The race is on
When everyone in the supply chain simultaneously starts trying to lock down their piece of it, tools become export controls, investment restrictions, and industrial policy — and everyone is using all of them at once.
Every major player in the AI supply chain is simultaneously trying to lock down their own piece of it, deny others access to critical capabilities, and build domestic alternatives to the things they currently depend on. The result is a supply chain that was built on the logic of global specialisation — and is now being used as an advantage by multiple parties, in multiple directions, at the same time.
Dependency as advantage #
The mechanism is simple. A chokepoint is a vulnerability if you are the buyer. It is a potential advantage if you are the seller — provided you are willing to pay the cost of using it.
For most of the history of the global semiconductor industry, that cost was considered too high. Suppliers did not want to lose customers. Governments did not want to disrupt trade. The integrated supply chain produced better products at lower costs for everyone. The incentive was to keep it running.
That calculation began to change around 2018 and broke decisively by 2022. The reasoning, simplified:
If the country you are supplying is building the technology that will determine military and economic dominance for the next generation, the customer relationship becomes a strategic subsidy to a competitor.
Once governments accepted that framing, the integrated supply chain stopped being a shared asset and became a map of pressure points.
The US offensive #
The United States fired the first major shot in October 2022, when the Bureau of Industry and Security published rules that fundamentally changed how semiconductor technology could flow to China.
The rules targeted three things simultaneously:
- Advanced chips: GPUs and other AI accelerators above specific performance thresholds could no longer be exported to China without a licence.
- Manufacturing equipment: Tools for making chips at advanced nodes — including equipment from US companies Applied Materials, Lam Research and KLA — were restricted
- Talent: US persons were prohibited from supporting advanced chip manufacturing in China without authorisation
The October 2022 rules were the opening move. They were followed by expansions in October 2023 and further tightening in 2024, as chip designers found ways to engineer products that met the letter of the original thresholds while preserving most of the restricted performance. Each update closed a loophole and extended the perimeter.
The entity list — a register of specific companies denied access to US technology — expanded to include Chinese chip designers, foundries, and AI companies. Being added to the entity list is, for a semiconductor company, a major slowdown. You cannot buy the equipment. You cannot license the software. You cannot access the IP.
The stated goal was to deny China access to the most advanced chips and the tools needed to make them — and to slow Chinese progress on AI systems with potential military applications.
Pulling in the allies #
The US restriction on US-made equipment was the straightforward part. The more complicated requirement was getting allied countries to impose equivalent controls — particularly the Netherlands (ASML) and Japan (Tokyo Electron and others). This took sustained diplomatic pressure. The Netherlands and Japan were reluctant. Their companies had significant China revenues. Their governments were wary of economic retaliation. Their industrial policy was oriented toward exports instead of restrictions.
By late 2022 and into 2023, both countries had agreed to restrict exports of advanced semiconductor equipment to China. The Netherlands barred ASML from shipping its most advanced EUV systems to China — ASML had never shipped EUV to China, but the restriction was formalised and extended. Japan announced restrictions on 23 categories of semiconductor manufacturing equipment in March 2023. The coordinated effect was to extend the US-designed perimeter around advanced chipmaking to cover the key equipment suppliers in allied countries. No single action was decisive. The combination was designed to be.
China's countermoves #
In July 2023, the Chinese government announced export controls on gallium and germanium — two materials where China accounts for a dominant share of global production. Gallium is used in compound semiconductors (gallium nitride, gallium arsenide) that appear in power electronics, radar, and communications equipment. Germanium is used in fibre-optic systems, night-vision equipment, and some semiconductor substrates. The restrictions required export licences. In practice, licences were selectively granted. The signal was clear: China has chokepoints too, and is willing to use them.
In October 2023, graphite was added to China's export licence list — relevant for EV batteries but also a warning about the range of materials China could restrict. In 2024 and 2025, additional rare-earth processing and magnet export restrictions followed, targeting the materials covered in article five of this series.
The pattern mirrors the US approach: identify the chokepoints you control, make clear you are aware of them, and begin exercising control in ways that impose costs without fully severing the relationship. Yet.
The industrial policy race #
Alongside the export control battle, every major bloc launched large-scale industrial policy programmes designed to reduce its most critical dependencies.
The United States passed the CHIPS and Science Act in August 2022: roughly $52.7 billion for domestic semiconductor manufacturing and research, plus an investment tax credit. The goal was to bring leading-edge fabrication back to American soil — specifically to attract TSMC, Samsung and Intel to build advanced fabs in the US.
The European Union announced the EU Chips Act in 2022, targeting €43 billion in investment to double Europe's share of global semiconductor production from roughly 10% to 20% by 2030. The goal was to reduce dependence on Asian fabrication for the chips used in European industry and defence.
China had been running its own programme for years. The National Integrated Circuit Industry Investment Fund — known as the "Big Fund" — had by its third phase committed hundreds of billions of yuan to domestic semiconductor development, targeting the entire chain from design to equipment to materials.
