What data we use

We collect monthly customs data for 16 destinations with their own records: the US, China, Hong Kong, the EU (as one bloc), South Korea, Malaysia, Singapore, the UK, Japan, Indonesia, India, Thailand and Mexico, plus Saudi Arabia, Australia and Canada through UN Comtrade. We also use Taiwan’s export records. We count four server and parts codes (8471.49, complete systems; 8471.50, servers; 8471.80, other computer units; and 8473.30, parts, including boards and cards), using 2021 as a baseline year, and supplement our analysis with other products and more detailed country-specific codes.

Four steps from customs records to chips by country

1. Hardware retained in each country. We count hardware only in the country where it stays, subtracting re-exports and the imported content of servers a country assembles and ships on (Section 3.2).
2. AI-related trade. We estimate AI-related trade as the excess above a baseline built from each country’s 2021 trade (Section 3.3).
3. Accelerator share of trade value. We estimate how much of the remaining value is the AI chip itself, rather than CPUs, memory, networking or chassis, using national tariff descriptions, and declared value per unit and per kilogram (Section 3.4).
4. Allocation of chips to countries. We split our chip sales estimates, by chip model and quarter, across countries in proportion to their chip dollars, assuming chips arrive about two months after they are sold. Huawei and Cambricon chips, which never cross a border, go to China. China’s dollars for diverted chips are adjusted for the premium paid above list price (Sections 4.1–4.3).

Countries whose customs data we do not collect directly, mostly those with a small share of chip trade or those that do not publish their own customs data, are estimated from what other countries declare shipping to them.

How well the estimates match chip sales

As a check, customs-derived chip spending over arrival quarters 2023Q3–2026Q2 comes to 103.4% of the value of the corresponding chips in our sales data; individual quarters differ by 14.6% on average (see the comparison with chip sales). The check isn’t fully independent, because estimating chip content uses our chip prices and product mix.

Allocating the model-level quantities from chip sales using customs geography gives an estimated 27 million H100-equivalents of cumulative arrivals through June 2026. This includes sales recorded since the start of 2022, with an assumed two-month delay before arrival. Some of these chips might not be operational yet, and some chips may be in transit in one country before being shipped to their end use location.

As of 2026Q2, the US has received 66.8% of allocated AI chips and 72.2% of H100-equivalent compute; China received 18.2% of units and 9.6% of compute. These differences depend on model eligibility and the assumed model mix within each eligible group.

How sensitive the results are

The US is robustly the destination of the majority of the compute across the tested modelling assumptions. China, the EU and several smaller destinations are more sensitive to how we distinguish ordinary trade from accelerator shipments. Their structural ranges are substantially wider than intervals obtained by varying parameters within the adopted model (see Section 4.5).

The data also catches some hardware being routed around export controls, as described in this piece on chip smuggling via Malaysia.

The rest of this methodology covers:

• The trade data (Section 2): product codes, sources, publication schedules and data limitations.
• Methodology (Section 3): each step from customs values to chip dollars, and the comparison with chip sales.
• Country estimates (Section 4): allocating chips and compute to countries, and uncertainty.
• Data quality checks (Section 5): how countries’ records of the same trade compare, a closer look at Taiwan, and why we use national records rather than BACI’s reconciled estimates.
• Robustness checks: at every step we report the variants we tested and the known weaknesses of each.

Note: parts of the sections below were drafted with AI assistance.

2.1 Why we use customs data

Goods are classified using Harmonised System (HS) commodity codes. The first six digits identify an internationally standardized category; countries can add digits to distinguish more specific products. For example, 847330 covers computer parts and accessories, while Japan’s nine-digit import code 847330011 identifies DRAM modules within that category.

Four properties make customs data attractive:

1. It has broad coverage. Customs statistics come from declarations required for international shipments, giving them broad coverage without relying on a voluntary survey. Coverage and accuracy nevertheless vary: records can contain missing shipments, incorrect classifications, valuation errors and other reporting problems, including in countries with strong customs enforcement.
2. It is high frequency. Taiwan publishes month M about 7–10 days after it ends; Korea, Malaysia, Singapore and Thailand within 3–4 weeks; the US at ~5–6 weeks; the EU’s detailed 8-digit data at ~2 months.
3. It is bilateral. Every record has a partner country, so flows are directional and traceable: Taiwan’s exports to Malaysia are a different observation from Malaysia’s imports from Taiwan, which can help us identify discrepancies. Importers’ declarations are also generally more reliable because of tariff and tax implications, while the same incentives do not exist for exports.
4. It is granular. National tariff schedules extend the six-digit Harmonised System codes and sometimes isolate useful products. Japan, Korea and the US identify memory modules that can be excluded from accelerator content. Korea’s multichip-memory line also helps trace HBM-related trade with Taiwan, although the code encompasses more than HBM. Product definitions and the limits of these proxies are listed in Section 3.4 and Appendix B.

The main limitation of customs data is that there is no “AI accelerator” code. The last meaningful revision of the HS was done in 2022, before the AI boom. This means that GPUs, AI servers, gaming graphics cards and ordinary rack servers share codes; a GB300 rack and a point-of-sale terminal can both be “automatic data processing units.” Extracting an AI signal therefore requires (a) picking a code basket where AI dominates the recent growth, and (b) explicitly modelling the non-AI part of the flow to remove from said basket.

2.2 Why we do not rely on UN Comtrade alone

UN Comtrade brings together countries’ official trade submissions in a common database, making it a convenient starting point for international comparisons. We use it for additional country coverage and cross-checks, but it alone does not provide the granularity, timeliness or Taiwan coverage this project needs:

• Granularity. Comtrade trims everything to HS-6. The national subdivisions that might isolate HBM, switch/router lines, or GPU add-in cards get discarded.
• Timeliness. Comtrade ingests national submissions and lags them by months, or sometimes over a year. There are also some other quirks: for example, China’s monthly submissions stop at 2024-12, and Thailand is simply missing 2025-03 → 2025-12.
• Coverage. Taiwan does not report to UN Comtrade at all (since it is not a member country). Its national portal is the only source.

Comtrade supplies partner coverage where national extracts are narrow, physical fields where available, and a cross-check on collection. Saudi Arabia, Australia and Canada enter the estimator through their own Comtrade submissions. We also use Comtrade to extend the historical records needed for the 2021 reference and netting warm-up. Taiwan’s history comes from its national portal, while the EU’s comes from Eurostat. The reconstruction of the UK’s 2021 EU imports is described in Section 3.3. For Mexico, the estimator uses INEGI trade values; supplementary INEGI quantities and weights support the sensitivity analysis in Section 3.2.

2.3 How customs records are structured

The unit of observation is a (reporter, flow, partner, commodity code, month) cell with a value in USD (sometimes converted from local currency using IMF data), and, where the source publishes it, a quantity and a net weight. “Reporter” is the country whose customs agency filed the record; “flow” is the direction (import, export, and, in some cases, domestic export vs re-export); “partner” is the other country. This means that every cell should be recorded twice, once by each side of the border, under each side’s own conventions.

