The competitive dynamics for artificial intelligence hardware in China took a sharp turn effective April 15. On that date, the U.S. Department of Commerce imposed an “indefinite license” requirement for the export of Nvidia’s H20 and AMD’s MI308 AI accelerators to mainland China, Hong Kong, and Macau.
Citing national security risks, particularly potential misuse in Chinese supercomputing endeavors, this action effectively blocked sales of the H20 – Nvidia’s main chip previously designed to comply with earlier US export rules that had already barred more powerful GPUs like the H100 from China. Affirming the administration’s position, Commerce Department spokesman Benno Kass stated the department is committed “to acting on the president’s directive to safeguard our national and economic security.”
The restriction delivered an immediate financial hit to Nvidia. The company disclosed it was taking a US$5.5 billion charge against quarterly revenue, linking it directly to unsellable H20 inventory and associated purchase commitments. Nvidia CEO Jensen Huang responded quickly, traveling to Beijing on April 17. During his visit, he met China’s head of international trade promotion, Ren Hongbin, and reportedly stated Nvidia hoped “to continue to cooperate with China.”
Meetings also occurred with DeepSeek founder Liang Wenfeng, potentially to discuss alternative, compliant chip possibilities. This activity followed earlier, conflicting reports around April 10th suggesting the Trump administration might have briefly considered pausing the H20 restrictions.
Huawei Ascends into the Void
Stepping directly into the market space vacated by the H20, Huawei Technologies is set to commence mass shipments of its Ascend 910C artificial intelligence processor to Chinese customers as early as May, according to sources familiar with the company’s plans cited by Reuters.
Some initial deliveries have already been made, the sources added. The Ascend 910C is a Graphics Processing Unit (GPU), a type of accelerator optimized for the parallel processing required by demanding AI workloads. Its imminent availability positions it as a key domestic hardware option for Chinese AI companies. Paul Triolo, a partner at consulting firm Albright Stonebridge Group, observed that the US curbs “will mean that Huawei’s Ascend 910C GPU will now become the hardware of choice for (Chinese) AI model developers and for deploying inference capacity.”
Sources describe the Ascend 910C not as a ground-up new design, but as an architectural advancement. It reportedly achieves performance comparable to Nvidia’s powerful H100 chip – itself already banned for sale in China since 2022 – by integrating two of Huawei’s previous-generation 910B processors into a single package.
This is accomplished using advanced chiplet integration techniques, where smaller, specialized silicon dies are combined. This approach effectively doubles the computational power and memory capacity compared to the standalone 910B.
The underlying 910C accelerator chiplet offers baseline performance around 780 TFLOPS using the BF16 numerical format (BFloat16, a format optimized for AI that balances computational speed and numerical precision) and includes specific enhancements for handling diverse AI data types. Huawei had already provided samples of the 910C to technology firms late last year and started accepting orders, according to previous reports.
Navigating Manufacturing and Sanctions
Producing chips like the Ascend 910C under existing US sanctions presents significant hurdles. China’s premier semiconductor foundry, SMIC, is reportedly manufacturing some core components using its N+2 7nm process technology. While representing a domestic capability, sources previously indicated SMIC faces challenges achieving high production yields with this process.
Adding another layer of complexity, Reuters reported, based on multiple sources, that at least some 910C GPUs incorporate semiconductors originally manufactured by Taiwan’s TSMC. These were allegedly produced for Sophgo, a China-based chip design firm, and potentially acquired by Huawei through intermediary channels.
This purported TSMC-Sophgo connection has attracted US regulatory attention. The Commerce Department is reportedly investigating TSMC’s work for Sophgo, following the discovery of a similar chip in an earlier Huawei 910B processor.
Research cited by Reuters suggested TSMC might have fabricated nearly three million chips matching Sophgo’s design, potentially exposing the foundry to substantial penalties. Responding to Reuters, Huawei reiterated that it has not used TSMC-made Sophgo chips.
TSMC said it complies with all regulatory requirements and has not supplied Huawei directly since mid-September 2020. Sophgo did not provide immediate comment. Similar indirect supply routes involving distributors like CoAsia Electronics are suggested for obtaining essential High Bandwidth Memory (HBM) – critical stacked memory needed for AI performance – potentially originating from Samsung.
Broader Context: Tech Rivalry and Self-Sufficiency
This development highlights the intensifying technological competition between the US and China. Washington’s export controls seek to limit China’s access to advanced technology, particularly for potential military applications.
However, these measures appear to simultaneously spur China’s ambitions for semiconductor self-reliance, an effort backed by large state investments such as the $47.5 billion “Big Fund.”
Even prior to the definitive H20 ban, Chinese AI firms were reportedly stockpiling the Nvidia chips, anticipating restrictions. The Ascend 910C’s rollout provides a powerful domestic alternative precisely when a key foreign competitor is sidelined, lending credence to analyst Patrick Moorhead’s earlier assessment: “Chinese companies are just going to switch to Huawei.” The situation also opens doors for other Chinese GPU startups, such as Moore Threads and Iluvatar CoreX, aiming to capture market share.
