AI Promises and Chip Precariousness, Policy Recommendations and a Changing World, Concerns and Counterpoints
Summary
- Ben Thompson’s core recommendation is to lift restrictions on finished chips—including leading-edge TSMC production and top Nvidia accelerators—while sharply tightening controls on chip-making equipment and its critical components. ZTE and Huawei showed that denial creates the existential demand needed to “pull a semiconductor industry into existence”; restored access would send Huawei, ByteDance, Alibaba, and DeepSeek back to TSMC and Nvidia immediately, blunting that urgency even if China continues indigenizing in parallel.
- The policy objective is not permanent technological containment but another 10 to 15 years in which both superpowers remain dependent on Taiwan. Ben argues that shared reliance on TSMC once supplied a “mutually assured destruction” deterrent; current policy leaves the US dependent while pushing China toward self-sufficiency, removing one reason not to attack an island only 70 miles from its coast.
- TSMC is the hidden single point of failure beneath the entire AI stack. Nvidia can theoretically be displaced, but AI progress remains dependent on advanced Taiwanese fabrication; COVID’s shortage of $1 power-window chips demonstrated how negligible unit prices conceal enormous fixed investment, irreplaceable process knowledge, and years of lead time. Remove TSMC, Andrew Sharp says, and “everyone is screwed and all this progress grinds to a halt.”
- The acknowledged price is that China could obtain AI systems equal to or better than America’s. Ben accepts that painful risk and argues DeepSeek has already weakened the premise that capable AI can remain contained: diffusion and a balance of power might ultimately prove more stable than monopoly. He finds the second sacrifice even harder—strict equipment controls could eventually let China displace Western suppliers globally over 10, 20, 30, or 40 years.
- A coherent equipment blockade would have to extend beyond EUV scanners to fewer than 100 indispensable components and coordinate the US, Japan, the Netherlands, and Germany. The cited examples include Zeiss EUV projection optics, precision mask stages, and radio-frequency plasma generators. EUV remains the meaningful wall: SMIC may now produce evolved seven-nanometer chips profitably through DUV quad patterning, with reported Huawei yields improving from 20% to 40%, but China has no EUV machine.
- US onshoring policy is aimed at the wrong side of semiconductor economics and is far too small. The CHIPS Act subsidizes supply, whereas Ben says customers and volume must be supported; “pat yourself on the back for $50 billion” misses what may need to be a trillion-dollar initiative. Even a strong Intel 18A process offers Apple or Nvidia too little advantage over TSMC’s two nanometers to make switching worthwhile.
- Andrew’s strongest objection is that renewed chip access may neither change Xi Jinping’s Taiwan calculus nor stop Chinese indigenization, while weakening China now may still be strategically necessary. Ben concedes his advice could be “catastrophically” wrong if China reaches AGI first, but counters that aggressive containment could create a “now or never” incentive to attack TSMC. Both land on the episode’s uncomfortable conclusion: every available policy exchanges one existential risk for another.
Deep dive
1. Geography makes Taiwan a structural flashpoint
Ben begins with his own compromised objectivity: he arrived in Taiwan in 2003 expecting one year, stayed for 22, and built a life and family there. That gives him both “a reason to listen to and dismiss my argument”—a Taiwanese-absolutist bias, but also decades of attention to the issue.
On that first flight, China Airlines’ route map showed zero routes to China; travelers had to connect through Hong Kong, then the world’s busiest air route. Beneath the politics was the enduring fact of an ethnically Chinese island 70 miles from the only coastline of an enormous country with which it remained, in Ben’s hedged wording, “still I think technically at war.”
His test is deliberately unsentimental: if America had one coastline and a geopolitical rival controlled an island directly across it, “we would have a problem with that.” Democracy, chips, and ideology intensify the stakes, but his 2020 “Chips and Geopolitics” argument began with the map.
