World trade is moving to the rhythm of AI orders

China’s exports rose 36.3% year on year in September 2026, measured in current US dollars. Imports increased by 31%. Released on 14 October, the figures suggest a powerful acceleration in cross-border business. Days earlier, the World Trade Organization had raised its forecast for 2026 merchandise trade volume growth from 1.9% to 3.9%. Investment linked to artificial intelligence is central to its explanation. ^{[1] [3]}

To understand the expansion, an order needs to be followed beyond its first buyer. A chip can cross a border, become part of a circuit board, travel again inside a server and support a computing service sold somewhere else. Several companies earn revenue and several economies record exports. Yet their business may depend on the same investment programme. Diversifying trade routes and diversifying ultimate sources of demand offer different kinds of resilience.

The documents assembled for this investigation illuminate that common exposure without providing an exact global measure. They also contain evidence that complicates a simple concentration story. In NVIDIA’s latest quarterly comparison, fewer direct customers exceed the 10% revenue disclosure threshold. Indirect demand remains much harder to observe. Part of the risk lies in the distance between those two views. ^[7]

The sources of trade growth are shifting

The WTO attributes 47% of the increase in global merchandise trade value in the first half of 2026, compared with the same period of 2025, to goods that enable AI deployment. Trade in those goods grew by 67% year on year. The category includes products such as semiconductors and servers. The 47% figure concerns the increase in trade: it identifies the source of momentum. ^[3] The basket represents 14.8% of world merchandise trade value in that period, a different denominator from its contribution to growth.

A category can explain much of an increase while accounting for a considerably smaller fraction of the total. The contribution also depends on what happens to other products. When some trade flows weaken, the contribution of the fastest-growing group becomes larger. This is evidence about the concentration of the recent impulse, which deserves to be distinguished from the composition of the whole economy.

Method and scope

l0g reconstruction: the nominal increase between H1 2025 and H1 2026 is normalised to 100. The WTO basket contributes 47; the remaining 53 is the difference. The 67% figure is growth within that basket; 3.9% is a full-year volume forecast for all merchandise. These denominators are not interchangeable.

Value and volume need to be kept separate too. Dollar-denominated trade combines prices and quantities, exchange-rate effects and changes in the mix of goods sold. Substituting more powerful, more expensive machines for ordinary equipment can raise the trade bill without a proportional increase in the number of units shipped. The WTO’s 3.9% forecast instead refers to a volume index. Putting those percentages on the same scale would obscure their different meanings. ^{[3] [12]}

China’s release provides a practical illustration. The official 14 October briefing transcript reports September export growth of 29.6% in yuan, whereas the dollar-denominated report gives 36.3%. Both are nominal comparisons with the same month a year earlier. The gap calls for attention to currency conversion; it says nothing by itself about the physical expansion of shipments. ^{[1] [2]}

The AI label also needs a scope check. The WTO revised its product list to focus more closely on deployment, removing some general-purpose goods and infrastructure inputs predominantly used outside AI. The basket remains a selection of customs products, rather than an itemised register of each final user’s purchases. A customs classification identifies the object shipped more readily than its eventual use. ^[3]

Components travel through the new destinations

Chinese customs reports that intermediate goods accounted for more than half of export value in the first nine months of 2026. These are goods used in a further stage of production. The same briefing reports increased trade with more than 180 economies, a measure covering imports and exports together. These government statements describe the reach of the network and China’s role in successive stages of production. The distribution of ultimate demand remains a separate question. ^[2]

A business shipping components to a Southeast Asian assembler may have changed its customs destination while still serving orders from the same final markets. Another may genuinely have found new local customers. Rising shipments to a region can accommodate both stories. Distinguishing them requires links between products, importing companies, processing activities and subsequent exports.

