Boo Kok Chuon

In April 2025, shortly after the Trump administration announced a sweeping new tariff regime, I wrote on this Bookcase that tariffs were unlikely to solve the structural problems facing the United States economy, not because I objected to them politically, but because I struggled to see how deliberately making imported goods more expensive could somehow reverse more than two centuries of economic logic built around specialisation, comparative advantage and the allocation of production towards locations where it could be carried out more efficiently. Tariffs could certainly protect selected industries, and in some circumstances governments may even have strategic reasons for doing so, but they could not simply erase the underlying reasons why production had moved across borders in the first place; at best, they would alter relative prices, redistribute costs and force businesses to reconsider how and where they traded.

I also made several more specific observations in Tariffs and Trade Myopia: Trump’s Disregard to 200 Years of Economic Wisdom. I expected China to respond not by simply accepting the loss of American demand, but by redirecting more of its trade towards ASEAN, Europe and other markets. I expected American importers eventually to pass at least part of the tariff burden to consumers, even if some of the cost was initially absorbed through narrower margins. I expected Singapore, being deeply integrated into international trade and supply chains, to feel some of the consequences of the disruption. More broadly, I expected the additional friction introduced into the global trading system to weaken world trade.

Eighteen months later, enough data have accumulated for me to revisit those observations with the benefit of hindsight, and the result is more mixed than a simple victory lap would suggest. Some of the mechanisms I identified have held up rather well, particularly the transmission of tariff costs into prices and the diversion of Chinese trade towards alternative markets. Other parts of the forecast have aged less well, especially my expectations concerning global trade volumes and Singapore’s economic performance. What makes the exercise interesting, however, is that the misses themselves point towards a broader economic lesson: even where the first-order effect of a policy is broadly predictable, the eventual macroeconomic outcome depends on how businesses, investors, engineers and governments adapt around that policy.

What I Got Right

The most straightforward prediction concerned prices. In my April 2025 article, I argued that tariffs could not be imposed on foreign goods as though the cost would simply remain with the exporter, because the burden would eventually be distributed through the commercial chain according to bargaining power, competitive conditions, exchange rates and the ability of firms to absorb lower margins. By April 2026, researchers at the United States Federal Reserve estimated that the tariffs introduced through November 2025 had raised core goods PCE prices by approximately 3.1 per cent and had contributed about 0.8 percentage points to core PCE prices overall.^[1] The transmission was not immediate, because retailers initially absorbed some of the additional costs while consumers remained price sensitive, but the direction of travel was broadly consistent with what one would expect from standard tariff incidence.^[2]

The more interesting prediction concerned China. In April 2025, I argued that Chinese exporters would not simply sit still and accept weaker demand from the United States, because trade relationships are not fixed arrangements and businesses will naturally search for alternative markets when one becomes less attractive. European Central Bank analysis subsequently found that Chinese exports to the United States fell by 20 per cent in 2025, while exports increased by 13 per cent to ASEAN, 8 per cent to the euro area and 26 per cent to Africa.^[3] The figures should not be interpreted mechanically as though every dollar of lost American demand simply reappeared somewhere else, but they do illustrate the broader point that the trading system adapted rather than merely contracted.

That distinction matters because tariff policy is often discussed in static terms. A government raises the cost of importing from one country and then estimates how much less will be imported from that country, but the real economy does not stop at that point. Importers search for new suppliers, exporters cultivate new markets, manufacturers reorganise production, logistics networks change routes and governments respond with new commercial arrangements of their own. The first-order effect may therefore be relatively easy to identify, but the second-order response can substantially alter the eventual outcome.

What I Missed

Where I was less successful was in assuming that all this additional friction would translate relatively quickly into weaker global trade. That did not happen, at least not in the aggregate figures that followed, and the most important reason was the extraordinary expansion of investment associated with artificial intelligence. The World Trade Organization now expects world merchandise trade volume to grow by 3.9 per cent in 2026 and has estimated that AI-enabling goods accounted for approximately 47 per cent of the increase in global merchandise trade by value during the first half of the year.^[4] In other words, while tariffs were applying pressure in one direction, a very large technology-driven investment cycle was pushing much harder in the other.

Singapore illustrates the same point even more clearly. In April 2025, shortly after the tariff announcements, the Ministry of Trade and Industry downgraded Singapore’s growth forecast from 1 to 3 per cent to 0 to 2 per cent, citing the significant deterioration in the external outlook.^[5] At the time, my concern about Singapore’s exposure to global trade therefore seemed entirely reasonable. Yet Singapore ultimately grew by 5 per cent in 2025, with strong performance in manufacturing, wholesale trade and other externally oriented sectors, while AI-related demand became an increasingly important support for the technology cycle.^[6]

My forecast was therefore not wrong because tariffs suddenly became economically harmless. It was wrong because I underestimated the scale of another force operating at the same time. The tariff shock increased friction and altered trade incentives, but the AI investment boom created such strong demand for chips, servers, networking equipment, power infrastructure and data centres that the net global outcome remained expansionary. The lesson is therefore not that the original tariff mechanism disappeared, but that macroeconomic outcomes are the result of several forces acting simultaneously, and one very powerful countervailing force can overwhelm another even where the underlying mechanism remains intact.