2. Chip denial created the Chinese industry it meant to contain
The chain starts with ZTE, which Ben accepts had violated restrictions on shipping to Iran. Because American technology is embedded throughout global chip production, Washington could deny ZTE chips and force foreign fabs to comply—a capability backed by US military power and access to its consumer market that effectively killed the company for roughly a year.
Ben warned that this leverage would produce a costly long-run response. China had unsuccessfully pursued a self-contained chip industry since roughly the 1950s or 1960s, while Japan, South Korea, and Taiwan prospered by entering the global chain. As long as Chinese companies could buy near-marginal-cost chips, reproducing the entire stack made little economic sense.
The semiconductor model explains why. Bob Noyce at what Ben thinks was still Fairchild Semiconductor could make a chip for $1 and sell it for 25 cents because profitability came from driving volume across enormous upfront costs, not maximizing initial unit margin. Assembly migrated to Hong Kong, Taiwan, and Singapore for cheap labor; eventually the broader manufacturing chain followed into Taiwan.
Andrew stresses that sanctions also focused Beijing and accelerated billions in subsidies and equipment purchases. Ben’s distinction is sharper: subsidies merely push supply, whereas ZTE—and, around 2020, Huawei—created existential customer demand. Once China’s most important technology company faced destruction, “we’re locked in”; industry had to pull an alternative supply chain into existence.
3. A $1 automotive chip revealed TSMC’s true leverage
During COVID, automakers behaved like supply-chain sovereigns and canceled deliveries. Their inexpensive power-window chips looked disposable, yet each depended on a fab that had been leading-edge around 2002, whose upfront cost was long depreciated and whose capacity nobody would rationally rebuild.
Suppliers found other customers; when car factories restarted into surging demand, automakers went to “the back of the line.” The supposedly trivial $1 component then shut down vastly more valuable production for weeks or months—the cleanest example of how chip prices obscure financial complexity, accumulated expertise, and a non-recoverable temporal investment.
Ben’s lesson from 2020 is that the US economy is “utterly dependent on the continued existence of TSMC,” and therefore Taiwan’s continued de facto independence. Computing also requires trust in whoever sits at the root; replacing Taiwanese dependence with Chinese trailing- and eventually leading-edge supply risks “making China root of everything.”
4. Restored access could rebind China to TSMC
Sanctions prompted Chinese firms to stockpile Western equipment, making suppliers’ China revenue a misleading proxy for sustainable demand. They also allowed SMIC to lose almost certainly billions while mastering seven-nanometer production through difficult DUV quad patterning—losses Intel could not have borne without risking bankruptcy.
Roughly three years into that effort, Ben finds it plausible that SMIC’s evolved seven-nanometer processes are profitable and materially faster than their predecessors. Andrew adds unverified reports that Huawei yields have risen from 20% to 40%; both treat the figures cautiously but see the products reaching market as evidence of genuine progress.
EUV nevertheless remains “a wall.” China has no EUV scanner, while TSMC moved from five to three nanometers, was coming out with two nanometers that year, and expected roughly one-nanometer-equivalent processes later. The names contain marketing, Ben says, but the density, efficiency, speed, and thermal improvements remain meaningful; China will eventually build EUV, yet having even one machine to examine would make that work easier.
Precisely because advanced chips are so valuable, Ben expects Huawei to order TSMC’s leading edge “tomorrow” if permitted, while ByteDance, Alibaba, and DeepSeek immediately buy top Nvidia accelerators. Chinese firms would still pursue independence, but “thank God we finally have advanced chips” would redirect scarce money and attention from fabrication toward competing in AI.
Andrew notes that even while Huawei works with SMIC, TSMC has been accused of fabricating for Huawei subsidiaries, underscoring the continuing appeal of the best available foundry.
5. Competitive AI still rests on one precarious island
Ben calls the AI buildout a “Goldilocks moment” for capitalism: Anthropic, OpenAI, Google, and xAI are all investing; hyperscalers and neoclouds absorb capacity; open models keep lock-in fears bounded; prices fall even as investment remains attractive. H100 utilization reportedly returned to extremely high levels after DeepSeek, echoing increased demand inside China.
