Memory is China's AI bottleneck, and smuggling can't fix it

US export controls on AI chips are often described as leaky. Chinese labs keep shipping strong models, Huawei keeps announcing new Ascend chips, and smuggling cases make headlines every few weeks. Two new pieces of research from Epoch AI, an independent research group that tracks AI compute, suggest the controls are biting harder than the headlines imply, and that the choke point is not the chip itself but the memory next to it.

More chips, much less compute

Huawei is not short of chips in the usual sense. Epoch estimates it will ship about 1.5 million Ascend units in 2026, a quarter of Nvidia’s 5.9 million. The difference is what each chip can do: Huawei’s best part, the Ascend 950, delivers roughly half the performance of Nvidia’s H100, a chip that first shipped in 2022. Measured in H100-equivalents (H100e), a standard way to compare compute, the gap is about 25 to one.

Chips shipped

1. Nvidia5.9M
2. Huawei1.5M

Compute (H100e)

1. Nvidia23M
2. Huawei<1M

Source: Epoch AI, Will Huawei catch up to Nvidia by 2030? (CC BY)

The memory wall

Modern AI chips are useless without high-bandwidth memory, stacks of DRAM that sit next to the processor and feed it data. The US restricts sales of advanced HBM and of the equipment to make it. China’s own producers, led by CXMT, are ramping quickly but from a low base. After accounting for yields, Epoch estimates Nvidia has access to about 73 times more usable HBM than Huawei in 2026.

The result shows up in Huawei’s product mix. Only a small share of its 2026 chips can be built with Chinese-made HBM.

1. Slower non-HBM memory (950PR)750K
2. Stockpiled and 2025 smuggled HBM (910C)430K
3. Chinese-made HBM (950DT)240K
4. Smuggled HBM (950DT)80K

Source: Epoch AI, Will Huawei catch up to Nvidia by 2030? (CC BY)

Why smuggling can’t close the gap

Smuggling is real and significant. Epoch’s earlier work put chips smuggled into China through the end of 2025 at a median of about 660,000 H100e, roughly a third of China’s AI compute. But the numbers needed to catch Nvidia are of a different order. Epoch models a scenario in which Huawei smuggles ten times as much HBM as China can make in 2028. At a conservative $15 per gigabyte that is a flow of more than $40 billion a year, which the authors call implausibly large. It would raise Huawei’s 2028 output from just under 1.5 million to about 16 million H100e: still only 16% of what Nvidia is expected to ship that year.

1. 2026, all memory sources<4%
2. 2028, Chinese HBM only~1.5%
3. 2028, plus 10× smuggled HBM~16%

Source: Epoch AI, Will Huawei catch up to Nvidia by 2030? (CC BY)

The Malaysia corridor

How much gets through in practice? Customs data offers a rare window. Every shipment is recorded twice, once by the exporting country and once by the importer, under the same product code. Servers fall under HS code 847150. Between April 2024 and June 2025, China recorded $3.8 billion of servers arriving from Malaysia at an average of about $106,000 a machine, the price of an AI server. Malaysia recorded only $0.6 billion of exports. The two sides counted roughly the same number of machines (about 36,000) but valued them very differently: around $17,000 each on departure and $106,000 on arrival. China’s importers have little reason to under-declare, since they pay duties on the value.

The gap is consistent with about $3 billion of AI servers routed through Malaysia, roughly 150,000 H100e. The timing is the telling part. The flow began a few months after the US banned Nvidia’s A800 and H800 chips for China in October 2023, and it collapsed within weeks of Malaysia requiring a permit for exports of high-performance US AI chips in July 2025.

Source: Epoch AI analysis of customs data (CC BY)

After memory, the chips themselves

The memory gap will narrow. Epoch expects the difference in usable HBM to fall from 73 times in 2026 to about 11 times in 2028 as Chinese production ramps. But the gap moves rather than closes: Nvidia’s advantage in compute per gigabyte of memory grows from nearly 3 times to 6 times over the same period, because Chinese chipmakers lack the most advanced manufacturing tools. Even under an optimistic assumption, an 18-fold increase in China’s HBM wafer capacity by 2030, Huawei would produce around 24 million H100e in 2030. That is roughly what Nvidia ships in 2026: a lag of about four years.

What it means

• For policy: the most effective lever is not the chip ban alone but control of memory and chipmaking equipment, plus enforcement at transit hubs. Malaysia’s permit rule shows that a single well-placed rule can shut a multi-billion-dollar corridor within weeks.
• For the AI race: Chinese labs will keep competing on efficiency and price rather than raw scale. That fits what we see in their business: fast growth, high margins and a constant shortage of compute, as in DeepSeek’s case.
• For anyone reading these numbers: H100-equivalents are an on-paper comparison. They ignore software, networking and how efficiently labs use their chips, and the estimates carry wide uncertainty. The direction is clearer than the exact size of the gap.
• The open question: leverage lasts only while enforcement keeps up. Smuggling routes can move to other transit hubs, and the policy debate over selling newer Nvidia chips to China could change the picture faster than any factory in China.