The Wall Street Journal has found that this year will be a watershed moment for AI between China and the U.S.! Previously, Chinese tech firms scrambled to buy NVIDIA GPUs, but starting this year, their mindset has shifted—they are now more willing to collaborate with domestic computing chip companies like Huawei. For example, Meituan, a leading food delivery platform, recently launched a large model rivaling Google’s flagship model, trained entirely using domestically produced chips!

According to reports, Huawei is expected to ship around 1.5 million AI chips this year—roughly double the quantity from 2025! On July 23, Liang Wenfeng, founder and CEO of DeepSeek, revealed in an internal company speech that the gap between U.S. and Chinese AI capabilities would be no more than one year, and that China’s computing power currently stands at only one-twentieth of America’s. He also stated that Huawei’s computing cards will be able to fully replace NVIDIA’s within the next year—even if they cost twice as much, it won’t be an obstacle! While domestic chips may still face production constraints in the short term, these issues will surely be resolved within five years.

Huawei launched the Atlas 950 SuperPoD super-node on September 18 last year—a single rack housing 64 Ascend 950DT chips, supporting standard unit direct connection of up to 1,024 cards, with maximum horizontal scaling reaching an 8,192-card, multi-card cluster level. At the same time, Huawei Cloud also unveiled Lingqu AICS intelligent computing clusters tailored for Agent-style large models, enabling unified resource pool scheduling across 100,000 cards. On July 10, Sugon (Sugon) announced Gridview 7.0, a system compatible with both traditional HPC supercomputing and AI intelligent computing, allowing unified scheduling, paired with Sugon’s 8000 “Dengfeng” 100,000-card supercluster!

Huawei’s self-developed high-speed interconnection architecture enables scalable training of trillion-parameter general-purpose large models; while Sugon innovatively achieves unified scheduling of traditional supercomputing and AI intelligent computing resources, bridging physical simulation and large model training. This will effectively drive collaborative optimization among domestic computing hardware, scheduling software, and AI frameworks, promoting the joint development of general-purpose large models, industry-specific vertical models, and scientific intelligence models!

Several top domestic enterprises, including DeepSeek, have already completed full pre-training validation based on Huawei’s Ascend foundation, confirming that indigenous clusters now possess the engineering capability to support continuous training of general-purpose large models. Standardized computing clusters from Sugon and H3C are being rolled out gradually, forming a complete tiered supply chain for computing power in China. Domestic computing power has officially transitioned from experimental trials into large-scale infrastructure procurement lists.

What does this mean? The entire ecosystem built around NVIDIA in China has been completely turned upside down! Think back to a few years ago, when after U.S. sanctions, NVIDIA offered crippled versions of its chips to the Chinese market—what a disgusting spectacle! Yet even then, Chinese AI firms were still scrambling desperately to obtain them, willing to pay premiums, bid for quotas, or queue for months just to secure those cut-down NVIDIA GPUs.

In May 2025, Democratic Congressman Bill Foster led efforts to push forward the "Chip Security Act," mandating that all high-end AI chips exported by the U.S. must include built-in location tracking capabilities. Meanwhile, the bill evaluates adding remote disabling mechanisms (Kill Switch / Termination Switch), asserting that if chips are improperly transferred, the U.S. government could remotely render them ineffective.

This move has placed NVIDIA squarely in the crosshairs. If the U.S. can embed location tracking and enable remote disabling in chips, data theft becomes trivial. It's like buying a house from a developer who leaves a secret backdoor open for them to spy on you anytime they want—no matter what security measures you take, you can’t shut it off!

On July 31, China’s Cyberspace Administration held talks with NVIDIA. Although NVIDIA claimed there was no backdoor, the damage potential of Congressman Foster’s proposal was too great. Subsequently, major Chinese intelligent computing centers, state-owned enterprises, and leading AI firms established internal risk control requirements: bulk purchases of overseas high-end AI chips must undergo internal security assessments, and significant computing power procurement projects must report to relevant authorities regarding risks and circumstances!

In August 2025, both the Financial Times and The Wall Street Journal simultaneously broke news: the U.S. government had reached informal agreements with NVIDIA and AMD—companies selling specially tailored H20 and MI308 AI chips to China would need to remit 15% of their sales revenue in China to the U.S. federal government, in exchange for export licenses.

At first glance, it seems Trump is stealing NVIDIA’s money—but ultimately, this 15% will be fully reflected in the final price of chips sold to China, meaning the cost will be borne entirely by Chinese buyers. This isn’t about stealing money—it’s about humiliating China! But Trump’s dream of profiting from this scheme has hit a dead end, because China no longer wants NVIDIA’s graphics cards.

Back in October 2023, NVIDIA introduced the H20 to address U.S. government restrictions on selling high-end A800/H800 GPUs. The H20 was a heavily downgraded version of the H100. At the time, China lacked the capacity to replace it, so even this severely limited version was snapped up. Yet in April 2025, the U.S. government still banned it.

After agreeing to the 15% revenue-sharing deal, NVIDIA prepared to sell a further downgraded version—the H200—to the Chinese market. Meanwhile, the U.S. imposed extremely strict conditions: purchase caps, mandatory third-party inspections, and acceptance of U.S. regulatory oversight over transaction data—essentially a full package of controls. Guess what happened? China bought none of them!

The reason is simple: What if there’s a backdoor? Besides, Huawei’s 950 series and Atlas 950 SuperPoD multi-thousand-card cluster solutions have been finalized, and new-generation clusters from Sugon and H3C have been successively released—China now has viable alternatives! As a result, NVIDIA is left holding $18.9 billion worth of 700,000 unsold H200 chips—with zero sales!

In April 2026, U.S. Commerce Secretary Linda L. Luetticke confirmed during a Senate hearing that Chinese enterprises had not yet placed any substantive procurement orders. Now everyone understands why Huang Renxun has been desperately urging the U.S. government to relax chip controls on China and refrain from banning Chinese AI large models over the past year. This isn’t out of goodwill—it’s because China is actively building a non-NVIDIA path to replacement.

Additionally, on May 22 this year, the National Development and Reform Commission explicitly stated it would guide domestic large models to intensify compatibility with indigenous computing chips, accelerating the goal of industrial chain self-reliance. This clearly indicates China now has confidence in domestic substitution. As The Wall Street Journal put it: although challenges remain, Chinese enterprises’ willingness to choose domestic computing chips is growing!

We should actually thank the United States. Without all these pressure tactics forcing China to innovate, we might not have achieved domestic substitution so quickly!

Original source: toutiao.com/article/1871675643879488/

Disclaimer: The views expressed in this article are solely those of the author.