China’s artificial intelligence (AI) development could face slower progress as new US export restrictions targeting remote access to advanced semiconductors threaten to cut off a key channel through which Chinese technology companies have been accessing high-performance computing capacity overseas, Maybank Investment Bank Research said on Thursday.
In a research note, Maybank said the latest restrictions would shift China’s AI constraint from access to computing capacity towards the availability of domestic chip supply.
The US Commerce Department’s Bureau of Industry and Security (BIS) on Nov 16 released an interim final rule titled “Addressing Remote Access to Advanced Semiconductors”, which effectively closes the remaining route for Chinese AI companies to access frontier US chips by renting computing capacity from data centers outside China.
The new rule introduces a worldwide licensing requirement under Section 744.24 of the US Export Administration Regulations, covering advanced chips and servers supplied to data centers in any country that provide more than 12 million aggregate tera operations per second (TPP) of computing capacity to companies headquartered in China or Macau, or whose ultimate parent is based in either jurisdiction.
The regulation also places greater responsibility on data-center operators and chip suppliers to verify the ownership of their customers.
Where a hosting provider cannot or refuses to provide customer ownership checks, BIS could treat this as a red flag and presume that suppliers have knowledge of the restricted end-use.
The rule also closes certain attestation routes previously available to Chinese parented chip designers with fewer than 30 billion transistors when engaging foreign foundries and packaging facilities.
However, the existing January 2026 channel allowing licensed exports of Nvidia’s H200-class chips into China remains open.
Citing Nvidia’s financial disclosures, Maybank noted that China-related Hopper shipments accounted for less than 1 percent of its data-center revenue in the second quarter of financial year 2027, while its third-quarter outlook assumed no revenue from China data-center computing.
It also highlighted that about 600,000 H200 chips were reportedly awaiting Beijing’s approval in Hong Kong to enter mainland China, with major Chinese technology platforms clearing only about 10,000 chips each.
With US chip supplies into China restricted and offshore computing rentals now subject to licensing, Maybank said Chinese AI developers unable to secure approval would have to rely more heavily on domestic semiconductor infrastructure, which is already facing significant capacity constraints.
According to Epoch AI estimates cited by Maybank, Huawei’s computing output in 2026 is expected to remain below 4 percent of Nvidia’s, while Semiconductor Manufacturing International Corp (SMIC) operated at more than 97 percent capacity utilization in the third quarter of 2026.
Huawei executives have also indicated that domestic demand for computing capacity may not be fully met until 2030.
Maybank cautioned against assuming that the restrictions would immediately help China narrow its technological gap with the US by accelerating domestic innovation.
While the restrictions could encourage greater software optimization and hardware-software co-design, they would not resolve shortages of high-bandwidth memory (HBM), advanced wafer capacity or power efficiency.
The research house said the likely outcome of tighter enforcement would be slower frontier AI training in China, as research laboratories lose access to the additional computing capacity previously obtained through overseas data centers.
It also expects a widening divide between Chinese technology companies, with vertically integrated hyper scalers that own their chips and have stronger financial resources better positioned than AI laboratories that depend on rented computing capacity.
China’s domestic chips remain significantly less efficient than Nvidia’s advanced processors.
Epoch AI estimates show Huawei’s Ascend 950DT delivers one petaflop per second (PFLOP/s) in FP8 performance, about half that of Nvidia’s 2022 H100 chip, while its FP4 performance trails Nvidia’s B300 by 6.75 times.
In terms of cost efficiency, the Ascend 950DT delivers 0.05 tera operations per second per US dollar, compared with 0.34 for Nvidia’s B300.
Maybank said China’s leading AI models had scaled their training computing capacity at about three times annually since US export controls began tightening in late 2021, compared with five times globally.
Despite the slower hardware expansion, Chinese models have lagged the US frontier by only about seven months on average, partly because Chinese companies have continued accessing overseas Nvidia computing clusters, including those in Southeast Asia.
The latest BIS rules seek to close this channel by linking restrictions to the headquarters of customers rather than the country hosting the data centre. They also aggregate computing capacity across digitally connected sites, preventing companies from circumventing the threshold by splitting clusters across different regions.
HBM shortage remains the biggest bottleneck
Maybank identified HBM availability as the primary constraint on China’s ability to expand domestic AI computing capacity.
Based on Epoch AI estimates, Nvidia has a 73-fold advantage over Huawei in usable domestic HBM supply in 2026, contributing to Nvidia producing nearly 200 times more AI computing capacity than Huawei can support using domestically sourced HBM.
Although the gap in HBM availability is expected to narrow to 11 times by 2028, Nvidia’s computing advantage per gigabyte of memory is projected to widen from three times to six times.
China’s domestic HBM supply is estimated to support only about 240,000 Ascend 950DT chips this year, forcing another 750,000 Ascend 950PR chips to rely on non-HBM memory, which offers lower bandwidth of 1.64 terabytes per second and is more suitable for inference prefill workloads than actual model training.
Demand is already exceeding available supply. DeepSeek’s order for 160,000 Ascend 950DT chips faces a delivery wait of more than a year, while Alibaba chief executive Eddie Wu has said the proportion of proprietary chips deployed on Alibaba Cloud remains relatively low due to limited domestic fabrication capacity.
Domestic accelerator chip prices have also risen by 20% to 30% amid HBM shortages, Maybank noted, adding that stronger demand would likely lengthen delivery queues and push prices higher rather than immediately increase production capacity.
Hyperscalers better positioned
Maybank said Chinese AI companies would increasingly have to focus on software optimisation and inference workloads, where domestic chips are already being deployed.
Z.AI, for instance, is serving inference workloads on more than 100,000 domestic chips, reducing unit token costs by 80% this year. DeepSeek has also ported its core computing kernels to Huawei’s Ascend 950 chips and reduced memory requirements per token to 890 bytes.
However, Maybank noted that these developments predated the latest BIS restrictions, while former White House official Ben Buchanan had said no capable Chinese AI system had yet been fully trained on Chinese chips.
The research house said the medium-term outlook would depend on whether the US licensing process results in routine denials for generative AI training and whether displaced computing workloads can migrate to alternative infrastructure.
If licensing restrictions are enforced tightly, Alibaba, Tencent, ByteDance and Baidu are likely to be better positioned given their financial resources, proprietary chip development and ability to fund substantial AI infrastructure investments.
DeepSeek could also be relatively more resilient given its close partnership with Huawei in developing computing kernels.
In contrast, AI laboratories that rely heavily on rented computing capacity and lack large consumer or advertising businesses to support research and infrastructure spending, including Moonshot, Z.AI and MiniMax, could face greater pressure as they compete for limited domestic computing resources largely controlled by China’s major technology groups.
New US chip curbs put Southeast Asian data center projects under greater scrutiny – BMI