2.4 Which product codes we track

Countries report their trade following the Harmonized System. The HS is maintained by the World Customs Organization and used by more than 200 countries covering ~98% of world merchandise trade. It standardizes commodity classification up to six digits and is revised every ~5 years, with the last revision in 2022.

We identify candidate codes from their product descriptions and national tariff subdivisions, then check where the observed accelerator supply-chain growth appears and whether classification differences reconcile across the broader family. The core basket combines processing units, complete computer systems, other computer units and parts. These codes contain both AI and non-AI equipment. We first estimate and subtract a baseline for ordinary computing and component trade. We then estimate how much of the remaining value belongs to accelerator modules, using national product descriptions and declared value per unit or kilogram. The rest can include host memory, CPUs, storage and other system components.

We divide the codes below into two tiers. Tier 1 contains the server, system and parts codes used to estimate accelerator trade and allocate it across countries. Tier 2 covers chips, memory, networking and storage; these codes are used only for diagnostics and supporting analyses, not for the allocation itself. The repository’s code definitions list the codes used by the estimator and explain why bare chips are excluded from the allocation.

2.5 The countries in the study

We have built fourteen national datasets so far, and three more on UN Comtrade. “Flows” uses the schema’s codes (M import, X total export, DX domestic export, RX re-export); “quantity” and “net weight” show the share of value that has that data.

The estimator uses import records for sixteen destinations, treating the EU as one bloc. Other destinations are estimated from selected exporters’ declarations. The main collection covers 2022 onwards, supplemented with 2021 trade and reference-index history to establish the baseline and initialize netting. Taiwan’s 2021 exports cover all four counted hardware codes. The historical extension supports the reference calculation; it does not provide the same national-code and physical-field detail as every current dataset. Partner coverage varies by source: Mexico’s monthly national extract covers nine corridors plus a World control. Each source has a standardized dataset and validation checks. Twelve of the fourteen national collection workflows are automated; Taiwan and China require manual downloads.

Coverage gaps include Vietnam, where tariff-line data would require a third-party source, and the UAE, which publishes no usable customs detail and sends no monthly data to Comtrade. Australia, Saudi Arabia and Canada are covered through their HS-6 Comtrade submissions. The same approach could provide destination customs records for Norway, Switzerland, Israel and Brazil.

The data guide lists the sources and required inputs. Each source module documents its publication, code coverage, flows, file formats and treatment of revisions; see, for example, the United States and Taiwan modules.

2.6 How to read customs records

These are properties of customs data that change how the numbers should be read:

1. Imports are generally attributed to origin, while exports are attributed to the destination known when the declaration is filed. These need not describe the same leg of a shipment. An exporter may name an intermediary hub or a destination beyond it, while the importing country records where the goods were made. Assembly can change origin; distribution or repackaging generally does not.

   For example, the US reports $85.7 billion of server exports to Mexico, while Mexico books only $10.4 billion of arrivals from the US, and books $60.9 billion as Taiwan-origin, more than double the $24.9 billion Taiwan itself reports sending to Mexico. These are Taiwan-built servers passing through US warehouses: the US sees a shipment heading to Mexico and records an export there, Mexico sees Taiwan-made hardware arriving and records an import from Taiwan, and Taiwan sees goods leaving for the US. The two systems record different partner concepts: immediate destination on the export side and country of origin on the import side. Their disagreement therefore does not, by itself, indicate a reporting error.

   

2. The same goods can be filed under different codes. Taiwan and the US show this most clearly. Compare them code by code and the two sides disagree wildly, anywhere from 0.18× to 2.90× on individual lines. Add the whole 8471 and 8473 family together and they agree to within about 4%: $295.7 billion against $283.2 billion. The two authorities simply sort the hardware into different boxes. Family-level comparisons can reduce discrepancies caused by different commodity classifications.

3. Imports and exports are priced differently. An import value usually includes the freight and insurance (Cost, Insurance, and Freight or CIF) needed to get the goods to the border, while an export value usually stops at the port of departure (Free On Board or FOB). So for the same shipment, the importing country’s figure should come out a little higher than the exporting country’s. Converting records reported in local currency to dollars also introduces exchange-rate noise.

   This turns out to explain very little of the disagreement we actually see. Take the corridors where the two sides disagree on value and which also report weight: in 89% of them, the two sides disagree on the tonnage as well. If two countries cannot agree on how many kilograms crossed the border, the gap is not about how the same goods were priced; it is about which goods got counted at all. That leaves freight and insurance able to explain at most about a tenth of the flagged corridors. We still put every country’s import values on the export basis before comparing anything. The US publishes both values for every line, so we measure the freight and insurance margin there, as a curve that falls as goods get more valuable per kilogram, and divide it out of the other countries’ import values. For this hardware the margin is 0.5–1.6% of value (Section 3.2).

4. Recent customs observations can be revised as filings arrive. We exclude incomplete customs quarters and flag the latest allocation as provisional. Estimates can also change when later exports offset previously counted arrivals: the netting rule allows revisions to a surplus during the following two quarters.

5. Access fragility. Two collection workflows are manual because of CAPTCHAs (Taiwan, China), one runs on an undocumented API that could change without notice (Malaysia METS), and one scrapes a fragile OLAP endpoint (Mexico BCMM). These are the collection workflows that need the closest monitoring, and our validation scripts help detect collection failures.

2.7 What customs data cannot capture

1. Hardware that never crosses a border. The dataset sees goods in transit between countries, so anything built and deployed in the same place leaves no record of itself. Taiwan is the pure case: chips fabricated there and installed in Taiwanese data centers generate no customs entry at any point in their life, and nothing recovers them.

2. Value added domestically leaves no trace in the data. Customs records only capture what crossed the border. Anything added to the hardware afterwards (a designer’s margin on a domestic chip sale, integration work, a distributor’s markup) is simply not in the record, so the data itself gives no way to size it. It can still be corrected for, but only with an outside estimate rather than from the trade figures.

   This may help explain the early shortfall relative to chip sales, although we cannot confirm the explanation. Boards crossing into the US through 2023H2–2024H1 were priced at only 40–90% of the chip sales average selling price, rising towards it in later quarters. One explanation that fits is that Nvidia was not yet leaning on ODMs and integrators as heavily as it later would, so hardware entered at intercompany cost and was sold on to the final customer only after clearing customs, leaving the margin as a purely domestic transaction. If that is what happened, the border values in those quarters understate the chip flow, and it would account for the estimate falling short of chip sales there (Section 3.5). Other explanations fit the same observation, though (a different product mix, or genuinely cheaper early hardware), and without transaction-level evidence we cannot separate them well enough to justify a correction.