He also records a real change of mind. Ben had defended staying with OpenAI’s familiar app despite Anthropic’s perceived quality edge, but now judges Grok “so much better” for general chat, while Deep Research remains amazing and he still likes OpenAI’s app.
DeepSeek establishes what Ben calls a minimum open model: capability at or below that level is simply available, making earlier caution such as withholding GPT-2 look obsolete. “The doomers have to adjust to a reality that DeepSeek exists,” though he explicitly acknowledges that broader diffusion is terrifying if their premises are right.
Beneath this flourishing ecosystem is one dependency more fundamental than Nvidia: TSMC. America values Taiwan’s democracy and fabrication capacity; China wants the island and, in Ben’s stark framing, could still accept it “bombed to smithereens.” That mismatch makes weakening chip-based deterrence particularly dangerous.
6. The proposal knowingly trades away AI exclusivity
Ben does not dispute the Biden-era premise, adopted before ChatGPT, that AI chips matter for national security. His recommendation means China “will have AI systems on par with ours and very well may have better ones.” He does not want that outcome; he is explicitly ranking Taiwan deterrence above a legitimate competing objective.
His uncomfortable analogy is nuclear proliferation: Americans who passed secrets to the Soviet Union were traitors, yet one could argue their actions contributed to peace because mutual assured destruction worked. Likewise, a durable AI balance might be safer than one state assuming it can permanently control a technology with superpower-level implications.
This extends Ben’s prior safety argument, not a dismissal of danger: AI is probably not containable, DeepSeek makes hand-waving about control less credible, and “we need good AI to defend ourselves against bad AI.” General diffusion could therefore be the only sustainable long-run balance, although he presents that as a possibility rather than certainty.
The harder concession for Ben is industrial. Cutting China off from all US equipment puts “the kiss of death” on Western manufacturing-equipment dominance: Chinese suppliers will climb the curve, allied firms may capture abandoned sales, and China could eventually undercut incumbents worldwide over 10 to 40 years. He had mentally accepted AI diffusion; this sacrifice still felt worse.
7. Equipment controls are simultaneously necessary and self-destructive
Andrew had found prior discussion incoherent because policy tightly restricted finished AI accelerators while leaving upstream equipment and components comparatively exposed. If technological dominance is the goal, banning chips while enabling the tools to make substitutes looked “frankly insane and incompetent.”
The SemiAnalysis prescription he cites is stricter upstream control over fewer than 100 critical items—not every nut and bolt. China cannot import EUV scanners, yet controls allegedly omitted Zeiss projection optics, “the number one critical component,” alongside precision mask stages and radio-frequency plasma generators.
Ben grants the Commerce Department one rationale for loopholes: SMIC making seven nanometers on Western equipment is bad, but it delays Chinese equipment learning. The article’s recommendation still closes them, knowingly sacrificing future equipment leadership to maximize dependence on Taiwan while opening downstream chip access.
8. China’s scale is breaking the old world order
Pax Americana worked because the US could use its consumer market to rebuild Europe and Japan, surrender some manufacturing, and guarantee peace. Those smaller economies could absorb only so much generosity. China is different: its geography and much larger population made comparable integration structurally too expensive for a system centered on one generous hegemon.
Ben connects China’s manufacturing rise with hollowed-out American industrial towns and exploding US debt, while acknowledging debates over subsidies, consumption, currencies, and the unusually weak Taiwanese dollar. Zoomed out, however, his answer is blunt: “China’s just too big.”
Apparent Trump-era weirdness—tariffs, aggression within the Americas, indifference to European preferences—looks less bizarre if Washington believes Pax Americana is ending. One alternative shifts US concentration to Asia’s first island chain; another accepts China as the eventual Asian hegemon. Both make Taiwan, the chain’s critical link, the unavoidable flashpoint.
