Analyst Ben Thompson: TSMC's risk is not in China, but in the AI industry being forced to hit the brakes.

When the market talks about TSMC (Taiwan Semiconductor Manufacturing Company) risks, it often points directly to the geopolitical situation in the Taiwan Strait and military threats from China. However, technology strategy analyst Ben Thompson pointed out that the real risk that has already occurred and is being paid for is not war, but rather TSMC's highly conservative capital expenditures, which are becoming a structural bottleneck for the global AI industry.
Monopoly actually makes Beijing more eager to eliminate TSMC.
Recently at the Davos Forum, Anthropic CEO Dario Amodei publicly opposed the U.S. allowing Nvidia to sell advanced AI chips to China, even describing this move as akin to selling nuclear weapons to North Korea. But Thompson believes that the biggest difference between AI and nuclear weapons is that AI has a physical dependency point that cannot be ignored: Taiwan. Almost all of the most advanced AI chips are manufactured by TSMC, and even without invasion, as long as missiles destroy key wafer fabs, the global AI supply chain could be paralyzed.
If the United States attempts to monopolize overwhelming AI advantages, then China's best strategy will become to eliminate TSMC. Therefore, his stance is contrary to the mainstream hawks. Allowing China to partially rely on TSMC for AI is, on the contrary, a stabilizing strategy.
Root of the problem: TSMC did not bet early during the ChatGPT moment.
Ben Thompson pointed out the harsh reality risk that the AI industry has already incurred huge opportunity costs. In the most recent quarterly report, nearly all tech giants gave the same statement: AI demand > supply.
Amazon: Computing power is immediately consumed once launched.
Microsoft: Azure AI is once again in short supply.
Google: Still under tight supply and demand before 2026.
Meta: Always underestimating internal demand for computing power.
TSMC's chairman, Wei Zhejia, also stated frankly: "The bottleneck is not power, but silicon from TSMC."
Ben Thompson pointed out that the release date of ChatGPT was November 2022, while TSMC's CapEx grew significantly in 2021 (mainly due to COVID and 5G), and capital expenditures will actually decrease in 2023 and 2024. He noted that when major giants fully believe in AI, TSMC does not. This makes TSMC effectively a brake on AI development.
Even if they increase investment now, it will not take effect until 2028.
TSMC 2025 CapEx: $52 billion to $56 billion, with a year-on-year growth rate of +37%, indeed starting to take action. But the problem is that new wafer fabs will take 2 to 3 years to reach production, with real ramp-up occurring in 2028 to 2029. Meanwhile, the capital expenditure growth rates of Amazon, Microsoft, Google, and Meta are still far higher than TSMC.
Ben Thompson emphasized that TSMC's conservatism is not wrong, but rather an extremely rational choice. Nearly all costs in wafer foundry are CapEx; once demand falls empty, depreciation becomes a long-term loss. What Wei Zhejia fears most is that over-investment will cause disaster, but this risk has not disappeared; it has instead been transferred to tech giants.
Only solution: Cultivate competitors to make TSMC feel the competition.
Ben Thompson believes that asking TSMC to expand production is useless; only competition will change its behavior. The real solution is to make Samsung or Intel viable alternatives, so that TSMC is not only worried about losing revenue but also about losing customers. However, the reality is that introducing new processes takes several years and carries extreme risks, while TSMC's services and yield rates remain the industry ceiling.
This article Analyst Ben Thompson: TSMC's risk is not from China, but from the AI industry being forced to hit the brakes. First appeared in Chain News ABMedia.