3. Movement that nobody declares. Goods leaving a country without an export declaration are invisible in that country’s records, although the destination may record their arrival. Malaysia is an explicit exception to our usual reliance on the importer’s own book: after declared-export deductions, we subtract 50% of the positive gap between China’s imports from Malaysia and Malaysia’s declared exports to China for server codes 847149 and 847150, per calendar quarter on a common FOB basis. The same adjustment applies to the 2021 reference. This is a judgement about possible undeclared onward shipments, not an estimated smuggling rate; routing, timing and origin attribution can also create gaps. Goods undeclared on both sides remain invisible.

3.1 Overview of the method

Our aim is to turn customs values into AI chip dollars by country that can be compared with our chip sales data. Our sales estimates value each chip at a fixed price per model: for example, $25,000 for an H100 in every quarter it sold. That price covers the accelerator module, including its HBM and packaging, but not networking, chassis, system integration or later markups. Customs values cover whatever crosses the border, which can be a board, a server or a full rack. Each step below moves customs values closer to our sales basis.

The method consists of four steps. Trade netting and chip-content classification use monthly observations; excess offsets are tracked quarterly:

1. Get the relevant trade dollars (3.2): estimate hardware retained in each destination by removing declared re-exports, the imported counted content of exported servers, own-origin returns and mid-chain bare-chip crossings.
2. Take the AI share (3.3): compare retained trade with an indexed 2021 reference, allowing nearby quarterly shortfalls and surpluses to offset.
3. Convert to chip dollars (3.4): estimate accelerator-content shares from national tariff labels, unit values and value-to-weight ratios, then apply those shares to the excess trade dollars.
4. Line up and compare (3.5): compare calendar-quarter customs totals with sales interpolated into the same arrival periods under a two-month lag.

Over arrival quarters 2023Q3–2026Q2, customs chip dollars total $408.2 billion, or 103.4% of the corresponding chip sales dollars. Section 3.5 reports the full-period and recent-quarter discrepancies. Section 4.5 describes uncertainty in the resulting country allocations.

3.2 Step 1: estimate retained trade

The principle is that each accelerator should be counted once, at the market where it is finally deployed. Gross customs imports break that rule in four ways. The same correction applies to every country’s records:

Netting mechanics

Three mechanics then apply throughout, per code and month:

• Carry. An export deduction that exceeds current imports is carried forward rather than discarded. Netting starts in January 2021, so the published 2022 estimates inherit the preceding year’s balances. A separate quarterly rule can revise previously counted excess when a later shortfall occurs; Section 3.3 explains that charge-back.
• Imported content of exported servers. A country that builds servers from imported boards and ships them out has sent the boards on in a different form. We subtract those exports at the share of their value that was imported, and we read that share from how heavy the exported machines are per dollar. Light, expensive machines are mostly imported accelerators; heavy, cheap ones are mostly locally assembled metal and CPUs. A server exported at $250 per kilogram nets 10% of its value, at $600 per kilogram 40%, at $2,000 per kilogram 90%. If the weight of the exported machines is not available we take the same share from their value per unit, and if neither is available we use the accelerator share found in the country’s own imports of that code. The deduction is taken first from that month’s imports of the same code, and any remainder from the country’s other counted import lines in proportion to their size. Exported boards and other units are never deducted: their inputs are bare chips and memory, which the basket does not count.
• Destinations without import records. We use selected exporters’ shipments to estimate arrivals in destinations whose own records are unavailable. These include hub re-exports, US domestic exports and selected finished-hardware exports from the EU, Malaysia, Singapore, Hong Kong, the UK, Indonesia and India, together with Taiwan’s residual destination coverage. These are gross trade values entering the baseline and content calculations, not amounts treated wholly as accelerator value.

Put together. For each country, code and month:

\[\text{retained} = \text{imports} - \text{own goods returning} - \text{declared re-exports} - k \times \text{own-made server exports}\]

where \(k\) is the estimated imported share. Malaysia also receives the server mirror-gap deduction described below. A negative month carries into the next.

The table shows the netting calculation for each reporting country or bloc for the finished-hardware basket (847149/50/80 and 847330), in billions of dollars. Totals start in 2023Q3 and end at the last available month for each reporting country or bloc, so the endpoints vary across rows:

	Gross $B	Re-exports and own goods returning $B	Own-made server deduction $B (k)	Retained $B	Removed
United States	668.7	134.3	20.5 (45%)	514.0	23%
China	211.2	47.7 (own goods returning)	9.6 (14%)	153.9	27%
European Union	157.6	59.7	7.6 (33%)	90.2	43%
Hong Kong	153.7	199.4	0.0 (28%)	0.0	100%
Mexico	138.4	0.0	149.1 (77%)	0.0	100%
Singapore	87.1	98.0	8.8 (68%)	6.4	93%
Malaysia	47.0	8.2	7.3 (38%)	29.3	38%
South Korea	35.6	0.0	0.3 (35%)	35.3	1%
Japan	30.2	0.0	0.2 (30%)	30.0	1%
United Kingdom	28.7	9.5	0.9 (48%)	18.2	36%
India	25.3	0.0	0.5 (54%)	24.8	2%
Thailand	19.9	0.0	6.4 (70%)	13.5	32%
Canada	18.5	1.3	0.4 (25%)	16.8	9%
Australia	16.1	0.0	0.1 (33%)	15.9	1%
Indonesia	9.7	0.0	0.1 (59%)	9.7	1%
Saudi Arabia	4.1	0.0	0.0 (59%)	4.1	1%

Across the reporting windows in this table, netting removes 42% of gross finished-hardware trade ($690 billion of $1,652 billion). Malaysia’s retained total includes the additional $2.04 billion server mirror-gap deduction, shown separately in the discussion below rather than in the own-made-server column. Mexico’s retained trade rounds to zero at this precision. Vietnam’s output is observed through its partners’ imports because the estimator has no Vietnamese destination customs dataset. Destinations without their own datasets are estimated from selected dispatch corridors.

Six cases are worth discussing in more detail:

