The Decoupling That Wasn't

The word “decoupling” entered the technology policy vocabulary around 2019 and has since been used to describe a process that is simultaneously real, incomplete, and persistently misunderstood. The argument for decoupling — that the US and China should reduce their mutual technological and economic dependencies to mitigate strategic risk — is coherent as an abstract proposition. The actual implementation has produced not a clean separation but a tangled reorganization that has created new dependencies while reducing some old ones.

This is not a failure of policy. It is a consequence of how supply chains actually work, which is very different from how policy documents describe them.

Consider the semiconductor supply chain, which is the primary target of decoupling efforts. The explicit goal has been to prevent advanced chip technology from reaching China. The actual structure of the supply chain makes this more difficult than a bilateral trade restriction implies.

An Nvidia GPU contains components from multiple countries. The chip die is fabricated at TSMC in Taiwan using equipment from ASML in the Netherlands, Applied Materials in California, Tokyo Electron in Japan, and Lam Research in California. The packaging uses substrates from Japanese and Korean suppliers. The memory is from Samsung in Korea or SK Hynix in Korea. Final assembly may occur in multiple locations. The software stack draws on open-source contributions from researchers globally, with significant Chinese academic contributions.

Decoupling this supply chain means deciding at what level of granularity the separation should occur. If China cannot buy the finished GPU, can it buy any of the components? Can it buy the equipment used to make components? Can it access the software libraries? Can its researchers contribute to the open-source frameworks? Each answer to each question has implications for both the effectiveness of the control and its economic cost to the countries imposing it.

The US has progressively moved its restrictions further upstream — from finished chips, to chip-making equipment, to the chemicals and gases used in fabrication, to the electronic design automation software used to design chips. Each step upstream increases the control’s potential effectiveness and increases its cost to American companies with supply chain relationships that the control disrupts.

The intermediary problem has grown at each step. Countries and companies that are not subject to the restrictions have found themselves in positions where they can, legally or through ambiguity, route technology through channels that complicate enforcement. Singapore, Malaysia, and several Gulf states have seen significant increases in semiconductor imports from the US and other restricted-technology suppliers — imports that do not all appear to end up in those countries’ domestic consumption.

The Bureau of Industry and Security, the Commerce Department office responsible for enforcing export controls, published data in 2025 showing that it had identified 57 companies in 17 countries that were acting as re-export intermediaries for controlled technology to China. The investigation led to additions to the Entity List, license denials, and several criminal prosecutions. It also demonstrated that the enforcement infrastructure for a global technology control regime is significantly less robust than the policy documents imply.

The deeper decoupling problem is at the application layer — the software, services, and data ecosystems that are less tractable to export control than physical hardware. TikTok, ByteDance’s video platform, was the subject of American legislative and regulatory pressure for years, culminating in a forced divestiture requirement passed in 2024. The platform continues to operate in the US under circumstances that remain legally contested. The underlying question — whether a Chinese-owned data platform constitutes a national security risk that justifies the disruption of free speech and commerce norms — has not been resolved in either direction.

The TikTok case is instructive about the limits of decoupling at the application layer. The platform is not a technology product in the semiconductor sense. It is a software product that runs on American cloud infrastructure, serves American users, employs American workers, and earns American advertising revenue. Decoupling it from the Chinese parent company is not technically impossible, but it involves separating software assets from the organizational knowledge, algorithmic infrastructure, and content moderation frameworks that make it function. Those assets are not cleanly divisible.

Similar dynamics apply to other domains where Chinese and American technology ecosystems have become intertwined: open-source AI research (where Chinese and American researchers co-author papers and share code freely), commercial software supply chains (where Chinese contractors contribute to products deployed in sensitive American environments), and academic research partnerships that predate the current political environment and have been wound down unevenly.

The practical effect of incomplete decoupling is not stasis. It is a reorganization of how technology flows across borders — through different intermediaries, with more documentation requirements, at higher transaction costs. The technology still moves. It moves less efficiently, with more friction, through more opaque channels. Whether the friction serves the strategic purpose depends on whether the purpose was to stop the technology flow (which it has not done) or to slow and monitor it (which it has partially accomplished).

There is a historical analogy in the Cold War export control experience. COCOM — the Coordinating Committee for Multilateral Export Controls — maintained technology restrictions on trade with the Soviet bloc for four decades. The restrictions were persistently violated, thoroughly documented by intelligence services, and regularly updated in response to known evasion patterns. They did not prevent the Soviet Union from developing sophisticated military technology. They probably slowed certain aspects of Soviet technological development by years rather than decades. Whether that delay was worth the cost to Western commercial interests is a question historians still debate.

The semiconductor export controls of 2022-2026 may turn out to have a similar historical verdict: real effect, partial enforcement, significant commercial cost, ultimate inadequacy as a comprehensive barrier. The more interesting question is what they are supposed to be a component of, because an export control regime that is not paired with affirmative investment in domestic capability — manufacturing, research, talent — is a restriction without a strategy.

The affirmative side of the ledger has been less consistent. The CHIPS Act committed $52 billion to domestic semiconductor manufacturing, which is real money that has produced real results: new fab construction at TSMC Arizona, Intel’s Ohio expansion, Micron’s New York DRAM facility. But $52 billion, spread across multiple projects over multiple years, is modest relative to the hundreds of billions that the private sector has been investing in AI data centers (which require chips) rather than chip manufacturing capacity. The ratio of investment between “using chips” and “making chips” has not shifted dramatically despite the policy intention.

The decoupling that has occurred is real in the highest-sensitivity domains — frontier AI chips, EUV equipment, certain categories of military technology. It is partial or nonexistent in the broader technology ecosystem. The result is a world that is neither the clean separation that decoupling advocates envision nor the frictionless integration that existed before 2018. It is something genuinely new: a deeply interdependent technology supply chain with an overlay of restrictions, enforcement asymmetries, and political signals that all parties are simultaneously complying with and working around.

That is not a stable equilibrium. It is a managed tension. The management requires constant attention, significant resources, and ongoing negotiations that do not fit neatly into either the “allies versus adversaries” or the “global free trade” frameworks that dominate the policy conversation.

Complexity is not the same as failure. But it is not the clean strategic victory that the decoupling narrative promised, and treating it as one is how strategic failures eventually arrive.