The policy can therefore only buy time. Trading AI advantage and semiconductor-equipment leadership “just for 10 more years, or just for 15 more years” is disgusting to Ben, but waiting until 2025 eliminated cleaner options. “There are no good solutions right now.”
9. Intel needs customers, not merely subsidized factories
Ben’s intentionally extreme recommendations—including a US data center using Gaudi AI accelerators made by Broadcom and Apple, open-source output, and copyright indemnification—were meant to expand the Overton window. Their implausibility carries the subtext: structural dependence on Taiwan may be unfixable in peacetime, so every demand-creation mechanism deserves consideration.
The CHIPS Act has two failures in his telling: it subsidizes supply instead of demand, and it is “dramatically too small.” A $50 billion program invites self-congratulation where the actual requirement may be “a trillion-dollar initiative” capable of making customers endure the cost and friction of adopting a US foundry.
Intel lost more than processor competitiveness when its process lead disappeared; it lost the bridge to foundry demand. Had it pursued foundry seriously around 2010 or 2013, Apple might have tolerated the pain to obtain the world’s fastest chips. Today, even if 18A is marginally better than TSMC’s two nanometers, the advantage is insufficient.
That is why neither Apple nor Nvidia has an economic reason to choose Intel, just as nobody rationally builds redundant trailing-edge capacity while TSMC is available. A semiconductor industry cannot exist because government wishes it into being: “You actually need the customers.”
10. TSMC cannot simply be transplanted to America
From a shareholder perspective, TSMC should be hightailing it out of Taiwan to a certain extent. The Taiwanese government has different incentives: it bars export of the leading-edge and N-minus-one nodes, permitting only N-minus-two. US production was five nanometers marketed as four and would add three nanometers as Taiwan moved toward roughly one nanometer, perhaps called A16—Ben hedges the naming.
That lag preserves the leading edge as Taiwan’s strongest reason to expect US defense. Ben agrees with Trump’s accusation in unusually direct terms: the government is trying to hold America hostage—“yes, they are”—because Washington asking for the industry so it no longer needs to defend Taiwan gives Taipei no reason to cooperate.
Trailing-edge diversification is even uglier economically. Chinese capacity is already crashing prices toward zero; a new legacy fab may never recover construction costs and could lose money operationally. Yet those old nodes support cars and much of the broader economy, while Intel historically repurposed or sold old equipment rather than maintaining TSMC-style legacy capacity.
Nor can TSMC simply inherit Intel fabs. Its transistors differ “all the way down to the molecular level”—chemicals, dimensions, density, and process knowledge all change—while R&D remains in Taiwan. “Everyone else gets a vote”: Taiwan, China, allies, customers, and competitors constrain what Washington can demand.
11. The counterarguments leave no safe side of the trade
Andrew doubts unlimited TSMC access would dissuade Xi Jinping or stop indigenous investment. Taking Taiwan already entails enormous costs, yet Xi may eventually decide they are worthwhile; if so, America should use every available tool now to diminish Chinese power rather than gamble on renewed interdependence.
He also rejects a stable Monroe Doctrine 2.0. China’s behavior over the prior decade, plus recent drills around Australia and New Zealand, suggests pressure would not end with Taiwan; Japan and other Pacific allies would come next, ultimately affecting America. Defending Taiwan therefore serves broader US interests, whether or not the PLA can currently seize it.
Ben doubts that unlimited access alone would sway Xi or stop China from indigenizing. Andrew says he would favor the US in an initial skirmish but is “a little more pessimistic” about a war of attrition. He also argues that China’s home-court advantage strengthens over time, while a near-term US AI breakthrough could make waiting a “now or never” mistake.
Andrew’s AGI objection is that even a future AGI would still rely on Taiwanese chips, so China could answer “Nice AGI you have there” with missiles aimed at Hsinchu. Ben concedes the thesis could be catastrophically wrong—China might get AGI first and invade trivially—but maintains that policymakers cannot ignore the possibility that tighter controls accelerate the very war that destroys TSMC.