DeepSeek's 160,000 Huawei Order Is A Real Threat To Nvidia In 2027

Roughly 350 kilometers northwest of Beijing, in the city of Ulanqab, DeepSeek is constructing a data center measured in gigawatts. Running at full capacity, the site would draw enough power to supply 750,000 homes. On Friday, Bloomberg reported that DeepSeek intends to fill a portion of this massive footprint with at least 160,000 of Huawei's newest AI processors, the Ascend 950DT. Once operational, it will stand as the largest known cluster of Chinese-made AI silicon in the world.

Yet, Huawei's processors have a notoriously poor track record when it comes to the heavy lifting of training AI models - a limitation DeepSeek experienced firsthand with the previous chip generation. Instead of building models from scratch, the Huawei chips will be relegated to inference: running DeepSeek's finished models to generate answers for end users. That is not the consolation for Nvidia it sounds like. Inference is the larger pool of compute, the faster-growing one, and increasingly the one that pays - and Huawei has just taken the biggest inference order in China.

What The Chips Won't Do

Every AI model leads a double life. The first phase is training: a grueling, months-long process where tens of thousands of chips work in perfect synchronization, digesting vast swaths of the internet to learn how to predict the next word. The second phase is inference, which encompasses everything that follows - answering chatbot queries, writing code, or summarizing documents. Training determines who holds the technological crown, but inference drives the sheer volume of compute and, increasingly, the revenue.

When Huawei laid out its roadmap last September, it positioned the 950DT as a dual-threat capable of both training and inference. However, Bloomberg's sources indicate DeepSeek has zero intention of using it to train. Last year, buckling under pressure from Chinese authorities to adopt domestic hardware, DeepSeek spent months attempting to train its R2 model on older Ascend chips with on-site Huawei engineers. According to the Financial Times, the effort failed to yield a single successful training run. DeepSeek ultimately reverted to Nvidia for training, keeping Huawei strictly for inference. A year later, and a full hardware generation newer, Friday's report confirms that division of labor remains intact.

This means China's premier domestic chip has successfully captured the inference market, but remains locked out of the training arena. Ironically, this is precisely the boundary U.S. export controls were designed to enforce. Four years of restrictions were intended to block China from training frontier models; Washington never actually set out to prevent them from running finished ones.

What Nvidia keeps in China, meanwhile, barely counts. Its remaining Chinese training demand is fueled by rented overseas compute and alleged smuggling routes, and neither pipeline shows up on Nvidia's balance sheet as legitimate Chinese revenue. The segment Huawei has taken is the one that books.

Huawei Can't Fill It

One issue - Huawei simply cannot manufacture these chips fast enough. According to Bloomberg, severe shortages of top-tier memory will bottleneck 950DT production to the low hundreds of thousands this year. Huawei is actively juggling DeepSeek's massive order against other domestic clients and a fledgling export initiative, meaning it could take well over a year just to fulfill this single contract. DeepSeek has even petitioned Beijing to pressure Huawei into accelerating deliveries.

The root of this bottleneck lies in a component often overlooked in AI coverage. An AI accelerator is only as fast as the high-bandwidth memory (HBM) stacked alongside it. The companies that matter here are South Korea's SK Hynix and Samsung, and the U.S.'s Micron. When Washington severed China's access to advanced HBM in December 2024, Huawei had to pivot to developing its own. The 950 series is the first Ascend generation to feature this in-house memory, and Bloomberg's sources confirm it is exactly what is crippling output.

China's national champion is having its flagship chip rationed by the state, while the nation's most prominent AI lab waits in line begging for a larger allocation. To put this in perspective, Bloomberg reported last year that Huawei's entire 2026 production target is roughly 1.6 million Ascend dies - enough for about 600,000 of the older 910C chips and a scattered mix of newer silicon. Furthermore, the 950DT is generally considered to be on par with Nvidia's Hopper generation (the H100s that fueled the 2023 AI boom). By the time DeepSeek's cluster is fully operational in late 2027, more than a year from now, Nvidia will be two full generations ahead in training - a race Huawei is no longer running.

A Fraction Of The Site

The bigger open question is scale. Bloomberg emphasizes that 160,000 950DTs will constitute "only one chunk" of the gigawatt-scale facility in Ulanqab. What will occupy the rest of the site remains a mystery.

When Elon Musk's xAI activated its initial 100,000 Nvidia H100s in Memphis in 2024, the cluster required roughly 150 megawatts of power. Applying that ratio to 160,000 Hopper-class Huawei chips yields a footprint of about 250 megawatts - roughly a quarter of DeepSeek's planned gigawatt, or perhaps a third if the 950DT proves more power-hungry than Nvidia's hardware (a metric Huawei keeps under wraps). Regardless of the exact math, the vast majority of the gigawatt site is unaccounted for.

Washington has its own theory regarding the missing hardware. In February, a senior Trump administration official claimed that DeepSeek trained its latest model on Nvidia's Blackwell chips - current-generation hardware strictly banned from entering China - and suggested these chips are likely housed at the Inner Mongolia facility. "We're not shipping Blackwells to China," the official stated, though they declined to elaborate on how DeepSeek procured them. DeepSeek did not respond to Reuters' requests for comment at the time. While Washington has repeatedly accused both DeepSeek and Moonshot of relying on smuggled Nvidia silicon, Bloomberg notes it hasn't independently verified these claims.

The Huawei order is the narrative Beijing wants public, while the Blackwell smuggling allegation is the narrative Washington wants public - and neither explains the true scale of the Ulanqab facility. 

Beijing Is The One Saying No

In December, the Trump administration cleared Nvidia to sell its H200 chips - the crown jewel of the Hopper generation - to vetted Chinese buyers, provided the U.S. Treasury received a 25 percent cut. The Commerce Department formalized this rule in January, and Bloomberg reports the administration was ready to authorize the export of up to a million chips. Beijing has approved a fraction of that volume.

Given the options, Chinese tech giants still prefer Nvidia. Both Alibaba and Tencent placed massive orders the moment U.S. restrictions relaxed, and Chinese firms continue to rent top-tier Nvidia compute by the hour from unregulated overseas data centers. The true blockade is being orchestrated in Beijing - as the Chinese government is aggressively weaning its domestic labs off foreign silicon, restricting Nvidia imports, and forcefully pivoting the industry toward Huawei.

So - Huawei is not winning Chinese inference on price and performance. It is winning on quota - which makes the threat to Nvidia a policy instrument rather than a product, and a far harder thing to compete against. 

Nvidia CEO Jensen Huang has warned for years that U.S. export controls would inevitably forge a self-sufficient Chinese chip industry. His prophecy is now being fulfilled - by the Chinese government itself. Regardless of what Washington permits, Beijing acts as the ultimate gatekeeper, utilizing import quotas to force its most critical AI labs to adopt Huawei hardware for their fastest-growing workloads.