The data we use comes from monthly customs records obtained directly from each national statistical agency’s own API or data portal, covering the period from January 2022 to June 2026. These records always include the dollar value and, in some cases, piece counts. We use the following sources:

We also track nine HS-6 codes connected to AI hardware, from bare processors (854231) through memory (854232) and switches (851762) to assembled servers (847150). The headline figures are based on server values and quantities, but we use the other eight codes to check whether the missing servers simply left Malaysia under a different label, which sometimes happens in other AI trade flows.

These five sources are part of a larger collection of country level customs data aggregated by Epoch (which we will release in an upcoming publication) that covers most of the world’s AI supply chain: the US, Taiwan, South Korea, Japan, the UK, India, Thailand, Indonesia, the EU-27, and Mexico. The broader dataset is used for two things. First, other countries’ import records show how Malaysian-origin servers were priced elsewhere and let us determine that China had a uniquely high-value distribution. Second, we can check whether re-routing is responsible for the discrepancy, by comparing China’s imports against those countries’ exports. That check covers 77% of China’s server imports across the window. The largest origin we do not have is Vietnam, but it receives only around $100 million of servers from Malaysia during the window, and arrivals from Vietnam into China are priced like ordinary hardware at ~$300 per unit.

Six of these reporters publish values only in local currency (Malaysia, Singapore, Hong Kong, Japan, the UK, and the EU-27); we convert them to US dollars using International Monetary Fund (IMF) monthly period-average rates, except the euro, for which we use the European Central Bank (ECB) reference rate. Imports are usually valued including freight and insurance, while exports are not, which inflates the importer’s side by a few percent — we do not adjust for this in our analysis, but it is negligible relative to the discrepancies observed.

Finally, the GPU prices we use to convert dollars into chip counts come from Epoch AI’s Chip Sales dataset.

The headline figures are based on Malaysia’s exports (to China) and China’s imports (from Malaysia) of servers (HS 847150) over the period between April 2024 and June 2025. We sum the dollar value on each side and then divide each side by its machine count, China’s from its own records, Malaysia’s from its UN filing. The counts agree within a few percent in aggregate over the window, though single quarters swing up to two-fold. The analysis therefore focuses on testing alternative explanations for the rise and fall of this flow and the gap in values between the two reporting countries:

1. Routing through a hub. Exporters report the immediate destination of goods while importers report the country of origin, so the disagreement between Malaysia and China could be due to those servers moving through Singapore or Hong Kong before reaching China. This is unlikely: the flows are too small; they are the wrong product, priced at $1,000–$12,000 per unit against the $106,000 China recorded; and the machine counts already match on the Malaysia-China lane.
2. Different commodity codes. Malaysia and China could classify the same goods under different HS codes, so we compare both countries’ records across all nine codes in our AI basket, looking at their direct trade and also what is routed through the hubs. If the discrepancy were driven by differing classifications, Malaysia would show a surplus on some other code, first spiking and then dying with the directive. No code does: the one code where Malaysia declares more than China records — storage (852351) — is matched in scale by a different storage category in China’s records (847170), and the one line that grew — chips into Singapore — is priced differently and kept growing after the Malaysian directive, when server flow fell.
3. Coincidence. The flow’s start and end could just happen to line up with the two policy events. But among all origins from which China booked at least $0.1 billion of servers, only two others have the same surge-and-collapse pattern in both value and price. These are Singapore and Mexico, which together are about a third of Malaysia’s scale. Each died down after an export-control event. China’s imports of Singapore-origin servers surged and collapsed earlier, consistent with the timing of Singapore’s own enforcement actions. Malaysia’s exports of servers to Mexico and China’s import of servers from Mexico also decreased substantially after Malaysia’s directive, suggesting that these were related. We also check how common similar patterns are, leveraging our larger dataset of 14 reporting countries: among the roughly 1,500 time series in our collection that have at least $10 million traded per quarter, about one in eighty shows a comparably sized surge and collapse in value and in price at the same time.
4. A change on China’s side. China’s demand for servers could simply have collapsed, in which case the fall in Malaysian imports would be unremarkable (though this still wouldn’t explain the surge or the discrepancy in records between Malaysia and China). We compare China’s server imports from all other origins before and after the directive: imports grew over the period of interest.