Taiwan’s figures show why the distinction matters. September exports reached $87.222 billion, up 60.9% on the year. Two groups, information, communication and audio-video products, and parts of electronic products, together accounted for $70.259 billion, or 80.6% of total exports after rounding. These categories extend well beyond AI. They nevertheless establish the scale of Taiwan’s electronics specialisation. ^[4]

A separate breakdown shows the United States taking 31.5% of exports, mainland China and Hong Kong 23.6%, and ASEAN 21.6%. ASEAN is the association of Southeast Asian nations. These are trade destinations: goods can be consumed, installed as productive equipment or processed further there. The attachment does not supply the detailed cross-tabulation needed to connect each product category to each destination. ^[4]

Method and scope

September 2026, preliminary data, current US$ million. Total 87,222. Electronics groups: 39,128 + 31,131 = 70,259, or 80.5519%. Other products: 16,963. United States: 27,461; mainland China and Hong Kong: 20,622; ASEAN: 18,881; other destinations: 20,258. Both bars use the same denominator; no product–destination cross-tabulation is inferred.

Our analysis of semiconductor bottlenecks separates fabrication, memory and assembly: rising shipments do not establish that every production stage has equal spare capacity.

This changes the design of a resilience test. A local transport disruption may be bypassed by reorganising routes. A slowdown in one investment category can affect several routes at once. The first question concerns alternative capacity. The second concerns the independence of the demand being served.

One investment can leave several customs records

Consider a wholly fictional supply chain, using one common monetary unit. A manufacturer in economy A produces a chip with 40 of value added and exports it to B for 40. The business in B adds 30 through labour, services and operating surplus, then exports a board to C for 70. Server integration in C adds another 30. The completed server is sold for 100 to a business in D, which uses it as productive equipment.

Assume that all transactions occur in the same period, with no inventories, taxes or other transactions. Gross exports sum to 210: first 40, then 70, then 100. Final investment expenditure is 100. Value added is distributed as 40, 30 and 30 across the three producing economies. Each border correctly records the value of the product passing through it; previously incorporated value travels with the product.

Method and scope

l0g simulation: four economies, one period, no taxes, inventories, separately priced transport or other intermediate purchases. Export prices are 40, 70 and 100; value added is 40, 30 and 30. The initial producer is assumed to use no purchased input. The final server is investment in economy D. No actual supply chain is quantified.

Within the example, every transaction is genuine. Each pays a business for productive activity. The analytical mistake would be to treat the three shipments as three independent final purchasing decisions. A change to the project in D can reach suppliers in A, B and C. The chain distributes income during the expansion, then transmits adjustments when its ultimate market changes pace.

This is the problem addressed by the OECD’s Trade in Value Added, or TiVA, approach. It reconstructs contributions by economies and industries using tables of their purchases from one another and of final demand. Final demand includes consumption and investment in equipment. This provides a view of interdependence that gross shipment records alone make difficult to interpret. ^[5]

The 2025 edition described in OECD country notes covers 1995–2022. In April 2026, the OECD also published experimental estimates for 2023 and 2024. These newer data remain aggregated; they do not reconstruct AI users’ September 2026 orders. Estimating a precise fraction of Chinese exports dependent on a few investors would require additional information and an explicit updating method. The fictional chain explains an accounting identity without estimating that fraction. ^{[6] [13]}

There is also a scope limitation: standard TiVA treats a server purchase as final investment. Tracing the value added embodied in the equipment does not, by itself, identify later users of computing services. A new OECD framework extends the analysis to capital used in production, without identifying private customers in 2026. ^[6]

Our analysis of the US trade deficit, from the port to the factory follows the same distinction between equipment purchases, inventories and domestic output.

The visible buyer and the demand behind it

NVIDIA’s quarterly filing provides evidence at another scale. For the three months ended 26 July 2026, its second quarter of fiscal 2027, the company reports that one direct customer accounted for 16% of revenue. A year earlier, two customers represented 23% and 16%. US accounting rules require information on customers accounting for at least 10% of revenue, without requiring their identities. This is the benchmark used in our chart, rather than a map of every buyer. ^[14] The relevant customers are unnamed in this section. ^[7]

The comparison shows fewer large direct customers in the published shares above that threshold. It does not measure a reduction in overall concentration. The latest disclosure weakens a story based solely on growing concentration among major direct buyers. Shares below the threshold can be distributed in many ways: an “other customers” segment in a chart represents several buyers, never a single one. These disclosures cannot produce a complete concentration index or establish which customers are the same across years.

