JD.com runs 700M-user item-AI stack on Huawei Ascend NPUs

A new paper on arXiv from JD.com's own AI group is worth reading twice, because the interesting claim is not the model, it is where the model runs. JD says its Oxygen AI Item Center, the platform that keeps the product catalog structured and understood, now processes hundreds of millions of item updates per day on Huawei Ascend NPUs. The user base behind that catalog is over 700 million active users and millions of merchants, sitting on tens of billions of SKUs.

The scale is eye-catching, but the more useful signal is that a full LLM and VLM stack, not a bolt-on classifier, is doing item-understanding work in production. The team reports 94.2% precision and 82.8% recall on knowledge production, covering tens of thousands of JD categories. Downstream, the paper claims search-traffic coverage of 80.4%, a 37% drop in item-information quality issues, and an automated fill rate above 80% for core attributes when a merchant lists a new SKU. Those are the operational numbers a category planner or a search PM actually cares about.

Why this matters beyond one company: the ongoing debate about whether China's domestic AI stack can carry serious production load, or whether it is a lab-bench exercise while the real work still runs on Nvidia, now has a concrete counterexample on the record. A 700-million-user commerce platform is stating in writing that its item-knowledge pipeline runs on Huawei silicon at hyperscale. Regulators, enterprise buyers, and the other Chinese hyperscalers will notice.

The honest caveat is that this is a self-published paper by JD's own group, posted to arXiv on 26 June 2026, and none of the headline numbers have been independently benchmarked. The write-up does not tell you which Ascend generation is in use, how many chips are in the fleet, whether the LLMs and VLMs were trained on Ascend or only served on it, or how the precision and recall figures compare to a Nvidia-hosted baseline. Those are the questions to keep in your pocket when someone waves this paper at you.

Still, the direction is the part worth watching. If the numbers hold up under third-party scrutiny, Huawei has its first credible hyperscale reference for the Ascend line, and Alibaba, Tencent and Baidu have a paper they can point to when they defend their own domestic-chip roadmaps.