US Remote-Access Chip Rule to Slow China's AI Push, Maybank Says
BIS's Nov 16 rule closes offshore compute rentals for Chinese AI firms. Maybank says HBM shortages and sub-4% Huawei output versus Nvidia will slow frontier training and favor hyperscalers.
By Grace Kim
5 min read

What's News
- BIS issued an interim final rule on Nov 16 requiring licenses for data centers worldwide providing over 12 million TPP to China- or Macau-headquartered customers.
- Epoch AI estimates Huawei's 2026 computing output at under 4 percent of Nvidia's, with a 73-fold Nvidia advantage in usable domestic HBM supply.
- About 600,000 Nvidia H200 chips are awaiting Beijing's approval in Hong Kong; major Chinese platforms have cleared only about 10,000 each.
- Domestic accelerator chip prices in China have risen 20–30 percent amid HBM shortages.
- DeepSeek's order for 160,000 Ascend 950DT chips faces a delivery wait of more than a year.
US export restrictions released on Nov 16 will slow China's frontier AI training and shift the country's bottleneck from compute access to domestic chip supply, Maybank Investment Bank Research said on Thursday in a research note.
The US Commerce Department's Bureau of Industry and Security (BIS) published an interim final rule, "Addressing Remote Access to Advanced Semiconductors," that effectively closes the last remaining route for Chinese AI companies to reach frontier US chips: renting computing capacity from data centers outside China.
The rule creates a worldwide licensing requirement under Section 744.24 of the US Export Administration Regulations. It covers 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 takeaway: restrictions now follow the customer's headquarters, not the data center's location.
What does the new rule change?
The regulation puts greater responsibility on data-center operators and chip suppliers to verify customer ownership. Where a hosting provider cannot or refuses to run those checks, BIS could treat it as a red flag and presume 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.
One channel stays open: the existing January 2026 framework allowing licensed exports of Nvidia's H200-class chips into China remains in place. Nvidia's own disclosures, cited by Maybank, show China-related Hopper shipments made up less than 1 percent of its data-center revenue in Q2 of financial year 2027, while its Q3 outlook assumes no revenue from China data-center computing. About 600,000 H200 chips were reportedly awaiting Beijing's approval in Hong Kong to enter the mainland, with major Chinese platforms clearing only about 10,000 chips each.
BIS also plugs a loophole: the rule aggregates computing capacity across digitally connected sites, preventing companies from slipping under the threshold by splitting clusters across regions.
How big is China's compute gap?
Large — and measured. Epoch AI estimates cited by Maybank put Huawei's 2026 computing output below 4 percent of Nvidia's, while Semiconductor Manufacturing International Corp (SMIC) ran at more than 97 percent capacity utilization in Q3 2026. Huawei executives have said domestic demand for computing capacity may not be fully met until 2030.
On performance, Huawei's Ascend 950DT delivers one petaflop per second in FP8 — about half of Nvidia's 2022 H100 — and trails the B300 by 6.75 times in FP4. On cost efficiency, it delivers 0.05 tera operations per second per dollar versus 0.34 for the B300.
Maybank cautioned against assuming the curbs will instantly accelerate Chinese innovation and narrow the gap. They may push software optimization and hardware-software co-design, but they won't fix shortages of high-bandwidth memory (HBM), advanced wafer capacity or power efficiency. China's leading AI models have scaled training compute at roughly three times annually since US export controls began tightening in late 2021, versus five times globally — yet still lag the US frontier by only about seven months on average, partly because Chinese firms kept accessing overseas Nvidia clusters, including in Southeast Asia.
Why is HBM the biggest bottleneck?
Maybank identified HBM availability as the primary constraint on China's domestic AI computing buildout. Epoch AI estimates give Nvidia a 73-fold advantage over Huawei in usable domestic HBM supply in 2026, which translates into nearly 200 times more AI computing capacity than Huawei can support from domestically sourced HBM.
The gap should narrow to 11 times by 2028, but Nvidia's computing advantage per gigabyte of memory is projected to widen from three times to six times.
China's domestic HBM supply can support only about 240,000 Ascend 950DT chips this year, forcing another 750,000 Ascend 950PR chips onto non-HBM memory with lower bandwidth of 1.64 terabytes per second — hardware better suited to inference prefill than model training.
Demand is already outrunning supply. DeepSeek's order for 160,000 Ascend 950DT chips faces a delivery wait of more than a year. Alibaba CEO Eddie Wu said the proportion of proprietary chips on Alibaba Cloud remains relatively low due to limited domestic fabrication capacity. Domestic accelerator prices have risen 20 to 30 percent amid the HBM squeeze, and Maybank expects stronger demand to lengthen queues and lift prices rather than immediately expand production.
Who wins and who strains?
Chinese AI companies will lean harder on software optimization and inference workloads, where domestic chips are already deployed. Z.AI serves inference on more than 100,000 domestic chips and cut unit token costs by 80 percent this year. DeepSeek has ported its core computing kernels to Huawei's Ascend 950 chips and reduced memory requirements per token to 890 bytes.
But Maybank noted these moves predate the latest BIS rules, and former White House official Ben Buchanan said no capable Chinese AI system has yet been fully trained on Chinese chips.
The medium-term outcome hinges on whether US licensing results in routine denials for generative AI training, and whether displaced workloads can migrate to alternative infrastructure. If enforcement stays tight, Maybank expects Alibaba, Tencent, ByteDance and Baidu to be better positioned through their finances, proprietary chip development and infrastructure budgets, with DeepSeek relatively resilient thanks to its Huawei kernel partnership. AI labs dependent on rented compute — Moonshot, Z.AI and MiniMax — face the most pressure as they compete for domestic capacity largely controlled by China's tech giants.
By Grace Kim
Source: technode.global
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,582 articles