DeepSeek V5 Eyed for Ascend 950 Training, Not Just Inference

China's open-weight champion may be about to cross the river that still separates domestic chips from true independence. On November 3, 2026, Chinese financial media including Cailian / Caiwen and STAR Market Daily, citing industry supply-chain sources, reported that DeepSeek plans to pre-train its next-generation model DeepSeek V5 with Huawei Ascend 950 as the core training compute, aiming at a 2027 landing. Ascend silicon has so far been used mainly for inference and some post-training on DeepSeek stacks. V5 would be the first DeepSeek flagship described as natively trained on Ascend 950.

Neither DeepSeek nor Huawei has officially confirmed the plan. Treat it as a high-signal leak, not a launch note. If it holds, it is the most important China AI hardware story since DeepSeek's V3 efficiency shock: the last workload that "had to be Nvidia" becomes negotiable.

What Was Reported

DeepSeek V5 Ascend training is a reported plan to run the model's pre-training primarily on Huawei Ascend 950-class accelerators, with project timing pointed at 2027.

Earlier open-sourcing matters here. On September 30, 2026, DeepSeek released Ascend-side infrastructure components (TileLang tooling, DeepGEMM / FlashMLA-class libraries, DeepEP communication) mirroring its Nvidia stack. That is what you do before you trust a months-long pre-training run.

Inference Was Solved. Training Was Not.

Chinese accelerators already carry a large share of domestic serving traffic. Pre-training is different: longer runs, harder all-reduce behavior, nastier failure modes, and less forgiveness for software immaturity.

A June research collaboration reported full-parameter post-training of DeepSeek-V4-Pro on at least 1,000 Ascend 910C chips. That is impressive and still not the same as birthing a frontier model from random initialization on a new ecosystem.

How This Connects to Ulanqab and 160,000 Chips

Bloomberg and follow-on coverage earlier described DeepSeek planning a roughly 1 GW class facility in Ulanqab (Inner Mongolia) with on the order of 160,000 Ascend 950DT accelerators, initially framed around inference. The new V5 training reports sit awkwardly against that framing. Possible reconciliations:

We covered the inference-cluster arc in DeepSeek's 160,000 Ascend / 1 GW Ulanqab plan and the chip family in Ascend 950PR.

Huawei's own roadmap splits 950PR (inference / prefill leaning) and 950DT (full lifecycle training and inference), with 960DT pulled forward toward early 2027 in some executive comments. A V5 training claim is also a Huawei capacity claim.

Our Analysis: Strategic Meaning

1. Export controls are working as industrial policy for Huawei. Every closed Nvidia door raises the ROI on Ascend software pain. The Nvidia 51% cloud rule and Taiwan's smuggling crackdown (chip bill hearing) push in the same direction.

2. Open weights plus domestic training is a bloc strategy. If DeepSeek can train on Ascend and release competitive open weights, China exports influence without exporting banned GPUs. That hits Global South builders who already optimize for cost (Altman's developing-country adoption thesis).

3. Single-vendor risk moves from Nvidia to Huawei. DeepSeek sources say Ascend is the primary training partner, not a multi-domestic GPU bazaar. That concentrates schedule risk in Huawei yields and SuperPoD networking.

4. Benchmarks will decide if this is sovereignty theater. A 2027 Ascend-trained V5 that lags V4-class Nvidia-trained quality becomes a patriotic demo. One that matches changes procurement forever inside China and renegotiates every "CUDA or nothing" assumption outside it.

Original Forecast: Three 2027 Scenarios

Our base case is hybrid with a loud Ascend marketing wrapper. Even that matters. Once procurement teams inside China can point to a named DeepSeek training run on 950DT, every provincial cloud tender gets easier for Huawei. Outside China, the effect is price: open weights that cost less to produce domestically keep undercutting API oligopoly margins.

Compare that path with Nvidia trying to keep SEA clouds majority non-Chinese. Both stories are the same story from opposite ends: the training graph is fragmenting into blocs whether executives admit it or not.

What "Native Ascend Training" Would Require Technically

A serious pre-train is not a weekend conversion job. DeepSeek would need:

1. Stable collective communication across thousands of accelerators without silent divergence
2. Operator coverage for attention, MoE routing, optimizers and checkpointing that matches CUDA maturity
3. Failure recovery that does not lose days of progress on flaky links
4. Numeric parity close enough that research recipes transfer without months of retuning

The Sept 30 open-source dump attacked pieces of (1) and (2). Hangzhou co-location attacks (2) and (4). Capacity and yields attack everything else. That is why unverified 2027 timing is believable as intent and still fragile as a calendar commitment.

For outside developers, the tell is not a press leak. It is whether DeepSeek's public model cards and server images start listing Ascend training tokens the way they once listed H800 hours. Until that appears, keep shipping on the stacks you can measure, and treat every Ascend rumor as a prompt to dual-track kernels rather than a reason to freeze product work.

What Developers Should Do

• Watch for DeepSeek operator commits and training blogs, not only financial-media leaks.
• If you serve China users: assume Ascend-compatible kernels become table stakes for serious onshore deploy.
• If you buy open weights globally: an Ascend-trained V5 may still run on Nvidia at inference time via conversion, but legal and perf paths will fork.
• Keep FinOps dual-track. Cheap DeepSeek APIs already force frontier vendors to compete (V4.1 Flash); domestic training strengthens that price pressure. Track live rates on LLM API Pricing.
• Re-read the China compute map via the AI chip supply chain hub and China data-center grid build.

What To Watch Next

• Official DeepSeek or Huawei confirmation
• Whether V5 training starts on 950DT before 960DT arrives
• Any disclose of chip count, tokens and wall-clock for the run
• Quality deltas vs V4 Pro / Flash on coding and agent benches
• Huawei allocation of 2027 Ascend output toward DeepSeek vs other labs

Key Takeaways

• Reports: DeepSeek V5 may pre-train on Huawei Ascend 950, targeting 2027
• Would mark Ascend's jump from mostly inference into DeepSeek pre-training
• Hangzhou joint engineering and Sept 30 Ascend open-source stack are the prep work
• Unconfirmed by either company; specs and date remain internal
• Strategic read: export controls accelerate a China-native full model lifecycle
• For developers: plan for Ascend-compatible tooling and continued open-weight price pressure
• What to watch: official confirmation and first training-quality disclosures

Sources

• Cailian / Caiwen and STAR Market Daily supply-chain reports (Nov 3, 2026)
• TrendForce and Commercial Times English/Chinese summaries (Nov 3 to 4, 2026)
• DeepSeek Sept 30, 2026 Ascend infrastructure open-source release
• Prior Bloomberg reporting on Ulanqab Ascend deployment plans
• Huawei Ascend 950PR / 950DT public roadmap comments