Reality Is Malleable

Essays on systems, society, and constructed futures.

The Chip Door China May Not Walk Through

Why reopening access to Nvidia chips may not restore the world export controls already broke

Chip connected to factories

Preamble

This essay is not about whether Nvidia’s H200 chips are useful to China. They are. It is not about whether U.S. export controls can slow China’s access to frontier compute. They can. The question is what happens after a chokepoint stops looking like ordinary commerce and starts looking like a dependency trap.

The H200 episode suggests that the chip war has moved into stranger territory. Washington can approve access, but it cannot easily restore trust in access. Beijing can still want the chips while also treating them as temporary, conditional, and unsafe to build around. That shift matters because once dependence becomes suspect, the response does not stay inside the chip layer. It spreads into domestic accelerators, offshore training, industrial buildout, rare earths, tungsten, transformers, refineries, and the long queues beneath modern technology.

TL;DR

The Door Opens, But the Deal Stalls

The strange thing about the H200 story is not that the door stayed closed. It is that the door appears to have opened, and the room still feels unsafe.

After the Trump-Xi summit, there was no clean breakthrough on AI chips. No dramatic reset. No sudden return to normal commerce. The U.S. had already cleared limited H200 access before the summit, with Reuters reporting that around ten Chinese firms had been approved to buy Nvidia’s H200 chips, including Alibaba, Tencent, ByteDance, and JD.com. Lenovo and Foxconn were also approved as distributors. But approval did not immediately become movement. Reuters reported that no deliveries had occurred, while Chinese hesitation appeared tied to Beijing’s desire to keep investment focused on domestic chip development.

That is the opening puzzle.

If Nvidia’s H200s are useful, and they are, why would Chinese firms not immediately rush toward them once Washington cleared a path? Why would access to a powerful American AI chip produce hesitation instead of relief?

The answer is not that the H200 is useless. The answer is that the meaning of access has changed.

On paper, the approval looks like easing. In a simpler world, that would be the story: Washington restricted, Washington relented, Chinese buyers returned, Nvidia recovered part of its China business, and the chip war softened into a commercial compromise. But that is not the world these firms are operating inside. The H200 approval arrives after years of export controls, downgraded China-specific chips, shifting rules, license uncertainty, and Chinese pressure to build domestic alternatives. It arrives after dependence itself has been politicized.

So the most important detail is not whether H200s eventually move in large volumes, token quantities, or not at all. The important detail is the pause. Even if orders appear later, the hesitation is diagnostic. It shows that access is no longer being evaluated as ordinary commerce. The transaction is now politically mediated. The chip is not only a chip. It is a test of dependency.

That distinction matters because the old chip-ban debate was too simple. It asked whether Washington would open or close the chip door. The H200 episode suggests a stranger possibility: the door can open after the room has become strategically untrustworthy.

This gives Washington a cleaner diplomatic defense than it had before. The U.S. can plausibly say: we approved the sales, we brought Nvidia into the conversation, we opened the door. If China does not buy, that is Beijing’s choice. That argument may be useful in a press conference. It may be useful in a negotiation. It may even be true in the narrow procedural sense.

But strategically, it is incomplete.

The deeper question is not whether the transaction was allowed. It is whether years of restriction changed what the transaction means. Access that arrives after dependence has been politicized does not feel like normal commerce. It feels like exposure with paperwork.

That is the first clue that the chip war has moved beyond chips. Washington may have cleared a transaction. It has not cleared the dependency risk. Reopening a door is easier than restoring trust in the hallway behind it.

The Chokepoint Loses Its Innocence

Advanced AI chips are no longer ordinary products. They have become infrastructure, leverage, bargaining objects, military-relevant inputs, cloud-enabling hardware, industrial-policy instruments, and symbols of technological dependency. That does not mean every chip purchase is a military act or every data center is a battlefield. But it does mean the category has changed. Once a chip sits inside the machinery of national power, buying it is no longer just procurement.

It is exposure.

That is why the H200 episode matters beyond the chip itself. The U.S. is no longer only deciding whether China can buy a commercial good. It is deciding the terms under which a strategic input can cross into a rival industrial system. The distinction may sound procedural from Washington. From Beijing, it is existentially different. A chip that can be licensed, capped, tested, delayed, revised, or withdrawn is not the same object as a chip that can simply be bought.

