cd /news/ai-infrastructure/deepseek-bets-2-6-billion-on-huawei-… · home topics ai-infrastructure article
[ARTICLE · art-122926] src=startupfortune.com ↗ pub= topic=ai-infrastructure verified=true sentiment=· neutral

DeepSeek Bets $2.6 Billion on Huawei Chips It May Wait a Year to Get

DeepSeek has committed to roughly 160,000 Huawei Ascend 950DT chips for a new gigawatt-scale data center in Ulanqab, Inner Mongolia, an order valued at about $2.6 billion that could take over a year to fill due to Huawei's production constraints. The Hangzhou AI lab plans to use the chips for inference, not training, as it continues to train models on Nvidia hardware, according to Bloomberg. Huawei's output of the 950DT is expected to be capped at 200,000 to 300,000 units in 2026 due to memory shortages, potentially delaying full operation of the data center until 2027 or later.

read4 min views2 publishedSep 8, 2026
DeepSeek Bets $2.6 Billion on Huawei Chips It May Wait a Year to Get
Image: Startupfortune (auto-discovered)

DeepSeek has committed to roughly 160,000 Huawei Ascend 950DT chips for a new data center in Inner Mongolia, a bet on Chinese silicon so large that Huawei's own factories may need more than a year to fill it.

Bloomberg reported that the Hangzhou AI lab is planning a gigawatt-scale facility in Ulanqab, Inner Mongolia, built around roughly 160,000 Ascend 950DT accelerators. At a street price near 111,000 yuan, about $16,000, per chip, the order carries a face value of roughly $2.6 billion. That's a serious bet on Chinese silicon. If it lands as described, Bloomberg noted, it would be one of the largest known clusters of Huawei AI chips ever assembled.

DeepSeek wants the chips for inference, not training. That distinction matters more than it sounds. The company is still training its models on Nvidia hardware, according to Bloomberg's reporting, because Huawei's silicon and software stack aren't yet trusted for the heaviest workloads. Ascend chips would instead run the finished models, answering the millions of queries DeepSeek's apps and API field every day. Nvidia's chips remain scarce in China after Washington's export controls and Beijing's own discouragement of foreign silicon, so a Chinese lab building at gigawatt scale has one real option left, and it's Huawei.

Huawei's Delivery Problem #

Here's the catch. Huawei can't make the chips fast enough. Component shortages, chiefly in advanced memory, are expected to cap 2026 output of the 950DT at somewhere between 200,000 and 300,000 units, according to industry estimates cited by TechNode and TrendForce. DeepSeek's order alone would eat most of a year's total supply. Filling it could stretch well past a year, which means a data center meant to serve today's users might not run at full capacity until sometime in 2027 or later.

The bottleneck isn't the chip design. It's the memory sitting next to it. High bandwidth memory accounts for roughly half the cost of an AI accelerator, and China's leading HBM producer, CXMT, is still running sample batches of HBM3 rather than mass production. Packaging capacity is tight too, with new lines from JCET and Tongfu not expected to add meaningful volume until 2027. Huawei can design a chip that rivals Nvidia's H200 on paper. Getting enough of them out the door is a different problem entirely.

Huawei Pours a Quarter of Its Revenue Into R&D as Profit Falls 36 Percent Huawei's first-half 2026 revenue rose 9.55% to 467.8 billion yuan, but net profit fell 36% as R&D spending jumped 25% to 121.4 billion yuan, over a quarter of revenue. The push is aimed at making Huawei's Ascend AI chips a real alternative to Nvidia for China's AI startups. - Huawei R&D spending strategy for AI chip development - Chinese startups reducing dependence on Nvidia GPU alternatives

A Market Already Tipping Away From Nvidia #

The scale of the shift is stark. Huawei expects roughly $12 billion in AI chip revenue this year, up 60% from $7.5 billion in 2025, driven largely by the Ascend 950 family after the 950PR entered mass production in March. Bernstein forecasts Nvidia's share of China's AI accelerator market falling from 40% to 8% in 2026, while Huawei's Ascend line moves toward controlling roughly half to 60% of the domestic market. Nvidia CEO Jensen Huang has already said publicly that the company's China market share sits at zero.

That's the real story behind DeepSeek's order. Washington built the export ban to slow China's AI progress. Instead, it handed Huawei a captive customer base with nowhere else to buy, and Chinese labs are now placing chip orders the size of small countries' defense budgets on hardware that still can't ship on time. The lockout didn't stop China's AI buildout. It just moved the bottleneck from export licenses to wafer allocation and HBM yield, which are Huawei's problems to solve now, not Washington's.

There's a longer-term wrinkle too. DeepSeek is reportedly in talks with a state-backed chip fund to lead its next funding round, according to a Financial Times report picked up by Bloomberg in May. A company that trains on Nvidia but runs on Huawei, backed by government chip money, is exactly the kind of vertically aligned national champion Beijing has been trying to engineer since the first export controls hit in 2022. It's taken four years, a lot of subsidy, and a chip shortage of its own making to get there.

None of this solves DeepSeek's near-term problem. The Ulanqab site can only run as fast as Huawei can ship, and right now that's the slowest part of the entire chain. A $2.6 billion order sitting in a production queue is still just a queue.

Also read: China's Exports Jumped 25% in August as AI Chip Demand Drove the SurgeHere's Why AI Agent Retry Logic Quietly Multiplies Your API CostsRobots Marched On Poland's Digital Ministry To Demand AI Regulation

Join the discussion #

Open in the community → Almost there. Sign in and your reply posts straight away.

── more in #ai-infrastructure 4 stories · sorted by recency
── more on @deepseek 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/deepseek-bets-2-6-bi…] indexed:0 read:4min 2026-09-08 ·