There is another part of my 2025 analysis that requires the same degree of caution. I expected less efficient reshoring and trade fragmentation eventually to weaken productivity, yet the aggregate American productivity data have not so far provided clear evidence of such a deterioration. That does not prove that tariff-related inefficiencies do not exist, because those effects may be concentrated in particular sectors or offset by productivity gains elsewhere, but I would be overstating the evidence if I claimed that the United States has already experienced the kind of productivity slowdown I anticipated. In that respect, the forecast remains unproven rather than vindicated.

The Variable I Underestimated: Artificial Intelligence

By October 2026, it is difficult to explain the resilience of Asian manufacturing and trade without acknowledging the scale of AI-related investment. The Monetary Authority of Singapore has itself described global AI investment as a major force supporting industrial production and trade, and much of the recent strength in the region has flowed through precisely the sectors that benefit from that cycle, including semiconductors, electronics, data infrastructure and related services.^[7]

The important qualification, however, is that AI investment should not be treated as a permanently stable offset against the costs of trade fragmentation. The present cycle is unusually large, and it depends on a combination of expectations concerning model capability, enterprise adoption, capital markets, electricity supply, data-centre construction, semiconductor availability and the eventual economic returns that businesses can earn from deploying AI. MAS has already warned that the present AI-related growth impulse is likely to moderate in 2027 and has specifically identified a sharper pullback in AI investment as a downside risk to Singapore’s economy.^[8]

That concern is not equivalent to predicting an AI crash. A transformative technology can continue changing the economy while still experiencing volatile investment cycles, particularly when capital expenditure runs ahead of proven revenue models. Railways, telecommunications and the internet all transformed economic activity, yet none of those transformations proceeded through a smooth and uninterrupted investment path. There is therefore no contradiction between believing that AI will have profound long-term productivity effects and questioning whether today’s pace of spending on GPUs, data centres and associated infrastructure can simply be extrapolated indefinitely.

This is where my May 2026 article on DeepSeek becomes relevant.

DeepSeek and the Problem With Linear Compute Logic

In May 2026, I wrote a Chinese-language article entitled 独家视角：DeepSeek背后的“捭阖之道”, in which I argued that much of the discussion surrounding the semiconductor contest appeared to rely on what I described as a kind of “linear compute logic”. The implicit assumption was that better AI required more GPUs, more GPUs enabled larger models, larger models required more capital expenditure and, therefore, restricting China’s access to the most advanced chips would eventually create an increasingly large technological gap.^[9]

DeepSeek complicated that story because its development placed much greater emphasis on architecture, optimisation and resource efficiency. DeepSeek-V3, for example, contains 671 billion parameters in total but activates only 37 billion for each token through its Mixture-of-Experts architecture, while its technical report repeatedly emphasises efficient inference and cost-effective training.^[10] The significance of this is not that computing power has ceased to matter, because it plainly has not, nor does it prove that engineering efficiency can indefinitely substitute for access to leading-edge hardware. The narrower point is that the relationship between AI capability and capital expenditure is not necessarily linear.

When computing resources become expensive, scarce or strategically uncertain, the value of using those resources more efficiently rises. Engineers have greater incentives to improve architecture, training methods, inference efficiency and system integration, because every improvement in utilisation becomes more economically valuable when hardware cannot simply be added without limit. In that sense, scarcity does not merely constrain technological development; it also changes the direction of innovation.

That is precisely why the current AI investment boom should be treated with some caution when used as a macroeconomic forecast variable. If AI models become substantially more efficient, the amount of capital required for each additional unit of capability may fall. On the other hand, cheaper and more efficient AI could create much more demand and therefore increase total compute requirements anyway. Both outcomes are entirely plausible, which is another way of saying that today’s extraordinary capital expenditure should not be mistaken for a fixed relationship between AI progress and hardware spending.

Huawei and the Other Side of Sanctions

The May 2026 article made another argument that now looks increasingly relevant, namely that restrictions do not merely restrict; they also alter incentives. If a country can purchase advanced technology freely from overseas, there may be very little economic justification for recreating an entire technological stack domestically, especially where the foreign product is cheaper, better and supported by a mature ecosystem. Once access becomes politically uncertain, however, the calculation changes because dependence itself acquires a strategic cost.