• European Union. Netted as one bloc, because members are not self-contained: Hungary exports $12 billion of servers against $0.7 billion of extra-EU imports, since its inputs arrive from stock cleared in the Netherlands. The re-export share of each member’s exports comes from its own statistics. What it reports producing itself, in Eurostat’s PRODCOM survey, sets the floor of own-made goods, and its imports cap what it can have re-exported. The Netherlands and Ireland re-export essentially everything they ship out and Germany 95%; Hungary, Czechia and Poland assemble rather than trade, at 4%, 20% and 10%. Weighted together, 58% of the EU’s server exports are re-exports.
• South Korea. Korea follows the same netting rule as other destinations. Its exports under these codes are dominated by memory modules on the parts line, whose own-made exports are not deducted. Explicitly labelled imported DRAM modules receive zero accelerator content in the classification step; other products retain their classifications from national tariff categories or physical measurements (Section 3.4).
• Malaysia. An assembler and a hub at once. Its $8.2 billion of declared re-exports are deducted at full value, and its own-made server exports are deducted at their estimated imported-content share of 38%, or $7.3 billion. We additionally deduct 50% of the positive China–Malaysia server mirror gap, by code and calendar quarter on the FOB basis. This adds a $2.04 billion deduction over the table’s reporting window and leaves $29.3 billion of retained trade. The quarterly gap is distributed over China’s recorded monthly arrivals in proportion to their value, then passes through the existing negative-carry rule. The same rule applies in 2021; no gap is deducted for boards or other units, and China receives no additional credit. Zero and full deductions are structural alternatives without assigned probabilities.
• Mexico. Export k uses the chip-content estimate for imports as a fallback. A supplementary INEGI dataset provides monthly quantities and weights for January 2023–June 2026 across nine corridors plus a World total, and annual observations for 2022–23. Its 5,218 overlapping trade values match the estimator’s input data exactly. Sensitivity tests use these physical fields, but they do not reliably distinguish ordinary-computer assembly from accelerator assembly, so the baseline estimate uses the import-content fallback.
• Thailand. Thailand does not separate re-exports from its own exports, so every export counts as own-made. Exported servers are deducted at the imported-content share implied by their physical evidence, rather than at full value.
• The other final markets. Japan, India, Indonesia, Australia and Saudi Arabia export little on these codes, and the deduction removes 1–3% of their gross imports. Canada declares its re-exports, which are deducted at full value, 9% of its gross imports in all. The UK is a distributor: its production statistics (the ONS ProdCom survey) say it makes little of what it ships, so 76% of its exports are deducted as re-exports, 36% of its gross imports in all; they go to the Netherlands, Germany, the US, Iceland and Ireland.

3.3 Step 2: separate AI trade from ordinary trade

There is no product code for AI hardware. The codes we use carry plenty of goods that have nothing to do with AI (racks of ordinary web servers, standard logic and memory modules), and netting in step 1 does nothing about that, because those goods really did arrive and really did stay. So we cannot treat every netted dollar as AI. We have to model the part that isn’t.

How much retained trade counts as AI-related

The table decomposes the reference adjustment over calendar arrival quarters 2023Q3–2026Q2. “Full window” covers 2022Q1–2026Q2; 2021 is used only for the reference and warm-up. Values are billions of dollars.

Destination	Retained trade, full window ($B)	Excess, full window ($B)	Retained trade, 2023Q3–2026Q2 ($B)	Excess, 2023Q3–2026Q2 ($B)	Share removed by baseline
United States	600.8	380.1	514.0	362.2	30%
China	180.6	74.2	153.9	71.3	54%
EU	113.2	49.4	90.2	46.1	49%
Malaysia	32.0	22.6	29.3	22.0	25%
South Korea	48.1	17.9	35.3	17.2	51%
United Kingdom	23.9	10.6	18.2	9.5	48%
Japan	41.0	9.9	30.0	9.0	70%
Australia	20.3	8.6	15.9	8.2	49%
Canada	23.0	7.4	16.8	6.8	60%
India	32.2	9.8	24.8	9.1	63%
Indonesia	12.2	7.5	9.7	6.6	31%
Singapore	8.7	1.2	6.4	1.1	82%

The reference and quarterly offsets remove different shares of retained trade across destinations. Over 2023Q3–2026Q2, the reduction is 30% for the US, 54% for China, 49% for the EU and 51% for Korea. It is 70% for Japan, 63% for India and 60% for Canada. Malaysia’s reduction is 25%, while Singapore’s is 82%. These differences depend on the 2021 reference, subsequent index movements and the quarterly offsets; they do not directly identify the products in the remaining excess.

How the choice of baseline changes the estimate

Over calendar arrival quarters 2023Q3–2026Q2, the indexed 2021 reference leaves $592 billion of excess trade. Fixed annual growth of 0%, 5% and 10% leaves $616 billion, $553 billion and $489 billion, respectively. With no reference subtraction, retained trade totals $1,003 billion. These are trade dollars before the accelerator-content adjustment, not chip-dollar estimates.

In the 2023Q3–2024Q1 trough, the indexed reference leaves $12–15 billion a quarter above the baseline, compared with $6–7 billion under 5% fixed growth. The measured index lowers the baseline during the ordinary-IT downturn, leaving more of the observed imports above it.

In recent quarters, rising component values increase the boards reference faster than the fixed-growth alternatives. The data do not separately identify how much of this movement reflects memory repricing, other component demand or accelerator trade.

The alternative-basket tests described above are not included in this chart, because they do not yet have consistent 2021 histories. This comparison uses current netting and changes only the reference used to calculate excess trade.

3.4 Step 3: estimate the AI chip share of trade value

The excess from step 2 is trade value attributed to AI hardware, not necessarily accelerator value alone. An HGX server reference configuration priced at $285k contains about $208k of accelerator modules; other value belongs to CPUs, host memory, storage, networking, chassis and integration. The accelerator basis includes the module’s HBM and packaging. We multiply excess dollars by an estimated accelerator-content share to make the comparison with the narrower chip sales basis.

We estimate chip content separately for each combination of importer, origin and month, then apply the resulting monthly share to the destination’s excess from step 2. Where a country’s tariff codes identify the product, we use them; for example, memory modules in Japan, South Korea and the US get no chip value. Otherwise, we compare the declared value per unit, or per kilogram where unit counts have insufficient coverage, with reference prices for accelerator boards and servers. Those reference prices come from our chip prices and product mix, which makes this the step where the customs estimates depend most on our sales data.

• Family constants and the product mix: each accelerator family gets its own content constant, and the constants are weighted by the quarterly product mix in chip sales to give the share for hardware that arrives as complete systems.

  System type	Chip content	Basis
  HGX servers	0.73	$208k of accelerator in a $285k server
  GB200 / GB300 racks	0.79	Built up per generation, below
  TPU and Trainium systems	0.70	In-house chips (TPU, Trainium) are valued near production cost, so no margin uplift

• Rack configurations: chip content depends on the system configuration and its valuation. A GB200 NVL72 reference has 72 GPUs at about $37k each: $2.66 million of modules in a $3.1 million rack, or 0.86. A GB300 reference has 72 at about $43k: $3.10 million in a rack just under $4 million, or about 0.79. The estimator uses 0.79 for the rack benchmark. Appendix A specifies the content perturbations used in the uncertainty model.

• Reading each cell: every combination of importer, origin and month is a cell (importing member, origin and month for the EU), and each is read from its own records. Where the number of units is available for at least 90% of the cell’s value, we divide dollars by pieces and compare the result with the accelerator price available in that market: roughly $12k for an H20 and $30k for an H800 on the China side, around $33k elsewhere. That price is the average across the models on sale, so a board carrying one of the cheaper models, or one declared at cost between related companies, can arrive at a third of it and still be one accelerator. The boards band therefore starts at 0.35 times the reference. The table gives the reference points used to assign accelerator-content shares; intermediate values are interpolated.

