DeepSeek plans a 160,000-chip Huawei cluster in Inner Mongolia
Monday's stories are about what happens after the models ship: a Chinese lab committing to domestic silicon at a scale nobody has attempted, a record graduating class walking into a hiring market AI is quietly reshaping, and a retailer giving away its AI stack to move more hardware.
DeepSeek plans to install at least 160,000 Huawei Ascend 950DT accelerators at a data center it is building in Inner Mongolia, according to people familiar with the matter — a deployment that would rank among the largest known clusters of Huawei's AI chips anywhere. The processors are earmarked for inference, not training, and DeepSeek currently has no plans to train on them. They are also only part of the site, which is scoped for power draw on a gigawatt scale. Timing depends on Huawei's production capacity, so this is a plan rather than a finished build.
The inference-only detail is the part worth reading twice. Training runs are one-off capital events; serving is the bill that never stops, and it is where a lab with mass consumer traffic actually bleeds money. Betting that load on domestic accelerators is a stronger statement than buying them for a showcase training run — it says DeepSeek expects Ascend to carry production traffic at a volume measured in hundreds of thousands of chips. That is a very different claim from "we have a compliance hedge." As we wrote when its books leaked, Deep Dive — DeepSeek's leaked P&L shows frontier AI can run on a shoestring — and cheap only stays cheap if the serving layer holds.
A record 12.7 million graduates are entering China's workforce this year, into a job market that already has more seekers than openings — and AI is thinning the entry-level end of it. Researchers quoted in the coverage describe the exposed work as repeatable thinking: drafting routine reports, sorting information, handling standard customer questions. The tasks that used to be how a new hire learned the job are the tasks that get automated first. The other side is real too: state media reported AI-related postings up 16.9% year over year after the Spring Festival holiday, with demand for robotics algorithm engineers up 57%.
Outside China the evidence is uneven rather than apocalyptic. A Dallas Fed analysis found only small aggregate effects so far, while a Stanford Digital Economy Lab update has young workers in the most AI-exposed roles falling further behind their peers through mid-2026. The risk in both readings is not mass unemployment — it's a slower, meaner path from school to a first job, and a cohort that never gets the apprenticeship that used to come with one. We saw the same pressure arrive for white-collar hiring from the other direction in UBS now makes AI skill a hiring test for junior bankers.
JD is giving its JoyInside AI stack away to appliance and home brands, then backing it with 1 billion yuan in marketing over three years. Announced at a Shenzhen forum on September 6, the "JoyInside AI home and appliance leader program" opens the platform — built on JD's own JoyAI model — to any brand at no cost, with Midea and TCL among the launch partners. The stated target is to help 300 brands ship JoyInside-powered products, supported by JD's 18,000-plus stores. The company says it has already signed roughly 200 brands and shipped close to 100 products, with connected devices up more than threefold since last year's Singles Day.
The strategy is the interesting part: JD doesn't need to sell the model, it needs the appliances to sell. Freeing the AI layer turns a differentiator into table stakes, and the margin moves to the hardware and the channel — which is exactly what a retailer is positioned to capture. The tell is that the 1 billion yuan is marketing budget, not research budget. The hard part was never getting a fridge to answer questions; it's whether anyone keeps talking to it after week two.
What to watch: whether Huawei's Ascend output can actually feed a 160,000-chip commit on any schedule, and whether China's graduate cohort shows up in unemployment statistics or simply in underemployment no one measures.
If domestic chips can carry a lab's entire serving load, does the export-control argument still hold — or has it just moved the problem? Tell us in the comments.