This is the problem with conditional access. It may look like a compromise, and in some ways it is. The U.S. is not saying an absolute no. It is saying yes, under conditions.

But conditional access is not normal access.

Reuters reported earlier this year that H200 exports to China were approved under rules involving third-party testing, security requirements for Chinese buyers, limits on Chinese entities’ share of total chips sold to U.S. customers, and Nvidia certification that enough H200s remain available in the United States. Chinese customers would also need to show sufficient security procedures and avoid military use. Those conditions may be defensible from the U.S. point of view. They may even be necessary if Washington wants to justify any easing at all. But they also change the meaning of the transaction.

For a Chinese firm, the chip does not arrive alone. It arrives with licensing rules, end-use review, customer limits, approval delays, compliance costs, diplomatic uncertainty, possible future rule changes, and reputational exposure inside China. The purchase may still make technical sense. It may still be commercially attractive. But the chip is now wrapped in policy.

That is not a return to normal. That is managed dependency.

This did not begin with the Trump-Xi summit. The summit merely made the condition visible. The path to this moment ran through advanced-chip restrictions, China-specific downgraded chips, the H20 compromise, changing export-control rules, licensing uncertainty, Beijing’s domestic chip-security push, Chinese firms adapting software around domestic hardware, and Huawei’s growing role as the most politically useful alternative. The H200 moment is not when chips became political. It is when the politics became impossible to ignore.

That is why the word “access” can mislead. Access sounds clean. Either the door is open or closed. Either the good is available or unavailable. But strategic access is messier than that. It can be partial, conditional, revocable, delayed, monitored, or poisoned by distrust. A firm may be allowed to buy a chip and still hesitate because the broader system around the purchase has become unstable.

From Beijing’s perspective, then, the question is no longer only: are these chips powerful?

They are.

The question is: what dependency arrives with them?

A chip with strings attached is not just a chip. It is a policy instrument disguised as hardware.

Preference Is Not Dependence

China may still prefer Nvidia chips.

That point should not be hand-waved away. Nvidia’s hardware, software ecosystem, developer familiarity, tooling, documentation, and performance advantages remain significant. A Chinese cloud provider or AI lab may have perfectly rational reasons to want H200s if they can get them. Preference is real.

But preference is not dependence.

That distinction matters because a country that desperately needs an input negotiates differently from a country that merely benefits from it. If Nvidia access is treated as irreplaceable, then every U.S. export-control decision becomes a pressure point. Every license becomes a favor. Every rule change becomes a threat. Every delayed shipment becomes leverage. In that world, China has to negotiate for relief.

But if Beijing believes it has a credible path around the chokepoint, even an imperfect one, the psychology changes. Domestic chips do not need to match Nvidia at the frontier tomorrow. Offshore training does not need to solve every workload. Model-efficiency work does not need to erase the compute gap. State-backed industrial mobilization does not need to be clean, elegant, or immediately profitable. The substitute path only needs to become plausible enough that Nvidia access stops looking existential.

A country that sees no alternative negotiates for relief. A country that sees an alternative negotiates for advantage.

This is where “good enough” becomes strategically dangerous. Chinese domestic chips do not need to beat Nvidia in every benchmark to change the bargaining position. They need to be good enough for enough tasks that dependence weakens. Inference workloads, government systems, telecom AI, industrial automation, public-sector procurement, surveillance systems, robotics, edge deployments, defense-support functions, domestic cloud workloads, and models adapted around Chinese hardware all create room for substitution below the frontier.

That is not the same as technological parity. It is something narrower and perhaps more important: operational sufficiency.

Once domestic systems can handle a growing share of ordinary workloads, Nvidia becomes less like oxygen and more like premium equipment. Still desirable. Still useful. Still worth acquiring under the right conditions. But no longer the only path to breathing.

This changes the entire negotiation. Washington may still control the best chips. It may still slow China’s frontier progress. It may still force Chinese firms to accept inefficiency, higher costs, and awkward workarounds. None of that disappears. But the nature of U.S. leverage shifts when the denied input is no longer treated as fully irreplaceable.

That is why the H200 hesitation matters. The pause does not prove China has solved domestic AI hardware. It does not prove Huawei has caught Nvidia. It does not prove Chinese firms can ignore foreign compute. The modest claim is stronger because it is more realistic.

