Every Outlet Ran 'China Trains Frontier AI on Domestic Chips.' What Actually Happened Was Post-Training. Pre-Training Hardware Is Undisclosed and Experts Suspect NVIDIA.

On June 5, 2026, a consortium including Huawei Technologies, the Shenzhen Loop Area Institute, Harbin Institute of Technology Shenzhen, and the Shenzhen Research Institute of Big Data announced that it had completed full-parameter post-training of DeepSeek V4-Pro on a cluster of more than 1,000 Huawei Ascend 910C chips. The run completed 1,500+ training iterations without interruption. The announcement came via Shenzhen government social media.

The headline that ran in nearly every Western and Chinese outlet was a variant of "China trains frontier AI on domestic chips." This framing is technically defensible and substantively misleading. Understanding the distinction is necessary to evaluate what the export control strategy has and has not accomplished.

Pre-training a large language model and post-training it are different operations by roughly two orders of magnitude in compute. Pre-training is the initial phase: ingest a corpus of hundreds of billions of tokens, run gradient updates across all model parameters for millions of GPU-hours, build the underlying predictive capability from scratch. For DeepSeek V3 — the predecessor to V4-Pro — this required 2,048 NVIDIA H800 GPUs running for 2.66 million GPU-hours. Post-training covers the alignment and instruction-following work: supervised fine-tuning, RLHF, DPO-class updates. This phase uses the same full set of model parameters but runs on smaller curated datasets, typically on much shorter timescales.

The Huawei announcement is about the post-training phase. DeepSeek V4-Pro's pre-training hardware is not disclosed in its technical report. Liu Zhiyuan, a professor at Tsinghua University, told MIT Technology Review that DeepSeek "appears to have adapted only part of V4's training process for Chinese chips, and the model may still have been trained mainly on Nvidia hardware." The China Academy's analysis agrees: "DeepSeek V4 hasn't fully cut ties with Nvidia." The working consensus among independent analysts is that V4-Pro's pre-training almost certainly ran on NVIDIA hardware — H800s acquired before the October 2023 export control escalation that banned them — and that the June 5 announcement describes the alignment work that followed.

The 60% H100 performance figure that appears in most coverage originated in a February 2025 self-assessment by DeepSeek, not from Huawei's June announcement. TrendForce reported it at the time as "DeepSeek reportedly reveals Huawei's Ascend 910C reaches 60% of Nvidia H100's inference power." The original claim is self-reported, the metric is "inference power" without a specified workload, and no independent third-party benchmark has confirmed it. Inference performance varies by 3-10x across different configurations (batch size, sequence length, precision format) for the same hardware; a claim that is true at one configuration may not hold at others. CSIS, in Gregory C. Allen's analysis, consistently qualifies this figure as "reportedly" — the appropriate epistemic standard for an unverified vendor claim with geopolitical significance.

The real engineering achievement is worth stating accurately. DeepSeek and Huawei spent months co-developing CANN-compatible code for V4, including fused operators and distributed training frameworks optimized for Huawei's software stack. Running 1,500+ uninterrupted post-training iterations on a 1,000-chip Ascend cluster at 1.6 trillion parameter scale is genuinely non-trivial engineering — the cluster stability and memory coordination required are not trivial at that parameter count. This is not a lab demonstration; it produced a commercial model. What it is not is evidence that China can pre-train a frontier model from scratch without NVIDIA hardware.

Cambricon Technologies, often mentioned alongside Huawei in this story, made a separate and distinct contribution. On April 24, 2026 — the date that appears as the signal date in most early coverage — Cambricon announced completion of inference adaptation for V4-Flash and V4-Pro on its own MLU chips, with code open-sourced to GitHub. This was inference adaptation (running an already-trained model for prediction), not training. Cambricon's work used the vLLM inference framework with five-dimensional hybrid parallelism optimized for MLU architecture. The April 24 Cambricon inference announcement and the June 5 Huawei post-training announcement are parallel milestones from different companies on different hardware. Coverage conflating them treats the combined picture as a single event.

Gregory C. Allen's CSIS analysis makes the most important argument about the export control strategy: the problem is not that the controls were poorly designed, but that enforcement failed before the controls were tightened. NVIDIA's H800 and A800 chips were sold to China at massive scale before the October 2023 export control escalation. Those chips are now in Chinese data centers. They cannot be recalled. DeepSeek's capability to pre-train frontier models almost certainly depends on compute acquired under the earlier, more permissive regime. The current export controls prevent future acquisitions; they cannot reverse past ones. Allen cites DeepSeek CEO Liang Wenfeng's July 2024 statement: "Money has never been the problem; bans on shipments of advanced chips are the problem" and that Chinese firms needed "two to four times more compute" for equivalent results. The efficiency gains that make DeepSeek's models economically viable are a response to scarcity — they do not demonstrate scarcity-independence.

The software gap is the less-discussed constraint. CANN — Huawei's answer to CUDA — is described by the China Academy as "an order of magnitude smaller" ecosystem than CUDA. PyTorch, TensorFlow, DeepSpeed, Megatron-LM, FlashAttention, and dozens of inference frameworks have accumulated years of CUDA-native optimization. CANN has none of that history. Huawei's CANN Next includes "CUDA-compatible programming abstractions that lower migration barriers," which is an acknowledgment that the compatibility problem remains live, not solved. Hardware performance gaps close with improved fabrication; software ecosystem gaps compound over time in the opposite direction.

The Ascend 950 is the forward-looking strategic signal that matters more than the 910C milestone. TrendForce's April 2026 analysis of the Ascend 950 projects 1 PFLOPS FP8 / 2 PFLOPS FP4 performance, 2 TB/s interconnect bandwidth, and 112 GB HiBL memory at 1.4 TB/s — positioning it "between H100 and H200." If these specs ship in volume in H2 2026 as projected, China moves from 60% inference parity to approximately 80-90% parity with H100. That is a qualitatively different competitive position. At 80-90% parity, the export control strategy's effectiveness in slowing frontier model development narrows to a margin that training efficiency gains can plausibly close. Volume production of the Ascend 950, not the post-training milestone that already happened, is the threshold that matters for the containment question.

What to watch: whether DeepSeek ever discloses V4-Pro's pre-training hardware; whether the Ascend 950 ships in the volumes and at the specifications TrendForce projects; whether an independent research institution publishes Ascend 910C performance benchmarks with specified methodology that can be evaluated against H100 in comparable workloads; and whether CANN accumulates third-party optimization work at the rate that would be required to close the software ecosystem gap within the 3-5 year horizon Huawei's public statements imply.