Industry executives confirm that these tech titans have initiated testing of alternative semiconductors to satisfy the escalating internal and client-driven demand for AI capabilities. The urgency of this pivot intensified last month when the Trump administration, building on existing Biden-era curbs, further tightened restrictions on sales of Nvidia's H20 chip – a modified version designed to comply with earlier US regulations.
Insiders familiar with the situation reveal that existing stockpiles of Nvidia chips held by Chinese tech groups are only sufficient to sustain current AI development until early next year. This limited window underscores the critical need for immediate contingency planning. The typical lead time for new chip orders ranges from three to six months, and uncertainty looms over when, or even if, Nvidia will be able to offer a new processor for the Chinese market that not only adheres to the tougher Trump-era export rules but also remains competitive against burgeoning local alternatives.
A Strategic Shift Towards Domestic Solutions
Chinese tech leaders are vocal about their commitment to exploring diversified chip solutions. Shen Dou, head of Baidu's AI cloud group, recently informed analysts that the company possesses a range of chip options, particularly for "inference processing" – the application of trained AI models – to replace Nvidia's offerings. "We believe that over time, domestically developed self-sufficient chips, along with increasingly efficient homegrown software stacks, will jointly form a strong foundation for long-term innovation in China’s AI ecosystem," Shen asserted.
Similarly, Alibaba chief Eddie Wu, on a recent earnings call, stated, "We are actively exploring diversified solutions to meet rising customer demand." Tencent president Martin Lau echoed this sentiment, explaining his company's focus on optimizing chip utilization while also considering alternative products. Lau assured analysts that Tencent possesses "enough high-end chips to continue our training models for a few more generations going forward," adding that the company could "potentially make use of other chips" to address growing inference needs.
Huawei's Ascend Chips Emerge as a Key Contender
The tightening US restrictions have inadvertently fueled a surge in domestic innovation within China's high-end AI chip sector. A notable beneficiary of this trend is Huawei's Ascend chip series, which a think-tank affiliated with China’s state security ministry, the China Institutes of Contemporary International Relations, highlighted as a prime example of independent innovation. The think-tank noted in a social media post that "Domestic entities in China have already begun large-scale procurement and use of Ascend chips."
Historically, the primary purchasers of Huawei's chips have been state-owned enterprises like China Mobile, along with companies operating in sensitive sectors such as defense, healthcare, and finance. However, the current landscape suggests that a much broader spectrum of domestic tech companies will now vie for access to Huawei's chips, solidifying its position as a national champion in the semiconductor space.
Despite the growing interest, companies exploring Huawei as an alternative have largely remained discreet about their testing of Ascend chips. This caution stems from recent guidance issued by Washington on export controls, which cautioned that the use of these chips "anywhere in the world" could expose companies to criminal penalties.
Challenges and the Road Ahead
While the move towards domestic chips offers long-term strategic advantages for China, it is not without significant hurdles. Analysts at GF Securities anticipate that Nvidia might begin producing its next generation of China-compliant chips for export as early as July. However, reports suggest that these new processors, though based on Nvidia's advanced Blackwell architecture, would likely lack high-bandwidth memory (HBM), a crucial component for efficient processing of large datasets. Furthermore, key details such as the inclusion of Nvidia's high-speed interconnect NVLink remain ambiguous. Nvidia CEO Jensen Huang, on a recent earnings call, acknowledged the limited options for a new China product, stating, "We don’t have anything at the moment."
A major challenge for Chinese tech companies lies in the substantial costs associated with migrating their existing systems from Nvidia chips to domestic alternatives. Porting training code, originally developed using Nvidia's proprietary CUDA software framework, to Huawei's CANN ecosystem is an extremely time-consuming endeavor. This process often necessitates significant support from Huawei engineers for debugging, optimization, and addressing various compatibility issues. One leading Chinese tech executive estimated that a full transition to Huawei's chips could lead to approximately three months of disruption in AI-related development.
In response to these complexities, most companies are considering a hybrid approach. This strategy involves continuing to run AI training on their existing Nvidia chips while simultaneously utilizing local processors for inference tasks, where demand is rapidly escalating due to the broader adoption of AI applications across China.
While Huawei is actively working to expand production capacity with its partners and is even launching its own fabrication plant, the current supply of its chips is struggling to keep pace with soaring demand. Beyond Huawei, other Chinese chipmakers, including Cambricon and Hygon, are also seeing their products being tested by the tech giants. Furthermore, both Baidu and Alibaba have been proactively developing their own in-house processors to meet the burgeoning demand for AI capabilities within their respective operations. The shift is underway, a complex and critical undertaking that will undoubtedly shape the future of AI development in China.