We also check whether similar inconsistencies show up in Malaysia’s server trade with other partners: Malaysia’s declared exports match the partner’s import record within 5% for Singapore, India, South Korea, and the EU-27. Four partners do disagree — Taiwan, Mexico, and the US record less than Malaysia declares, while Hong Kong records more. The discrepancies of Taiwan, Hong Kong, and the US are consistent with simple rerouting, while Mexico’s trade with both China and Malaysia also has the same pattern of sharply increased price during the period between the export controls and the Malaysian directive. We explore this in more detail in the companion notebook.

Could these machines be something other than AI servers? The HS code covers a wide range of goods, from desktop towers to mainframes, but we find evidence for only two types of products costing $100,000 or more. The first is GPU servers, including Nvidia’s H20, the slower chip designed for China and sold at $140,000–$195,000 in its 8-GPU server configuration, and similar H100 systems at $250,000–$320,000. The second is a small set of enterprise systems such as mainframes and large database servers, but they do not fit the scale or origin of this trade: Oracle and IBM are the main producers of such systems, and their entire yearly worldwide hardware revenue is around $12 billion. Nothing in Malaysia’s trade with other countries, described above, suggests it handles machines of this kind.

The data cannot prove that the AI servers contained H100s rather than H20s, but the evidence we have points to that. The $106,000 average is almost certainly lower than the price of the machines themselves. Before the window, China imported tens of thousands of cheap units from Malaysia each quarter, weighing less than a kilogram, and it seems reasonable to assume that at least some of this trade continued. The weights support that: the average weight per unit swings between 22 and 86 kilograms from month to month while the value per kilogram remains stable, which is what a mix of heavy and valuable machines and near-weightless, low-value units would look like. If that mix is right, the machines weigh at least 86 kilograms and cost at least $186,000 each, which already exceeds average H20 prices.

The timing of trade through the corridor does not seem to have been impacted by the H20 ban, with three-quarters of the value arriving before April 2025, when H20s were still legal to sell to China and there was no reason to route them through Malaysia. Additionally, the legal H20 business operated through Chinese server makers that built the systems in China from imported boards, rather than passing through Malaysia as finished machines.

Finally, we estimate how many GPUs this trade could account for. Even if every one of the corridor’s 35,490 declared units were a cheaper machine at $11,600 — the price of the commercial servers that appear under this code after the directive — they would carry at most $0.4 billion of the $3.75 billion, so we treat the dollars as AI hardware. Dividing by the chip prices in Epoch AI’s Chip Sales dataset — $25,000 for an H100 accelerator — puts the corridor at roughly 150,000 H100-equivalents (H100e). Weighting each quarter by the mix of Nvidia GPUs shipping at the time (H100/H200, B200, B300, and the H20, which was then still legal to sell to China) gives about 180,000. These H100e estimates are rough and meant to help convey scale.

The companion notebook reproduces all these checks step by step.

• We take declared values and piece counts at face value. Customs records are not direct measurements, and all the data we use here inherit their errors.
• The unit counts reported by China and Malaysia match over the window but swing quarter-to-quarter, with a ratio between 0.5 and 2. Discrepancies like these are normal, as it takes time for goods to reach their destination and be cleared by customs; what matters is that they offset rather than accumulate, which two unrelated series would have no reason to do.
• We do not have insight into which AI chips were being smuggled. The main compute estimate assumes GPUs in the Hopper family. If the trade had instead been made up mostly of H20s, the same dollar value would imply as few as 50,000 H100e. This is unlikely for the reasons given in the analysis section. A high prevalence of Blackwells would result in a higher compute estimate.
• Since the origin of goods is recorded by the importer, China’s data claiming Malaysia as the origin of the servers could be either wrong or deliberately false. This by itself would not restore a benign reading of the data; it would just change the routing by which export controls were being evaded. However, the fact that these imports returned to baseline within weeks of Malaysia tightening its transit rules suggests this was real physical trade, since a different smuggling scheme would have no reason to respect that.