The filing adds information that ordinary trade accounts would struggle to show. By NVIDIA’s estimate, one AI research and deployment company contributes meaningfully to its revenue by purchasing cloud services from its customers. The company thus distinguishes direct purchases from indirect demand. It provides neither an identity nor a usable percentage for this particular customer. ^[7]

Method and scope

Total revenue; Q2 FY2026 ended 27 July 2025 and Q2 FY2027 ended 26 July 2026. Disclosed direct shares: 23% and 16%, then 16%. Calculated remainders of 61% and 84% contain multiple customers. Names and continuity across years are unknown. The lower network is illustrative; provider count, shares and user identity are not estimated.

A relationship diagram can explain the mechanism without assigning any role to a real company. One user spreads its computing requirements across several cloud providers. Those providers purchase equipment through manufacturers or integrators, or directly from a processor designer. There are several buyers at the first invoicing layer. Further downstream, part of their business may rely on the same programme. Conversely, one large direct customer can serve thousands of users with unrelated needs. Both possibilities have to remain visible.

NVIDIA’s statement establishes a measurement problem and a disclosed indirect dependency. It leaves the exact share of that demand, its stability and the contractual allocation of adjustments unknown. It cannot establish one universal customer structure for the entire global industry.

A further convention matters when matching sources. NVIDIA assigns geographic revenue according to the headquarters of its direct customer. Shipment locations and ultimate users may be elsewhere. Mechanically matching this table to Taiwanese or Chinese customs data would create spurious precision. The sources become useful together when their respective units of observation are preserved. ^[7]

Inventories can magnify a change in pace

An order also passes through inventory decisions. Companies buy to meet expected sales and maintain a buffer. When both objectives change, the movement reaching a supplier can be larger than the change observed at the final customer. Daniel Rees and Phurichai Rungcharoenkitkul describe this amplification mechanism, often called the bullwhip effect, in their 2021 BIS bulletin on bottlenecks. The text expresses the authors’ views. That episode illustrates a logic, rather than supplying a coefficient that can be applied automatically to AI in 2026. ^[9]

A second fictional example isolates the arithmetic. An assembler normally sells 100 units per period. It starts with 40 in inventory and wants to finish with 40, so it receives 100 from its supplier. For the next period it expects 90 sales and chooses to reduce its ending buffer to 20. With 40 already available, it needs new purchases of only 70. Receiving 70 and selling 90 indeed leaves 20.

Sales fall 10%; purchases fall 30%. The difference comes from the additional reduction in desired inventory. Had the business retained its 40-unit target, it would have bought 90. The example therefore contains two explicit decisions: a lower sales forecast and a smaller buffer. No comparable movement is being attributed to an actual company.

Method and scope

l0g simulation in homogeneous physical units, over one period, with no lead times, losses, returns or outstanding orders. Closing inventory = opening inventory + purchases received − sales. Baseline: 40 + 100 − 100 = 40. Adjustment: 40 + 70 − 90 = 20. Retaining the original buffer would imply purchases of 90. BIS documents the general mechanism, not these parameters.

Each stage can then respond to its own information. Some contracts cushion changes in the short run; others permit deliveries to be deferred. Shipments can continue even after new orders weaken. Customs statistics record goods actually crossing a border, potentially well after the investment decision that set them in motion.

The timing matters when reading industrial announcements. Delivered equipment, reserved capacity and a purchase planned for a future period occupy different places in the chain. Combining them into one headline amount erases the points at which schedules can still change. An investigation of order books needs delivery dates, deferral clauses and inventory movements, rather than an assumption that current revenue fixes future demand.