Japan launched its own programme, partly through a joint venture called Rapidus that aims to develop domestic 2nm chip fabrication — a highly ambitious target for a country without a recent history of leading-edge chip manufacturing. South Korea announced its own semiconductor cluster investments.
The global picture by 2025: every major player simultaneously claiming it will build more of the supply chain at home, using public money to subsidise it.
The boomerang problem #
Here is the thing that makes this more complicated than a simple arms race. Every export restriction hurts the restrictor as well as the target.
NVIDIA lost significant China revenue after the October 2022 rules. Its data centre segment had been selling large volumes into Chinese hyperscalers and AI companies. Those sales did not disappear cleanly — some were replaced by slower-restricted products, some by Chinese-designed alternatives, some by workarounds. But the restriction was a real cost to a US company.
ASML had been selling deep-ultraviolet (DUV) equipment to Chinese foundries, including SMIC. Extending restrictions to more DUV categories reduced revenue from a major customer. ASML's management stated publicly that the restrictions would cost the company billions in annual revenue.
Applied Materials, Lam Research and KLA — the US equipment companies — faced the same problem. China had been a major market. Restricting it was, to use a blunt term, self-harming.
This is the fundamental tension within an integrated supply chain: the lever works because the relationship is valuable. But pulling the lever degrades the relationship, which shortens the lever. The more you use it, the more the target finds alternatives, the less reach you retain.
The economic calculation that justifies the restriction anyway is that the long-term strategic cost of not restricting — a competitor gaining technological parity or advantage — outweighs the near-term revenue loss. That may be correct. It is not cost-free.
The fragmentation trajectory #
The combined effect of export controls, industrial policy, and strategic repositioning is to push the integrated global supply chain toward something more fragmented.
The term used in policy circles is "friend-shoring" — building supply chain relationships preferentially with geopolitical allies rather than the most cost-efficient suppliers. Instead of sourcing from wherever is cheapest and best, you source from wherever is safe.
Friend-shoring has a real cost. The reason the supply chain globalised in the first place was that different locations had genuine comparative advantages. TSMC fabricates better than anyone else for reasons that have accumulated over decades — process knowledge, engineering depth, scale, infrastructure. Moving fabrication to Arizona does not instantly replicate that. It replicates the machines. The knowledge accumulates more slowly.
The McKinsey Global Institute estimated that full decoupling of the US-China tech supply chain would cost both economies hundreds of billions of dollars annually. The actual trajectory is selective fragmentation, targeting the most strategically sensitive layers while preserving commercial relationships in others. But selective fragmentation still carries costs, and those costs compound over time.
Integrated global supply chain
produces: maximum efficiency, minimum cost, maximum interdependence
Selectively fragmented supply chain
produces: reduced efficiency, higher cost, reduced interdependence in critical layers
Fully fragmented supply chain
produces: minimum efficiency, maximum cost, strategic independence — maybe
Nobody is heading for full fragmentation. But everyone is moving away from full integration. The question is how far the middle ground holds.
The change #
It is worth being precise about what the last several years have actually accomplished — and what they have not.
China still cannot manufacture leading-edge chips. The EUV wall has held. SMIC made headlines in 2023 by producing a 7nm chip for Huawei's Mate 60 Pro using DUV equipment — an impressive engineering achievement, but not a route to volume production at competitive yields. The gap between the Chinese semiconductor ecosystem and the leading edge has not closed.
The US still depends on Taiwan for most of its most advanced chip fabrication. TSMC Arizona is producing at N4 nodes as of 2025, but at a fraction of TSMC Taiwan's capacity, and at higher cost. The dependency has been reduced at the margin; it has not been resolved.
China's gallium and germanium restrictions have not caused supply chain crises in the West — alternative sourcing has been developed, at higher cost. But China has demonstrated willingness to act, and its leverage over rare-earth magnets and processing is considerably larger than its leverage over gallium.
The export controls have slowed China's AI development. They have not stopped it. Chinese companies — Huawei, Cambricon, Biren, Moore Threads — are building AI accelerators using what they can access. The chips are less capable than NVIDIA's current generation. They are getting better.
The industrial policy programmes are building semiconductor infrastructure in new locations. TSMC Arizona, Samsung Texas, Intel Ohio. That infrastructure will take years to reach full capacity and more years to approach TSMC Taiwan's efficiency. It is real progress.
The drive #
If restriction is costly and fragmentation is inefficient, why is everyone doing it?
Because the alternative — allowing an adversary unrestricted access to the most powerful AI infrastructure — is considered worse. And because first-mover advantage in general-purpose AI, if it materialises, could be large enough to justify significant economic costs to prevent it accruing to the wrong party.
This is the logic of the arms race. You do not race because you want to bear the cost. You race because you fear the consequence of losing more than you fear the cost of racing.
Every country in this story is making the same calculation. The US is spending tens of billions to maintain its lead. China is spending hundreds of billions to close the gap. Europe is spending tens of billions to avoid irrelevance. Japan and South Korea are spending to protect positions they currently hold.
And underneath all of this spending, the same integrated supply chain keeps running — because neither side has yet found a way to fully replace what the other provides, and because even adversaries, it turns out, still need each other's components.