For years, the premise of U.S. technology policy toward China has rested on an important foundational assumption: without access to Western semiconductors, China’s artificial intelligence ambitions would inevitably stall. The strategy of export controls was designed to exploit this vulnerability, cutting off the supply of advanced chips and the equipment needed to manufacture them.
A synthesis of recent developments across the Chinese technology sector suggests that this assumption is becoming increasingly fragile. The theoretical goal of Chinese semiconductor independence is rapidly materializing into a functional, albeit imperfect, domestic supply chain.
The most recent evidence of this shift is the commercialization of DDR5 memory modules by ChangXin Memory Technologies (CXMT). As we reported earlier today, CXMT has successfully brought its DDR5 chips to market, providing Chinese tech companies with a viable domestic alternative for a critical component of AI servers. While CXMT’s chips are one generation behind the cutting edge produced by Samsung and SK Hynix, their commercial availability fundamentally alters the strategic landscape.
The Ascend 950 Gains Traction
The progress in memory is mirrored by advancements in processing power. Huawei’s Ascend series of AI accelerators has emerged as the cornerstone of China’s domestic compute infrastructure, and the latest generation is now entering production at scale.
Huawei’s Ascend 950 chip entered mass production in March 2026, and Tencent Cloud has already deployed DeepSeek V4 preview services on the Ascend 950 platform, a significant commercial validation of the chip’s capabilities. As we reported a few weeks ago, Tencent Cloud’s deployment of the Ascend 950 represents a major step toward building a domestic AI compute stack that does not depend on Nvidia hardware.
Huawei has also confirmed that its Ascend SuperNode, based on the Ascend 950 AI chips, fully supports DeepSeek’s V4 models. The adoption of Huawei’s chips is not merely a matter of national policy preference; it is a practical response to the reality that Nvidia’s most powerful GPUs are entirely off-limits to Chinese companies, and that domestically available alternatives must be made to work.
This forced adaptation is creating a meaningful ecosystem effect. As more Chinese AI labs and cloud providers optimize their software stacks for Huawei hardware, the practical performance gap between domestic and foreign chips narrows, not because the hardware has caught up, but because the software has learned to extract more from what is available.
The Software-Hardware Co-Design Effect
The most significant consequence of the U.S. export controls may not be the temporary restriction of compute power, but the catalyst it has provided for software-hardware co-design in China. Because Chinese AI labs cannot rely on brute-force compute to train their models, they have been forced to innovate at the algorithmic level.
This dynamic was vividly illustrated by the success of DeepSeek, which achieved state-of-the-art performance by developing highly efficient training methodologies that required significantly less compute than comparable Western models. When these algorithmic efficiencies are combined with improving domestic hardware, the result is a surprisingly resilient AI ecosystem.
As noted in our coverage of the Quartz analysis published today, China’s AI race no longer looks like second place. Huawei’s AI chip revenues are projected to reach $12 billion this year, even as the company’s hardware remains at least two generations behind the American frontier. The combination of scale, software optimization, and algorithmic efficiency is producing a domestic AI compute ecosystem that is less capable than the American one in absolute terms, but far more capable than U.S. policymakers anticipated when the export controls were designed.
The Limits of Containment
The emergence of a viable domestic supply chain for AI hardware poses a profound challenge to U.S. policymakers. The strategy of containment relies on the existence of irreplaceable chokepoints. If China can successfully work around these chokepoints, through domestic innovation, algorithmic efficiency, or alternative architectures, the leverage of export controls diminishes significantly.
This does not mean that China has achieved parity. The U.S. still maintains a significant lead in the design and manufacture of the most advanced logic chips, and the ecosystem surrounding Nvidia remains the global standard. The recent blockbuster IPO of Cerebras Systems, which priced above its expected range to raise $5.55 billion, underscores the massive capital Western hardware startups have access to.
The gap is real, and it is not closing as fast as Beijing would like. But the trajectory is clear. China’s AI chip independence is no longer a distant, theoretical goal; it is a rapidly developing reality. As the domestic supply chain matures, the U.S. will increasingly need to rely on out-innovating its rival, rather than simply trying to restrict its access to technology. The era of easy containment is over; the era of intense, symmetric competition has begun.