  Dollars per unit	What is crossing	Content share
  Under 15% of the accelerator price	Ordinary electronics	0.05
  A third to 2.5 times it	Accelerator boards	0.93
  More than five times it	Complete servers	The product-mix share above

  When piece counts do not meet the coverage rule, the classifier uses dollars per kilogram where available. The assumed boards ramp rises from 0.05 at $1,000/kg to 0.93 at $3,000/kg; the machines ramp rises from 0.05 at $600/kg to the eligible system-mix share at $2,000/kg, with log-linear interpolation. These are classification assumptions, not product-identification tests: valuable memory can have densities similar to accelerator boards. Each monthly cell is weighted by its positive excess above its own 2022 pace. We use unsmoothed monthly content when classified excess reaches $50 million/3 (about $16.7 million) and the selected physical field covers at least 90% of that month’s value. With insufficient monthly evidence, we retain the existing quarterly reading (a three-quarter median), then the pooled prior for that destination and code. Where physical data provide no content estimate, machines use the eligible product mix and boards receive the default share of 0.93. Native tariff-code content weights are also monthly; explicitly identified memory stays at zero.

  Historical quarterly diagnostics illustrate the scoring rule. Boards arriving in China from Vietnam in 2024Q3 were declared at about $2,100 per kilogram, between the two density anchors, implying a score of 0.66. Boards arriving in Malaysia from Taiwan through 2025 were declared at $460–740 per kilogram, below the lower anchor, implying 0.05. These pooled examples explain the classifier; the baseline estimate applies it to each month’s evidence without smoothing where the monthly gates are met.

  The classifier assigns different effective content scores to different observed mixes. Low-value server cells receive low scores, while ambiguous parts can fall well below the 0.93 accelerator-board benchmark.

• National tariff lines take precedence where their definitions identify product scope. We assign no accelerator value to separately identified standalone memory modules in Japan, the US and South Korea. HBM and packaging already included in an accelerator’s selling price remain in scope. The three memory-module codes are listed below.

Japan’s module and remainder cells preserve each origin-month’s total FOB value, allocated using the national tariff lines’ declared-value shares, with quantities and weights kept in their own groups. This separates identifiable memory before estimating content in the remainder. Identified memory modules remain in the trade totals used for netting but contribute no accelerator dollars. Broad non-memory categories can still contain heterogeneous products.

How much of the excess is accelerator value

Over calendar arrival quarters 2023Q3–2026Q2, $592.0 billion of excess trade becomes $408.2 billion of accelerator dollars, an effective content share of 0.69. The table covers destinations with at least $2 billion of excess trade over that window.

	Excess, AI window $B	Chip, AI window $B	Chip share
United States	362.2	277.3	0.77
China	71.3	49.1	0.69
European Union	46.1	22.2	0.48
Malaysia	22.0	11.6	0.53
South Korea	17.2	1.8	0.11
United Kingdom	9.5	5.8	0.61
India	9.1	7.1	0.78
Japan	9.0	6.7	0.74
Australia	8.2	6.6	0.81
Canada	6.8	2.6	0.38
Indonesia	6.6	3.8	0.57
Norway	3.3	2.2	0.67
Thailand	2.8	1.9	0.66
United Arab Emirates	2.7	1.2	0.44

Effective shares range from 0.11 for Korea, after explicit DRAM-module exclusion, to 0.81 for Australia. The US is 0.77, Japan 0.74, the EU 0.48, Canada 0.38 and Malaysia 0.53. These averages pool monthly readings over 2023Q3–2026Q2 and remain estimates of product composition.

3.5 How the estimates compare with chip sales

We compare calendar-quarter customs chip dollars with chip sales interpolated into the same arrival quarters under the two-month lag. The full AI comparison covers 2023Q3–2026Q2 arrivals; the recent comparison covers 2025Q1–2026Q2.

Cumulative match divides total customs chip dollars by total corresponding sales dollars. Median quarterly match is the median of the quarterly ratios. Mean absolute percentage difference averages the absolute quarterly gaps relative to sales. These measure discrepancies against chip sales, not errors against known truth.

How the baseline choice affects the match with chip sales

The October 19, 2026 rerun uses the 2021 reference and calendar arrivals with a two-month sales lag. Over 2023Q3–2026Q2, median quarterly customs-to-sales ratios are 0.99 for the indexed reference and 0.96, 0.86 and 0.75 for fixed annual growth of 0%, 5% and 10%. Without reference subtraction, the median is 1.65. Netting and physical content evidence are held fixed across these comparisons. The alternative-basket lines are omitted because they do not yet have consistent 2021 histories. This rerun includes the Australia historical-value and Singapore piece-coverage corrections; it is a descriptive sensitivity test, not independent validation or a rule for selecting the baseline.

4.1 How chips are divided among countries

We first estimate how many chips of each model arrived worldwide in each quarter, using chip sales and the assumed two-month delay. We then divide those chips among countries according to their shares of customs-derived accelerator dollars, subject to the model eligibility rules below. The country estimates add up to the worldwide total for each model.

The latest complete customs quarter is April–June 2026. Under the two-month lag, its sales budget covers February–April. Cumulative arrivals through June therefore include all recorded sales from 2022Q1 through 2026Q1, plus one-third of 2026Q2 sales. Residual-country splits and EU member allocations use the same calendar customs quarters.

The series begins with sales recorded in 2022, so its first arrival quarter includes only the March arrivals attributed to January sales. Missing or incomplete source-sales windows are flagged, and incomplete customs quarters are excluded.

Not every model can go everywhere, so units are split into three pools before distribution:

• China-fabbed models (Huawei, Cambricon) never appear in Taiwan-supply-chain trade data, so they are allocated to China outright.
• China-market variants (A800, H800, H20, MI308X, i.e. the models designed for lawful sale to China) are split across China and Hong Kong by their corrected dollar shares, which in practice means China.
• Everything else, the global pool, is allocated across all destinations by their share of corrected chip dollars, after netting out of China’s dollars what the first two pools already absorbed. Within the pool every destination gets the same model mix, with some exceptions as described below.

We exclude allocations below one unit per destination-quarter and redistribute their dollars within the same pool. Each destination-quarter also carries an evidence score: its excess relative to the variability of its pre-window quarterly gaps between retained trade and the baseline, through 2023Q2. The noise scale has an absolute floor and a floor proportional to the mean baseline. Cumulative scores adjust for the number of quarters accumulated. The score is published alongside the estimate and is not used to remove allocations.

4.2 Which chips can go where

Without eligibility restrictions, proportional allocation assigns every global-pool model to every destination, including countries where we have no evidence that those models are deployed. Google’s TPUs and Amazon’s Trainium are not sold on the merchant market, so they can only sit where their designers operate them. AMD’s merchant volume is small and committed to Western hyperscalers. Over arrival quarters 2023Q3–2026Q2, an unrestricted rule would allocate China and Hong Kong 0.27 million H100e of Google TPUs, 0.09 million H100e of Amazon Trainium and 0.13 million H100e of AMD chips.

