A gigawatt Ulanqab site, Huawei Ascend 950DT for running models, Nvidia for training. Beijing's quota is the competitor Nvidia cannot underprice.
Via ZeroHedge: DeepSeek's 160,000 Huawei Order Is A Real Threat To Nvidia In 2027
Bloomberg reported on September 4, 2026 that DeepSeek plans to deploy at least 160,000 of Huawei's next-generation Ascend 950DT accelerators at a data center in Inner Mongolia. ZeroHedge's write-up of that reporting puts the site in Ulanqab, about 350 kilometers northwest of Beijing, at gigawatt scale. If it lands as described, it would be one of the largest known clusters of Chinese-made AI silicon. Neither company has confirmed the plan on the record.
The number that matters for an owner-led firm is not the chip count. It is the job. Bloomberg's sources say DeepSeek intends to use the 950DT for operating its models — inference — not for training, even though Huawei marketed the chip for that heavier work. Training, the piece that still needs Nvidia-class interconnect, stays on Nvidia for now. Inference is the volume layer: chat answers, code completions, document summaries. That is the pool that pays.
This article is not investment advice and not a recommendation to buy, sell, or hold Nvidia, Huawei, or anyone else. It is a continuity note. If your workflows already call DeepSeek, or you treat “Chinese open weights plus a cheap API” as the default workhorse, the silicon under that API is becoming a policy object.
ZeroHedge, citing Bloomberg, says Huawei cannot fill the order quickly. In-house high-bandwidth memory on the 950 series is the bottleneck; 950DT output this year is described as the low hundreds of thousands, and fulfilling DeepSeek's contract could take well over a year. DeepSeek has reportedly asked Beijing to pressure Huawei to accelerate deliveries. Partial operation is framed as late 2027 or early 2028, not next quarter.
That delay is not comfort. It is a schedule. The 160,000 chips are also, per ZeroHedge's reading of Bloomberg, only one chunk of the gigawatt site. Applying the power ratio from xAI's 2024 Memphis cluster of 100,000 Nvidia H100s (~150 MW) to 160,000 Hopper-class chips yields about 250 MW — a fraction of a gigawatt. What fills the rest is unknown. A U.S. official has alleged DeepSeek trained on banned Nvidia Blackwell chips; Bloomberg has not independently verified that claim, and it should not be treated as established fact.
The training/inference split is the same split U.S. export controls accidentally drew. ZeroHedge notes DeepSeek spent months last year trying to train R2 on older Ascend chips with Huawei engineers on site; the Financial Times reported that effort failed to produce a successful training run, and DeepSeek went back to Nvidia for training. A hardware generation later, Friday's report says the division of labor is intact. Huawei takes the workload that books. Nvidia keeps the workload that crowns.
Beijing, not Washington, is the tighter gate on what Chinese labs may still buy from Nvidia. ZeroHedge reports the Trump administration cleared H200 sales to vetted Chinese buyers with a 25 percent U.S. Treasury cut, and that Beijing has approved only a fraction. Alibaba and Tencent still prefer Nvidia when they can get it. Huawei is winning Chinese inference on quota. That is a harder competitor than a price list.
Do the inventory this month, while the cluster is still a plan. Waiting until late 2027 to notice that inference moved is how a convenience default becomes a jurisdiction decision.
Huawei is not winning Chinese inference on price and performance. It is winning on quota. — ZeroHedge, on the Bloomberg-reported DeepSeek order
AgentsROI.ai is a managed AI services provider for owner-led SMEs. We do not sell Nvidia, Huawei, or DeepSeek. We match the model to the job — cloud, hybrid, or local — with a fallback when a lab, a chip, or a government quota changes the path.
Lead with Model Selection & Continuity Planning. DeepSeek's reported order is an extreme case of “right silicon, right workload, forced geography.” Most firms still have one default chatbot. The lesson is selection: which jobs may sit on a PRC-hosted inference API, which stay on a U.S. or EU path, and which never leave a machine you control.
Pair it with Managed AI Operations if staff already routed production traffic through a cheap DeepSeek endpoint because last quarter's price list looked clever. Routing fees, model IDs, and data-residency rules drift. Someone has to notice before the invoice or the regulator does.
DeepSeek's planned 160,000 Huawei Ascend 950DT chips, as reported by Bloomberg and relayed by ZeroHedge, are an inference bet under a quota. Nvidia still trains, for now. Huawei cannot ship overnight. Beijing is the gate. None of that is a reason to panic-rip a working tool on Monday. It is a reason to write down where your answers are generated, and what you do if that building changes owners.
If DeepSeek, Huawei-class silicon, or any single-country API is already in the stack without a named fallback, start with Model Selection & Continuity Planning. Book a no-pressure assessment when you want that map owned, not hoped for.
This article summarizes publicly reported information and is for general informational purposes only. It does not constitute legal, tax, financial, investment, security, or compliance advice. AgentsROI.ai is not a law firm, accounting firm, or registered investment adviser. Nothing here is a recommendation to buy, sell, or hold any security. Facts, pricing, statistics, and product capabilities cited here reflect the sources listed at the time of writing and may change. Readers should verify current information independently and consult qualified professionals regarding obligations specific to their industry, jurisdiction, and circumstances—including applicable New York State and New York City requirements. AgentsROI.ai may have commercial relationships with vendors mentioned; where material, such relationships are disclosed. Nothing in this article is an endorsement of any specific AI product, model, chip, or provider.
