to Nvidia’s latest high-end AI chips, previously approved H-series GPUs are reportedly struggling to meet the training demands of models such as DeepSeek V4 and V5, making domestic alternatives increasingly important.

The move would be a departure from DeepSeek’s earlier dependence on Nvidia. TrendForce points to a January Reuters report saying Nvidia engineers helped DeepSeek optimise its algorithms, software frameworks and hardware, and that DeepSeek-V3 required 2.788 million H800 GPU-hours for full training. Winzheng similarly reports that DeepSeek mainly used Nvidia H800 chips before gradually moving to domestic hardware in 2026.

For broader context on how China is approaching rules around frontier AI, see Unicorn’s earlier report on China’s first draft export rules for frontier AI models and agents.

How the DeepSeek–Huawei partnership got here

The relationship is not new. Around Chinese New Year in early 2025, Huawei reportedly deployed more than 3,000 Ascend chips to support DeepSeek V3 and R1 on Huawei Cloud.

When DeepSeek released preview versions of V4 in April 2026, the Associated Press reported that the new open-source models, in “pro” and “flash” versions, were supported in part by Huawei chips, reducing DeepSeek’s reliance on US chipmakers like Nvidia. Huawei said in a separate statement that its Ascend chips and related technology were compatible with the V4 models. Geopolitechs adds that Huawei announced its Ascend SuperPoD systems could fully support the V4 family through joint hardware-software optimisation. Both V4 versions have a 1 million token context window, up from 128,000 tokens in V3, according to DeepSeek.

The next step came on September 30, when DeepSeek open-sourced infrastructure components built for Huawei’s Ascend platform: the TileLang compiler toolchain, a high-performance computing library and a distributed communication library. DeepSeek said these map one-to-one onto the components it had previously open-sourced for Nvidia’s platform. Huawei, in parallel, published the companies’ joint work around Ascend 950 and its super node systems, covering model deployment, large-scale training, KV cache pooling for very long text and agentic reinforcement learning.

There has also been progress on training, though not yet the full kind. A research consortium in Shenzhen recently completed full-parameter post-training of DeepSeek V4-Pro, a 1.6-trillion-parameter model, on a cluster of Huawei Ascend 910C chips. Post-training refines an already-built model; pre-training from scratch remains considerably more demanding.

The software groundwork behind a move off Nvidia

Training a frontier model depends as much on software as on chips. TileLang is DeepSeek’s high-level language for writing operators, the low-level routines that do a model’s maths on a chip. DeepSeek says TileLang already carries most of the operator implementations used in training the V4 series on Nvidia hardware, and that every TileLang operator used in its training has a corresponding high-performance implementation on Ascend.

The September release also included five libraries: DeepGEMM for matrix operations, DeepEP for cross-device communication, TileKernels for general vector computation and memory access, FlashMLA for sparse attention and DeepSelect for data selection. On a 128-card Ascend 950 super node, Huawei says tests showed dispatch bandwidth of 375 GB/s and combine bandwidth of 347 GB/s for DeepEP’s communication, describing both as close to the hardware ceiling. Those are the companies’ own figures.

Independent coverage has stressed the limits. Superpower Daily notes that DeepGEMM-Ascend requires Ascend 950-series hardware and Huawei’s CANN 9.20 software, while DeepEP-Ascend also requires UBMEM connectivity, so the tools cannot be tried across Huawei’s existing chip fleet. Published benchmarks were run on a proof-of-concept test kit that DeepSeek said was not publicly distributed, which limits reproducibility. The CUDA-like interfaces offer a migration path, the outlet argues, not proof of a performance advantage.

Winzheng goes further, describing the toolchain as giving Chinese labs, for the first time, a developer toolchain that can fully replace Nvidia’s CUDA. That is the outlet’s characterisation; it also cautions that the real effect depends on developers’ willingness to migrate and the maturity of the ecosystem.

Huawei’s accelerated Ascend chip roadmap

The report lands as Huawei pulls its chip schedule forward. Citing Huawei Rotating Chairman David Wang, STAR Market Daily reported that the Ascend 960DT is now expected to be ready in the first quarter of 2027, three quarters ahead of its original schedule. According to Guancha, the Ascend 960PR is slated for the third quarter of 2027, one quarter early, followed by the Ascend 970 in 2028 and the 980 in 2029.

The reported V5 plan centres on the Ascend 950, not these later chips. The sources do not say whether DeepSeek would move V5 training onto newer Ascend parts as they arrive.

What it means for developers building on DeepSeek models

Nothing changes for builders today. V5 has no published specifications, capabilities or launch date, so there is no new model, context window or price to plan around. DeepSeek’s current models describe themselves as open source, letting developers modify and build on the core technology, and the company offers a free web and mobile chatbot.

The report matters as a signal about supply and continuity:

• A test of Chinese hardware at the frontier. Successful pre-training on Ascend would show whether Chinese AI accelerators can handle the full pre-training of a frontier model, rather than mainly inference and post-training.
• Less dependence on one supplier. If it works, Chinese AI developers could reduce reliance on Nvidia hardware, and Huawei’s position as a domestic alternative would strengthen.
• Portable tooling. Teams that already use DeepSeek’s open libraries on Nvidia now have Ascend counterparts with matching interfaces, though they require specific Ascend 950 setups.

The risks are equally concrete. Geopolitechs notes that the plan’s technical and commercial viability will depend on training stability, cluster performance, software maturity and the resulting model’s capabilities. Caiwen’s sources say the transition is still in its early stages and that timing will hinge on training progress. Even for V4, a Morningstar analyst told the AP that independent evaluations were needed before final conclusions could be drawn about DeepSeek’s claimed performance against US models.

Readers following DeepSeek’s developer tools can also see Unicorn’s report on DeepSeek open-sourcing Super-agent, a background coding agent prototype.

What to watch next

• Confirmation. Whether DeepSeek or Huawei publicly confirms that V5 is being pre-trained on Ascend, rather than reports from unnamed sources.
• A technical report. Whether a V5 release documents the training hardware, cluster size and stability, the factors Geopolitechs identifies as decisive.
• Timing. Industry sources point to 2027, contingent on training progress.
• Reproducible tooling results. Whether developers outside DeepSeek can obtain the required Ascend 950 setup and reproduce the published library benchmarks.
• Huawei’s roadmap. Whether the Ascend 960DT arrives in the first quarter of 2027 as now expected.