For these reasons, AMD’s models are allocated only across destinations outside China, Hong Kong and Macau, and Google’s and Amazon’s silicon to the eligible locations in our geography table: TPU v4, v4i, v7 and v8i and Trainium in the US alone; v5p in the US and the EU; v5e in the US, the EU, Japan, Chile, Singapore and Canada; v6e in the US, the EU, Japan, Chile, Singapore, India and Brazil. Within each model’s eligible set, shares are renormalized. China retains its dollar allocation, but within the global pool the model assigns those dollars to Nvidia GPUs assumed to reach China through diversion after export controls. The replacement is not one-for-one in units: captive silicon is priced near cost while merchant Nvidia carries the full margin, so the same dollars buy fewer chips. Over the same quarters, these eligibility constraints lower the cumulative China + Hong Kong allocation by about 10%: 2.58 million H100e against 2.88 million under the uniform mix. The withheld capacity is partly replaced by 0.18 million H100e of additional merchant Nvidia. These are allocation counterfactuals, not independently observed inventories.

TPU v8i is assigned to the US as a placeholder because the geography table has no public region evidence for it. This should not be read as a confirmed deployment location. About 31% of the US allocation in H100e comes from TPU and Trainium geography assumptions; about 39% of China’s comes from Huawei and Cambricon assigned to China. Customs shares determine the distribution of the remaining eligible supply.

4.3 The China price premium

Hardware reaching China after export controls does not change hands at the chip sales list price: an intermediary margin sits between what the customs record shows and what the designer originally charged. Converting China’s customs dollars into units at list prices would therefore overstate its chip count, so we adjust China’s global-pool dollar allocation for the estimated premium before converting it to units. The dated premium applies to the customs-dollar budget for the arrival quarter before allocation to eligible merchant models. Its effect depends on both customs geography and the sales budget interpolated into that quarter.

The premium is a dated series based on reports of diverted-chip sales, benchmarked against the per-model prices in chip sales:

4.4 Results: AI chip arrivals by country

Cumulative allocated arrivals through June 2026 total 27.391 million H100-equivalents and 24.184 million AI chips. These estimates use the 2021 reference, historical warm-up, retrospective charge-backs, monthly chip content and fixed calendar customs quarters. The primary estimate reflects the adopted assumptions; it is not a simulation median. The table uses the October 15, 2026 allocation and its matching provisional 144-case pilot, reporting primary estimates, conditional 90% intervals and structural envelopes (Section 4.5), with compute in millions of H100-equivalents. The primary estimate need not be at the midpoint of the structural envelope, which spans alternative assumptions without assigning them probabilities.

The US receives 66.8% of allocated units and 72.2% of compute; China receives 18.2% and 9.6%, respectively. South Korea’s primary estimate is 0.105 million H100e. Country uncertainty endpoints cannot be summed: shares are normalized within each draw, and “All others” is summed within each draw before its percentiles are computed.

4.5 How uncertain the estimates are

Our estimates can be wrong in two ways. The inputs can be off: sales volumes, prices and imported-content shares. And we can misread what the trade contains: expensive memory can look like accelerator boards, and an average server price can hide very different products.

We report two ranges. The 90% interval varies the inputs under our main assumptions. The structural range repeats this across 144 alternative sets of assumptions about trade content, the baseline and server product mix, and runs from the lowest 5th percentile to the highest 95th percentile across them. We don’t assign probabilities to the alternatives, so the structural range is not a 90% interval, and it only covers the alternatives we tested.

Each draw varies the indexed reference, imported content of exports (k), physical-classification thresholds, selected chip-content fractions and global sales inputs. We recompute trade netting, baseline weights, quarterly offsets, retrospective charge-backs, chip content and country allocation. The adopted model uses the 2021 reference without an AI carve; one structural alternative restores the 2022 reference with its carve. Sales are sampled in their original reported quarters before interpolation into arrival quarters.

Parameter draws vary around the values used for each case. Baseline and k errors combine shared and country-specific components that persist across quarters; classifier thresholds and category-content shifts also persist. National tariff classifications retain precedence, and import-derived k follows the content model rather than receiving a second independent k shock. On rows covered by an explicit server-mixture model, the mixture replaces the additive server-content shift. Appendix A gives the priors and structural cases.

Our chip sales estimates include uncertainty ranges for each model’s sales, but those ranges do not specify how errors move together across models or quarters. We assume that an overestimate of one model’s sales tends to persist across quarters, and that models from the same designer share some uncertainty. We fit positive two-piece lognormal distributions to the model-level quantity and H100-equivalent quantiles in chip sales. Dependence is introduced through a shared designer component, with a correlation of 0.5 between the underlying Gaussian variables. This is an assumption of this project about joint uncertainty, not an estimate taken from our chip sales data; Appendix A.2 gives the construction.

Price and quantity uncertainty are connected. If hardware revenue is $1 billion, a price of $25,000 implies 40,000 chips, while $20,000 implies 50,000. Those quantity estimates represent different prices for the same spending. Drawing price and quantity independently would also generate combinations inconsistent with that revenue estimate.

For Nvidia, we therefore draw quantity and recover the corresponding price using a shared hardware-revenue factor informed by our chip sales notebook. For AMD, Google and Cambricon, we assume that a higher quantity reflects both lower prices and higher spending, dividing the change equally on a logarithmic scale. That split is our modelling judgement. Designers with direct-volume estimates receive a separate illustrative price prior. Appendix A.3 specifies the equations and alternatives.

Each of the 144 cases uses the same 64 proposals: 32 from each of two independently scrambled Sobol sequences, giving 9,216 case–draw combinations. All passed the warning-based acceptance filter, so filtered and unfiltered endpoints coincide. Each draw conserves its own global units and H100-equivalent supply. These checks establish internal consistency, not the validity of the priors or product interpretations.

Through June 2026, the world primary estimate is 27.39 million H100e, with a conditional p5–p95 interval of 24.64–30.21 million. The structural envelope has the same world endpoints because every case allocates the same sampled global arrival budget.

With 64 proposals per case, each 5% tail is informed by roughly three draws. The maximum half-to-full endpoint movement is 8.13%, and the maximum gap between the two independent scrambles is 15.95%. These results remain a provisional small-sample pilot, not a convergence-certified release. The diagnostics measure numerical stability, not the validity of the priors or whether the model covers the true uncertainty; broad comparisons are more reliable than precise endpoints.

The structural cases vary ambiguous parts and other-unit content, the baseline index, server product mixtures and which products remain after baseline subtraction. They also include a combined low-content case, the 2022 reference and a high-China case that removes the diversion premium and excludes Vietnam from China’s boards reference. The 48 content, reference and selection specifications include an isolated merchant-only classifier-price alternative; crossing each with 0%, 50% and 100% Malaysia server-gap deductions gives 144 cases. No probabilities are assigned to these structural choices. All retain monthly content and the two-month sales-to-arrival convention. Reporting errors, country-specific model mixes, premiums and shipment timing are not comprehensively covered.

