The new 7nm-class chip, reportedly in testing, signals a shift to domestic fabrication and aims to rival Nvidia’s China-compliant GPUs while maintaining CUDA compatibility.
Alibaba is reportedly developing a new AI chip designed for inference workloads. The AI chip is currently in testing and will be manufactured in China, according to a report.
Alibaba has not yet disclosed the specifications, such as the process node, power efficiency, or even the benchmark performance of the AI chip. But, unlike the previous chips that were designed for specific applications and workloads, the new chip will be capable of handling inference tasks, according to a WSJ report.
The move will also mark a major shift from Alibaba’s previous silicon strategy, which has so far focused on fabricating chips overseas at Taiwan Semiconductor Manufacturing Company (TSMC). Alibaba will reportedly manufacture the chip locally, although it is unclear who will manufacture the chip. Currently, the only domestic source of fabrication of AI chips in China is SMIC’s 7nm capacity. But it is already overcrowded by competing players like Huawei, Biren, and Cambricon.
Alibaba did not immediately respond to a request for comment.
“Alibaba’s older Hanguang 800 chip, which was 12nm, lagged behind Nvidia by 40–60% in inference throughput,” said Pareek Jain, CEO at EIIRTrend & Pareekh Consulting. “The new chip, expected to be 7nm, is designed to be more flexible, targeting broader workloads compared to Hanguang 800, and is expected to rival Nvidia’s H800/H20, but not the flagship H100 or Blackwell chips. The chip would be versatile, optimized for inference workloads like recommendation systems and language processing.”
Danish Faruqui, CEO at Fab Economics, said Alibaba could leverage mathematics to overcome physics at the chip level. “For instance, for Huawei’s Cloud Matrix 384 system, Huawei has five times as many Ascends 910 C chips in each system, which more than offsets each GPU being only one-third the performance of an Nvidia Blackwell B200 to deliver 50% more system-level PFLOP performance as compared to Nvidia GB200 NVL72 system.”
Experts say Alibaba’s AI Chip could also address the shortcomings of Huawei’s 910B and 910C AI chips, as the new chip is expected to be compatible with Nvidia’s CUDA and PyTorch ecosystems.
“Nvidia’s dominance isn’t just in its hardware; it is in its robust software ecosystem, particularly CUDA. A seamless compatibility layer would allow developers to easily port their existing code, bypassing the steep learning curve associated with new, fragmented hyperscalar software stacks from domestic rivals like Huawei’s MindSpore or Baidu’s PaddlePaddle. The software development and optimization on custom chips are the major bottleneck,” said Neil Shah, vice president at Counterpoint Research.
Geopolitical pressures, domestic strategy
Alibaba has been using Nvidia’s AI chips for a long time, but increasing US export restrictions over the last few years have prompted Alibaba to accelerate its domestic semiconductor strategy.
In October 2022, the US imposed restrictions on advanced semiconductors exported to China. To this, Nvidia and AMD responded by scaling down the GPU versions, tailored to stay within export compliance. But later, even these chips faced licensing barriers.
In April this year, under President Donald Trump’s administration, Nvidia was informed about securing an export license to sell its H20 processor in China. Barely three months later, the US administration announced the resumption of the sales of H20 AI chips to China as part of a rare earth critical minerals deal. Later, China’s Cyberspace Administration (CAC) summoned Nvidia representatives regarding serious security issues related to potential tracking and remote shutdown capabilities embedded in the H20 chips.
“In 2023, Alibaba Chairman Joe Tsai, after blaming uncertainties created by US export restrictions on advanced computing chips for halting Alibaba’s cloud unit’s spinoff, made a bold vision about how it will handle the setback. He focused on developing a sustainable growth model based on emerging AI-driven demand for networked and highly scaled cloud computing services,” Faruqui said. The new AI chip by Alibaba, he said, is the outcome of this sustainable growth model to counter US sanctions.
China’s AI race
Alibaba’s reported development of a new AI inference chip marks another milestone in China’s race to build a self-reliant semiconductor ecosystem. Chinese firms like Huawei and Baidu are also developing their own AI chips to cut reliance on US companies.
“Alibaba’s investment, combined with efforts from Huawei and Baidu, is fostering the rapid growth of China’s AI chipmaking ecosystem. With more capable local chips, Chinese enterprises are less reliant on global clouds such as AWS, Azure, keeping data within Chinese regulatory frameworks,” Jain said.
To start with, Alibaba’s AI chip will most likely be produced for in-house data centers. Once the production is scaled, the chip can be used outside China to reduce reliance on Western clouds for some workloads, especially where Alibaba Cloud already has footprints, including Japan, South Korea, Southeast Asia, the Middle East, and Latin America.
While Alibaba’s long-term intention may be to build a CUDA-compatible chip, at this point, it could be a tactical move to gain leverage in negotiations with Nvidia. According to Shah, as China faces supply shortages of Nvidia’s H20 chips, Alibaba’s chip, specifically designed for the Chinese market under the US export controls, could be signalling its ability to produce a viable domestic alternative. This could influence the pricing and availability of Nvidia’s products in the region.