In that environment, the relevant question is no longer simply whether the foreign technology remains superior today, but whether it can still be relied upon tomorrow. Investments that previously appeared commercially inefficient may become rational once the possibility of future denial is incorporated into the decision. This is the broader economic context in which China’s investment in domestic accelerators, software tools and AI infrastructure should be understood.

Recent data suggest that this adjustment is becoming more visible. IDC figures reported by Reuters showed Nvidia holding about 34 per cent of China’s data-centre AI accelerator market in the first half of 2026, while Huawei had reached approximately 28 per cent.^[11] More recently, analysis cited by the South China Morning Post suggested that DeepSeek could become increasingly reliant on Huawei’s technology ecosystem over the next few years.^[12]

It would be premature to conclude from these figures that Huawei has achieved technological parity with Nvidia, especially because Nvidia’s competitive position depends not only on hardware but also on CUDA and the broader software ecosystem built around it. The more interesting point is that restrictions designed to preserve a technological gap may simultaneously increase the incentive for the restricted economy to develop substitutes that narrow that gap over time.

That does not mean sanctions are ineffective. They can raise costs, delay development and deny access to capabilities that cannot easily be reproduced. The mistake would be to evaluate them only by examining the immediate technological gap at the moment the restriction is imposed, because the system being restricted is itself adapting to the new constraint.

The Same Economic Story

Looking back, I increasingly think my April 2025 article on tariffs and my May 2026 article on DeepSeek were examining the same economic phenomenon from two different directions. In the first case, the subject was trade. In the second, it was technology. In both cases, the underlying mechanism was the same: once policy changes relative prices, access or incentives, economic actors begin searching for alternative routes.

A tariff raises the relative cost of importing from one country, so businesses search for other suppliers, markets and production locations. A technology restriction raises the relative value of domestic substitutes, even when those substitutes remain technically inferior at the beginning. Scarce computing resources raise the value of optimisation, while uncertainty over foreign technology raises the value of self-reliance. The common thread is adaptation, because the policy does not merely affect the immediate transaction; it changes the incentive map on which subsequent decisions are made.

This sounds almost obvious when stated in abstract terms, yet it is easy to overlook in policy analysis because governments often begin by estimating the first-order effect of an intervention. A tariff makes imports more expensive, a sanction makes technology harder to obtain and a subsidy makes investment more attractive, but the real economy does not remain frozen after the first-order effect occurs. Exporters find new customers, importers change suppliers, engineers redesign systems, investors reconsider where capital should flow and governments themselves respond with new policies. Once those responses accumulate, the eventual outcome can look very different from what a static analysis would suggest.

Singapore and the Other Side of Exposure

My mistake on Singapore also looks different when viewed through this lens. I saw our openness primarily as a source of vulnerability, which was not unreasonable given how deeply Singapore is connected to global trade, finance, manufacturing and logistics. When the external environment deteriorates, those connections allow shocks to reach us quickly.

What I did not give enough weight to was that the same openness also creates optionality. An economy connected to many markets can suffer when global trade weakens, but it can also benefit rapidly when capital, production or technology investment shifts in its favour. Singapore’s exposure to global networks therefore cuts both ways, and the same economic structure that made us vulnerable to tariff fragmentation also allowed us to participate strongly in the AI investment cycle.

In 2025 and 2026, the positive effect happened to be larger than the negative one. That does not make the original vulnerability disappear, because if AI-related investment merely normalises from its present pace, some of the underlying costs associated with trade fragmentation and duplicated supply chains may become more visible. I would therefore resist both extremes. It would be wrong to say that my original concern about Singapore was vindicated simply because the risk existed, because the realised outcome was plainly much stronger than I expected; it would be equally wrong to conclude that the risk was imaginary simply because another force ultimately outweighed it.

The more accurate conclusion is that openness creates both exposure and optionality, and the balance between the two can change very quickly depending on what is happening in the global economy.

What I Would Change Today

If I were rewriting the April 2025 article now, the main change I would make would be methodological rather than ideological. I would distinguish much more clearly between identifying a structural economic force and forecasting the final macroeconomic outcome. Tariffs increase trade friction, but that does not mean global trade must fall every year after tariffs are introduced. AI investment increases demand for semiconductors and infrastructure, but that does not mean the present pace of AI investment will permanently overwhelm every other economic force. What ultimately appears in GDP, export volumes or market performance is the combined result of several forces operating simultaneously, some of which reinforce each other while others offset one another.

This is why getting the mechanism right does not necessarily mean getting the forecast right. My April 2025 article correctly identified several mechanisms associated with tariffs and trade diversion, but underestimated the scale of the AI investment boom. My May 2026 article, in retrospect, captured another part of the same picture by focusing more explicitly on how technological systems adapt when constraints become binding. Read together, the two pieces point towards a broader lesson that I probably appreciate more clearly now than I did when either article was written.