For missing exporter physical fields, low and high imputation scenarios use donor-content percentiles, expanded where necessary to include the mean used by the primary estimate. These are alternative content assumptions, not confidence bounds. Separate retention scenarios test the possibility that inferred arrivals move onward: capacity not retained locally is reported as unlocated rather than reassigned to a guessed destination. These scenarios preserve global supply and carry no assigned probabilities. They remain separate from the 144 cases used for the reported envelope.

5.1 Comparing exporter and importer records

For every pair of countries where we hold both sides, and for each of the nine codes in the basket, we compare A’s declared exports to B against B’s declared imports from A, month by month. That gives 1,190 directed country pairs and about 10,400 combinations of pair and code. Together they cover roughly $6 trillion of trade that both sides recorded over 2022–2026. A perfect mirror is a ratio of 1.00. We count anything within ±20% as agreement, which is about the room that freight, insurance and currency conversion should need.

The two countries’ records rarely agree. Only 18% of the combinations land within ±20%, and the median disagreement is a factor of 2.3.

Three facts narrow down what that disagreement is. It is not a bias in one direction: the median ratio is 0.88, close to 1, so importers as a group are not reporting systematically more or less than exporters. It is not a calendar effect either. Scored monthly, annually, or pooled over the whole window, the agreement rate stays at 18–20%, and if shipments were simply straddling a month or year boundary then pooling would fix it. And it is mostly not about pricing. In 89% of the disagreeing corridors that report weight on both sides, the two countries disagree on the tonnage as well, and freight and insurance move value rather than kilograms. That leaves pricing able to explain at most about a tenth of the flagged corridors.

Pricing differences alone cannot explain these discrepancies. Differences in coverage, commodity classification and partner attribution also matter. We examine three mechanisms behind these discrepancies. Section 2.6 introduced the first two: countries file the same hardware under different codes, and exporters record where a shipment was sent while importers record where it was made.

The third mechanism is the largest, and it belongs to the re-export hubs. A hub records goods leaving for a partner, but the partner does not record them as arriving from the hub. It records them as arriving from wherever they were made, so the hub’s export appears nowhere in the partner’s customs records. Hong Kong reports $323 billion of processor re-exports (8542.31) to mainland China over the window, and China books almost none of it as coming from Hong Kong: it books those chips as arriving from Taiwan and Korea, where they were made. Across the collection, 823 combinations of pair and code are one-sided in this way, with one country recording a flow that the other records nothing of.

5.2 How much of the disagreement we can explain

Each mechanism has its own fix, and because the mechanisms are different, the fixes add up rather than overlap. We apply them in order. Each one is scored the same way: what share of the trade dollars now sits in corridors where the two countries’ records agree within ±20%? We score dollars rather than corridor counts, because a fix aimed at the largest flows should be judged on the largest flows.

1. Record both sides by where the goods were made. The importer side already is, since customs books imports by origin. The exporter side takes two moves. Five of our sources separate goods a country made itself from goods it is only passing on: the US, Hong Kong, Singapore, Malaysia, and Canada. Where that split exists, we stop crediting the hub with goods it did not make. Then we credit each hub’s re-exports back to the countries the goods came from, in proportion to what the hub itself imported under the same code. That proportion is an assumption, so we tested it on the largest hub: Hong Kong re-exports goods in almost exactly the origin mix in which it imports them, with a correlation of 0.999.
2. Compare families of codes, not single codes. Grouping the nine codes into five families (servers and boards together, logic chips, memory, networking, storage) cancels the classification disagreement inside each family, because both sides are then added up over the same set of codes. The grouping has to be chosen with care. The biggest single disagreement is Taiwan filing servers under 8471.80 that the US books as parts under 8473.30, and those two codes sit under different four-digit parents, so aggregating to four digits cannot cancel it. The family has to span both parents. Coarser grouping also always scores better for a trivial reason, because fewer and larger buckets average out noise. To check that ours wins on economic sense rather than on arithmetic, we scored it against all 1,260 possible regroupings of the same shape. It comes out at the 99.8th percentile.
3. Where the invoices still disagree, ask about the goods instead. Any remaining failing corridor whose sources publish piece counts on both sides can be split exactly into a quantity and a unit price. That separates two questions: did the two agencies disagree about what moved, or only about what it was worth? Where the goods can be counted, about two-thirds of the failing dollars fail mainly on price rather than on quantity.

The first step closes the widest gaps in the dataset. Taiwan’s bare-processor exports to China disagreed by a factor of 48, because about $161 billion of them physically travel through Hong Kong and Singapore. Credited to the Taiwan-to-China corridor, that disagreement falls to 2.2×. US servers to Mexico go from 0.12× to 1.4×, and Singapore to Taiwan from 0.04× to 1.1×.

The clearest example of the third step is also the largest failing corridor in the dataset. Taiwan and Hong Kong disagree by 1.7× on the dollar value of the logic chips moving between them, while their piece counts agree to within 1%: 163 billion units against 165 billion.

The price gap exposed by this diagnostic is persistent. Taiwan-made chips are booked into Hong Kong and Singapore at 1.5 to 2.2 times Taiwan’s declared values across the historical sample, including before the AI boom. Related-company invoicing is a plausible explanation: the two declarations may value different transactions in the same supply chain. Where comparable piece counts agree, they help distinguish valuation gaps from quantity gaps. This does not establish that physical fields are accurate on other corridors or that quantities can identify accelerator content by themselves.

Some remaining discrepancies cannot be investigated because piece counts are missing; others remain unexplained despite the available physical data. Nearly half of the unreconciled dollars cannot be diagnosed at all, because one side publishes no usable piece counts. Korea publishes none, which is the highest-value data gap left in the project. Most of the rest is the pricing gap just described. Corridors where the two countries still appear to disagree about physical goods, after every correction, hold about 6% of the trade we can check from both sides; this is the same trade the chart below shows as the 14.5% of unreconciled value that is unexplained. The largest of them is Korea’s memory exports to Taiwan, which Section 5.3 returns to.

To check for errors in our collection or processing, we repeated the comparison using UN Comtrade and separate code. Comtrade reproduces the aggregate agreement rate, at 17–18% against our 18–20%, and it reproduces the individual corridor gaps, with a correlation of 0.88 in logs. Customs mirrors disagreeing by a factor of two is a general property of trade statistics, not something our basket or our processing introduced.

One difference is specific to our basket, though. We also ran a control basket of nine ordinary goods: cars, packaged medicines, t-shirts, coffee, furniture, footwear, toys, plastic bottles and aluminium. For ordinary goods the largest flows mirror well, so 45% of their trade dollars agree within ±20%. For chips only 25% do. Ordinary goods disagree in their small flows, and chips disagree in their biggest ones, because the biggest chip flows are the hub-routed, related-company ones.