From technology orders to macroeconomic exposure

At its 13 October briefing, presented on a page published the following day, the IMF describes the technology boom as a demand impulse, particularly evident in trade and investment. It also highlights supply-chain concentration across countries, industries and companies. That structure allows a localised impulse to travel far, while creating common dependencies. ^[10]

The potential transmission can be traced without forecasting a crisis. Suppose several purchasers revise their capacity needs. New orders slow; assemblers then adjust procurement and inventories. Suppliers defer some of their own investment. Depending on the local value-added share, wages, profits or tax receipts become weaker than expected. In an exporting economy, foreign-currency receipts may also slow, with the consequences depending on associated imports, available financing and other export industries.

The important word is common. A company can diversify direct customers, and an economy its destinations, while retaining exposure to the same shift in technology spending. The exposure may be large or small; available accounts do not measure it fully. The case of Crux AI’s chip collateral and bank risk examines another connection: contracts and expected revenue supporting repayment of financed equipment. The implication is to test together activities whose revenues depend on the same ultimate decisions.

Method and scope

Conditional diagram developed by l0g from mechanisms discussed by the IMF and ECB. Lines indicate relationships, not measured volumes. Local income depends on value added retained; currency and funding effects depend on other flows. No macroeconomic multiplier or crisis scenario is calibrated.

There is also a path towards broader demand. Different users served by the same infrastructure can gradually stabilise revenue. Better software and organisation can make installed equipment more productive. In his 23 March speech, Philip Lane emphasises the role of AI diffusion in delivering productivity gains. The pace of hardware purchases and the pace of economic benefits can consequently follow different trajectories. ^[11]

Evidence that would change the assessment

The current expansion deserves to be taken seriously. NVIDIA announced $96.2 billion of revenue for its quarter ended in July. Equipment has been sold and capacity deployed; some of that investment may deliver useful services for years. September’s trade increases and the WTO revision are also documented outcomes or forecasts, on their respective scales. ^{[8] [1] [3]}

The favourable scenario would gain credibility from sustained growth in the user base, equipment actually put to work and benefits spreading beyond infrastructure suppliers. Stronger activity in sectors with little connection to the technology cycle would also diversify the expansion. Productivity improvements could support income or reduce prices, depending on competition and the distribution of gains. Those channels can extend growth beyond the initial installation phase. ^[11]

A concerning signal would be more specific than one weak export month: converging cancellations or deferrals, inventory accumulating despite weaker expected sales, followed by lower orders at suppliers across several economies. Each observation would require its own scope check. A new hardware generation, price changes or a delivery problem can produce temporary movements quite different from a broad weakening of final demand.

The investigation ultimately proposes a way of reading the expansion. Borders show where goods move. Company accounts show who receives the invoice. Assessing resilience requires following the uses, commitments and decisions that trigger those purchases. World trade can acquire new routes while remaining tied to the same rhythm of orders.

Sources and method

Documentary research closed on 14 October 2026. Chinese figures come from the official transcript reproduced by Beijing Daily and indexed Reuters republications. The original Reuters dispatch and detailed customs table were inaccessible; no independent audit of the collection process is claimed. The Taiwanese statistical attachment was read and visually checked across all three pages. NVIDIA’s disclosures are company publications; its quarterly statements are unaudited and its assessment of indirect demand remains its own estimate.

The original WTO release and executive summary were consulted, along with the full report’s AI-enabling-goods chapter, printed pages 30–32. Its calculations average available reported imports and exports, with mirror data for Vietnam and Russia. The WTO technical notes explain the value/volume distinction. The IMF’s complete briefing-page text was read; the full WEO report is not presented as having been consulted. The full BIS bulletin was also read. Conflicting press accounts of some detailed Chinese product categories led to their exclusion. No interviews, access to private contracts or identification of unnamed buyers were undertaken. Figures 3 and 5 are wholly hypothetical. Figure 6 shows conditional transmission channels, with no probability or numerical forecast attached.

