DeepSeek Bets $2.6B on Huawei as NVIDIA 'Cedes' China

160,000 Chips, $2.6 Billion — This Time DeepSeek Isn’t Buying From NVIDIA

According to Bloomberg, DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT chips at a data center in Ulanqab, Inner Mongolia, with the facility designed for a full 1GW (gigawatt) of capacity. Each 950DT ships with 144GB of HBM memory and 4.0 TB/s of memory bandwidth, priced at roughly $16,000 per unit on the street market — putting the total order value at an estimated $2.6 billion. These chips are meant for inference (the actual computation of responding to user requests), not training — DeepSeek still relies on NVIDIA chips to train its models and isn’t planning to switch that over yet.

Huawei can’t even make that many chips fast enough

The problem is that Huawei itself can’t currently produce chips at this volume. Reports indicate that component shortages, particularly in high-bandwidth memory (HBM), will cap Huawei’s total 2026 production of the 950DT at roughly 200,000 to 300,000 units — meaning fulfilling DeepSeek’s order alone could take more than a year. DeepSeek has reportedly asked Beijing for help securing priority allocation amid the shortage — a telling sign that even a company at DeepSeek’s scale has to get in line for Huawei’s output. The 950DT chip itself isn’t expected to reach commercial launch until Q4 2026, and the data center’s partial capacity is targeted to come online sometime between late 2027 and early 2028.

How big is this order, really?

160,000 chips sounds abstract on its own, but the comparisons make it concrete: it’s roughly 30% of Huawei’s publicly disclosed 520,000-plus-unit “SuperCluster” specification, and it dwarfs the cluster Huawei deployed in Shenzhen this past March, which used just 10,000 Ascend 910C chips — this order is 16 times that size. If it goes through as planned, it would be one of the largest known clusters of Huawei AI chips anywhere, sited in Ulanqab, roughly 350km from Beijing.

Even NVIDIA’s CEO admits China’s basically been ceded to Huawei

This order comes against the backdrop of NVIDIA’s rapidly deteriorating position in China’s AI chip market. Market forecasts cited in the reporting show NVIDIA’s share of China’s AI chip market plunging from 40% in 2025 to just 8% in 2026, while Huawei’s share is projected to climb to roughly 50%, with Huawei’s AI chip revenue expected to grow from $7.5 billion in 2025 to around $12 billion in 2026. NVIDIA CEO Jensen Huang himself admitted as much in May 2026, acknowledging that NVIDIA has “largely conceded” the Chinese market to Huawei.

Policy is blowing in the same direction

There’s a policy tailwind behind all this. China’s Cyberspace Administration was already discouraging companies from buying NVIDIA chips as of September 2025, and official guidelines from November 2025 went further, requiring state-funded data centers to prioritize domestic chips. On the US side, there’s been a subtle loosening — in January 2026, export review policy for H200 chips to China shifted from a “presumption of denial” to “case-by-case review,” reopening a narrow window. But for a Chinese AI company like DeepSeek — already betting on Huawei with one hand while still training on NVIDIA with the other — that narrow window probably isn’t enough to reverse an industry-wide shift toward domestic chips that’s already well underway.

DeepSeek can afford this bet

Part of why DeepSeek can make a wager this size comes down to its own finances. The company closed a funding round of 50 billion yuan (roughly $7.4 billion) in June 2026, pushing its valuation to an estimated 350.9 billion yuan (about $51.8 billion). Against that backdrop, a $2.6 billion chip order is a large but manageable strategic investment.

What matters more is the cost structure. The reporting cites NVIDIA CEO Jensen Huang’s own estimate that building an equivalent 1GW data center with NVIDIA’s top-tier chips would cost roughly $50 billion — though the reporting also notes that estimate is based on US construction costs, and data centers in China are generally cheaper to build, so the real-world gap is probably narrower than $50 billion versus $2.6 billion suggests. Even discounted for that, it may still be part of why DeepSeek is willing to risk waiting over a year for the full chip delivery to bet on Huawei: it’s not just a wager that chip supply will eventually catch up to its ambitions, but that the cost structure itself is an incentive too.