5.3 Taiwan’s customs records

Taiwan gets a closer look for three reasons. It is where essentially every advanced accelerator is made, so its export dataset is the one record of everything that later arrives somewhere, and the manufacturing-side check the destination estimate is compared against (Section 3.5). It is also the one major reporter that no international dataset can check, because Taiwan is not a UN member and does not report to Comtrade, so global datasets carry it only as a residual rebuilt from its partners’ declarations. And it sits on more of the remaining disagreement than any other country, with $2.3 trillion of checkable trade and six of the ten largest unresolved corridor gaps.

As a source it is among the best we hold, as Section 2.5 showed: the fastest publisher in the collection at 7–10 days after month end, net weight on all of its value, piece counts on about a third, and 11-digit national detail. Its tariff tree carves out no AI-specific lines, so all of this hardware rides the standard six-digit codes.

Taiwan’s declarations agree with partner records more often than the collection average. Weighted by dollars, 83% of Taiwan’s trade either mirrors outright (60%) or fails only on the related-company price gap while the goods themselves agree (23%). The comparable figure for the collection as a whole is 66%, and only 6% of Taiwan’s trade is genuinely unexplained. Its export side is the stronger half at 89% reconciled, and its memory exports reconcile at 94% on value alone, with no pricing story needed.

Taiwan’s use of 8471.80 accounts for a large share of its code-level disagreement with importers. It files roughly half of its basket exports under 8471.80, “other ADP units”, a code most importers barely use. That is the single largest code disagreement in the dataset, and it is why comparing Taiwan’s flows one code at a time misleads. Against the US, the individual codes range from 0.18× to 2.90×, while the family total agrees to within 4%.

Taiwan’s declared export destinations are weaker evidence of final location than of departure. The destination known at export can differ from the country where the hardware eventually remains. The estimator therefore prefers destination import records where available. Elsewhere it uses selected exporters’ declarations, including Taiwan’s. These estimates have weaker coverage and cannot establish local retention. They may also contain overlapping declarations when an original exporter names the final destination and an intermediary reports the same shipment as a re-export.

Korea’s memory exports to Taiwan (8542.32), which include HBM used in accelerators, remain unreconciled after these adjustments. That corridor disagrees by 1.32× on value and by 2.14× on weight, and those two point in opposite directions from every other memory corridor we hold. It is the only large corridor in the project where the physical evidence itself disagrees. Taiwan’s imported-memory line reconciles at just 6%, with more than half genuinely unexplained. Korea publishes no piece counts, so we cannot yet separate a weight-convention artifact from real under-declaration. Comparable Korean piece counts would help distinguish these explanations.

5.4 Why we do not use BACI’s reconciled data

Section 2.2 explained why raw Comtrade cannot answer this project’s question. The obvious next candidate is BACI, CEPII’s reconciled trade database, which is built on Comtrade and is the standard dataset in empirical trade research.

BACI addresses exactly the problem this section measures. Where two countries declare the same flow differently, it produces one number: it removes estimated freight and insurance from the import side, scores each country’s reporting reliability against all of its partners, and publishes a single value per exporter, importer, product and year. Because we hold both raw declarations, we can watch that reconciliation work on our own basket. We joined BACI’s current release (HS22, version 202601) to our data for 2022 to 2024, which is about 26,000 combinations of corridor, code and year, and $3.5 trillion of trade.

On the basket as a whole, BACI is a reasonable compromise between the two countries’ records. Its value falls between the two national declarations on 75% of the joined dollars. It tracks the importer’s figure more closely than the exporter’s, landing within ±20% of the importer on 53% of dollars against 39% for the exporter, which is a sensible default because imports are the better-policed direction.

On the corridors this methodology relies on, the two countries’ records differ for the structural reasons of Sections 5.1 and 5.2, and averaging them produces a figure that matches neither the shipping record nor the origin record.

The table illustrates three limitations. First, reconciling individual codes can understate trade when countries classify the same goods differently. BACI reconciles one product code at a time. That works when both countries file the same goods under the same code. Taiwan and the US do not: Taiwan puts servers under 8471.80 that the US books as parts under 8473.30. So on each of those codes BACI has to choose between two figures that are far apart, and on both it lands near the smaller one: $12 billion where Taiwan declared $45 billion, and $23 billion where the US booked $48 billion. Adding its codes back up gives $65.9 billion for the server-and-parts family, against $94.6 billion in Taiwan’s records and $91.0 billion in the US’s. The reconciled number comes out below both of the numbers it was reconciling.

The same pattern appears on other corridors. We checked it across every corridor where the two countries agree at family level, which is the closest thing to ground truth this data offers. BACI recovers 94% of the agreed value overall and 86% on servers and boards, with about a fifth of the agreed dollars landing more than 20% below both countries’ records.

The second is what happens to hub flows. BACI correctly refuses to credit Hong Kong with $205 billion of re-exported processors, adopting China’s near-zero instead. But the goods do not fully reappear where they came from. For Taiwan to China it averages the $243 billion of Taiwan-made arrivals that China booked against the $5 billion Taiwan declared sending directly, and lands at $91 billion, which is 63% below what China’s customs actually recorded arriving. Most of the re-exported value goes missing rather than moving to the right corridor. We keep the two declarations apart instead, and credit hub flows using the hubs’ own re-export data.

The third is what happens where the two countries genuinely disagree about goods. On the Korea-to-Taiwan memory corridor in the last row, BACI splits the difference on a disagreement whose cause is still unresolved on physical evidence (Section 5.3). The available evidence cannot validate the reconciled figure.

BACI reconstructs Taiwan’s trade from partner declarations. BACI has no Taiwan reporter, so its code 490, “Other Asia, not elsewhere specified”, is rebuilt from what Taiwan’s partners declare, and CEPII’s own FAQ recommends it as a Taiwan proxy. On this basket the import side of that proxy holds up well, at 0.96 times Taiwan’s own customs records. The export side is the one that matters for AI hardware, and it runs at 1.31 times Taiwan’s own declarations, with a median corridor discrepancy of 1.48× and only 40% of dollars within ±20%. The reconstruction inherits its partners’ view of Taiwan’s exports.

In a historical test using the earlier quarterly-content estimator, averaging importer and exporter declarations cut China’s customs dollars so far that they could no longer cover the China-specific chips that chip sales records as lawfully sold there; the typical-quarter match fell from 1.03 to 0.86. Current measured-country flows therefore retain the destination’s own records. The exceptions are the UK’s reconstructed 2021 EU-import history and associated indices, and the 50% Malaysia server-gap deduction (Section 3.2). These are targeted adjustments rather than a general averaging of importer and exporter declarations.

The last limitation is timing. BACI is annual and comes out each January. Version 202601 is the first to include 2024, and CEPII warns that the newest year is revised substantially in the following release. In the earlier quarterly-content diagnostic, 74% of customs-derived chip dollars over 2023Q3–2026Q1, about $240 billion of $325 billion, arrived in quarters after December 2024. None of that is in BACI yet, and it will not be until 2027.