1. Reuters, China’s exports surge 36.3% y/y in September, imports up 31%. 2026-10-14. Chinese customs data reported by Reuters: annual changes in current US dollars. Press cross-check, not an independent audit of customs collection. Indexed passages from the Reuters republication on Investing.com consulted; original dispatch inaccessible. Supporting document.
2. SCIO / Douanes chinoises, Conférence de presse sur le commerce extérieur des neuf premiers mois de 2026. 2026-10-14. Transcript of the official 14 October briefing at 10 a.m., reproduced by Beijing Daily at 12:54 p.m. Chinese government source: yuan-denominated data and intermediate-goods share. The original detailed customs table was not obtained.
3. WTO · AI boom and trade resilience to Middle East crisis lift outlook, but gains to be uneven. 2026-10-08. Original release and executive summary consulted. The full report, printed pages 30–32, distinguishes the 47% contribution to the nominal increase, 67% growth within the basket and 14.8% share of the total. Volume forecasts of 3.9% and 4.1% concern 2026 and 2027. Revised HS6 list; average imports/exports, available TDM reporters, mirror data for Vietnam and Russia.
4. Ministère des Finances de Taïwan / Taiwan Ministry of Finance, Exports and Imports in September 2026. 2026-10-08. Release at 4 p.m. local time; all three pages of the attachment read and visually checked. Tables in millions of current US dollars. Product groups and destinations are separate breakdowns of the same total. Supporting document.
5. OCDE / OECD, Trade in value added. 2026-10-14. Methodology page accessed on 14 October 2026; this is an access date. Domestic and foreign value added, final demand and inter-country input-output tables.
6. OCDE / OECD, Trade in Value Added – Country notes: United States. 2026-06-01. Overview of the 2025 TiVA edition, covering 1995–2022. Used for scope and time lag, not to extrapolate a global AI share in 2026. Seeing the capital behind trade, 2026-09-18, explains standard TiVA’s treatment of capital.
7. NVIDIA / SEC EDGAR, Form 10-Q, quarter ended July 26, 2026. 2026-08-26. Unaudited interim financial statements; Note 13, printed pp. 23–24, and Concentration of Revenue, p. 29. Direct-customer shares, estimated indirect demand and geography by direct-customer headquarters. The relevant customers are unnamed.
8. NVIDIA, NVIDIA Announces Financial Results for Second Quarter Fiscal 2027. 2026-08-26. Company release: quarterly revenue rounded to $96.2 billion. The SEC filing supplies detail; this communication is not independent confirmation.
9. BRI / BIS, Bottlenecks: causes and macroeconomic implications, Bulletin 48. 2021-11-11. Full bulletin read, particularly printed pages 1 and 3. Daniel Rees and Phurichai Rungcharoenkitkul express their own views on inventory amplification and cross-border propagation. No historical elasticity is applied to the 2026 AI cycle. Full bulletin.
10. FMI / IMF, World Economic Outlook press briefing. 2026-10-14. Page published on 14 October covering the 13 October briefing: technology demand impulse and concentration across countries, sectors and firms. Full briefing-page text consulted; the complete WEO report is not presented as having been read.
11. BCE / ECB, AI and the euro area economy, Philip R. Lane. 2026-03-23. Original speech: technological diffusion, productivity, investment and the timing of demand. Discussed scenarios are not realised outcomes for the euro area as a whole.
12. WTO · Technical notes, international merchandise trade statistics. Accessed 2026-10-14; no publication date assigned. PDF pages 5–6: current values, deflated volumes and unit values; customs primarily record goods actually crossing borders.
13. OECD · Global value chain repositioning. 2026-04-08. Experimental TiVA estimates for 2023 and 2024 update the standard series, while remaining aggregated; they do not identify September 2026 AI buyers. Full report.
14. FASB / FASAC · Meeting materials, 25 September 2018. 2018-09-25. Summary of US GAAP requirements, paragraph 280-10-50-42, printed page 22, PDF page 47: disclosure for customers representing at least 10% of revenue, without mandatory identification. General accounting convention, distinct from NVIDIA’s reported quarterly shares.