The Economy Routes Around Obstacles

Perhaps the most dangerous assumption in economic policy is not misunderstanding the immediate effect of a policy, because in many cases that immediate effect is reasonably predictable. Raise the cost of importing something and fewer firms will want to import it through the affected channel. Restrict access to an advanced technology and acquiring that technology becomes more difficult. Subsidise an activity and more capital will flow towards it. The harder question is what happens next, because businesses do not remain where policymakers leave them: exporters search for new markets, importers find new suppliers, investors seek new opportunities, engineers redesign systems and countries build capabilities that may previously have appeared uneconomic.

Some of those adjustments are efficient, some are costly and some create new vulnerabilities of their own, but the system continues moving. My April 2025 forecast underestimated how powerful some of those other forces would become, particularly the scale of AI-related investment and the extent to which it would support trade and Singapore’s economy. The lesson I take from that is not that the original economic logic was useless, but that static analysis is rarely enough when policy changes the incentive structure of a dynamic system.

The economy routes around obstacles, and the most interesting question is often not whether the obstacle works in the first instance, but where the new road eventually leads.

---

Footnotes

1. Robert Minton, Madeleine Ray and Mariano Somale, “Detecting Tariff Effects on Consumer Prices in Real Time – Part II,” Board of Governors of the Federal Reserve System, FEDS Notes, 8 April 2026. The authors estimated that tariffs implemented through November 2025 had raised core goods PCE prices by approximately 3.1 per cent through February 2026 and core PCE prices overall by about 0.8 per cent.
2. Sinem Hacıoğlu-Hoke, Sarojini Malladi and Leo Feler, “The Slow Climb: How Tariffs Gradually Raised Retail Prices in 2025,” Board of Governors of the Federal Reserve System, FEDS Notes, 5 March 2026.
3. Julien Le Roux and Tajda Spital, “Global Trade Redirection: Tracking the Role of Trade Diversion from US Tariffs in Chinese Export Developments,” European Central Bank, Economic Bulletin, Issue 1/2026. Chinese exports to the United States fell by 20 per cent in 2025 while exports increased by 13 per cent to ASEAN, 8 per cent to the euro area and 26 per cent to Africa.
4. World Trade Organization, “AI Boom and Trade Resilience to Middle East Crisis Lift Outlook, but Gains to Be Uneven,” 8 October 2026, and Global Trade Outlook and Statistics, October 2026. The WTO forecast merchandise trade volume growth of 3.9 per cent for 2026 and estimated that AI-enabling goods accounted for 47 per cent of global merchandise trade growth by value in the first half of 2026.
5. Ministry of Trade and Industry Singapore, “Singapore’s GDP Grew by 3.8 Per Cent in the First Quarter of 2025,” 14 April 2025. MTI downgraded the 2025 GDP growth forecast from 1.0–3.0 per cent to 0.0–2.0 per cent following the deterioration in the external environment.
6. Ministry of Energy, Trade and Industry Singapore, “MTI Upgrades 2026 GDP Growth Forecast to ‘2.0 to 4.0 Per Cent’,” 10 February 2026. The Ministry reported that Singapore’s economy expanded by 5.0 per cent in 2025.
7. Monetary Authority of Singapore, Macroeconomic Review, Volume XXV Issue 4, October 2026. MAS described the Singapore economy as expanding at an above-trend pace, buoyed by robust global investment in AI.
8. Monetary Authority of Singapore, “MAS Monetary Policy Statement – October 2026,” 14 October 2026. MAS projected moderation in the AI-driven growth impulse during 2027 and identified a sharper pullback in AI-related investment as a downside risk to growth.
9. Boo Kok Chuon, “独家视角：DeepSeek背后的‘捭阖之道’,” Iconomy Bookcase, 25 May 2026.
10. DeepSeek-AI et al., “DeepSeek-V3 Technical Report,” arXiv:2412.19437. DeepSeek-V3 contains 671 billion total parameters, with 37 billion activated for each token, and the authors report 2.788 million H800 GPU hours for full training.
11. Reuters, “Nvidia’s China AI-Chip Share Falls as Huawei Closes Gap, Report Finds,” 13 October 2026, reporting IDC data showing Nvidia at approximately 34 per cent and Huawei at 28 per cent of China’s data-centre AI accelerator market in the first half of 2026.
12. Ann Cao, “‘Jensen’s Nightmare’: Huawei, DeepSeek Narrow AI Chip Ecosystem Gap with Nvidia,” South China Morning Post, October 2026, reporting analysis by SemiAnalysis.