China does not need parity for leverage to decay. It needs a credible exit path.

The Double Bind Is Not Symmetrical

It would be too easy to say that every U.S. option is equally bad.

That would make the argument cleaner, but less true. Export controls may still buy time. They may still delay Chinese access to the best hardware. They may still force Chinese firms onto less efficient chips, more awkward architectures, and more expensive workarounds. A year or two of delay at the frontier is not meaningless when the technology in question may shape military planning, scientific research, industrial automation, cyber operations, and state capacity.

The problem is not that the choices are identical. The problem is that each choice creates a different failure mode across a different time horizon.

If the U.S. keeps restrictions tight, it preserves pressure on China’s access to frontier compute. That pressure may matter. It may slow some training runs. It may raise costs. It may give American firms and allied governments more time to prepare. But pressure also creates incentives. The tighter the restriction, the stronger the reason for China to build around it. Chinese firms have more reason to adapt to Huawei and other domestic hardware. Domestic procurement becomes politically favored. Model efficiency becomes a strategic priority. Cloud firms redesign around available chips. Infrastructure gets built for the hardware China can control rather than the hardware it wishes it could buy.

Restriction may buy time.

But it also teaches substitution.

If the U.S. partially eases access, the failure mode changes. Chinese firms may gain useful hardware. H200s may support mid-tier training, fine-tuning, benchmarking, and research workflows. Nvidia may preserve some relevance in China. U.S. firms may keep a foothold in a large market rather than surrendering it entirely to domestic alternatives. In the short term, that may reduce the pressure for abrupt decoupling.

But conditional easing carries its own poison. If access comes through licenses, caps, reviews, security certifications, shifting diplomatic moods, and future uncertainty, then the chip arrives as proof of the dependency problem it is supposed to soften. The transaction says: yes, you may have this, but only through a gate we control. That may be preferable to a locked gate. It is still a gate.

Conditional easing may reduce immediate pressure.

But it also proves the dependency is politically mediated.

The third option would be unconditional easing, or something close to it: a broader return to commercial normalcy where Chinese firms can buy advanced American AI chips without the thick layer of strategic supervision. That path may be the least compatible with U.S. politics. If Washington fully reopened access, critics would argue that the U.S. was feeding China’s AI ecosystem, strengthening a rival military-industrial base, and sacrificing national security for Nvidia’s market share. Whether one agrees with that argument or not, it is politically potent enough to make full normalization difficult.

So the U.S. is not facing a neat choice between control and surrender. It is trapped between imperfect instruments.

Restriction strengthens the case for Chinese substitution. Conditional access strengthens the case for Chinese suspicion. Openness strengthens Chinese competence and may be politically unacceptable in Washington.

That is the double bind, but it is not symmetrical. The costs arrive in different forms. Restriction creates a slow-burn substitution problem. Conditional access creates a trust problem. Openness creates an immediate capability problem. Policymakers can choose which danger to emphasize. They cannot make the others disappear.

The summit did not resolve this structure. It clarified it. Washington can now say it opened the door. Beijing can still look at the room and decide it is unsafe to build inside.

China’s Multi-Vector Adaptation

The weakest way to understand China’s response is to reduce it to a single contest: Nvidia or Huawei.

That frame is too small. It imagines China standing at a fork in the road, choosing either American chips or domestic chips, either dependence or self-reliance, either the best hardware or national pride. Real industrial systems do not adapt that cleanly. They segment. They route around constraints. They use what is available, protect what is sensitive, and keep building toward the future they would rather have.

China does not need one answer to the chip ban. It can pursue several at once.

For mass inference, domestic chips become more attractive because the workload is large, repetitive, and sovereignty-sensitive. If a system is running inside Chinese cloud services, government platforms, telecom networks, industrial automation, surveillance infrastructure, logistics systems, or public-sector procurement, supply continuity may matter more than absolute frontier performance. A slower domestic chip that remains available can be more strategically useful than a faster foreign chip that might disappear under the next rule change.

For government and other sovereignty-sensitive workloads, the logic is even clearer. If the data, deployment context, or political function is sensitive, foreign hardware may be treated as a risk even when it performs better. The question becomes less “which chip wins the benchmark?” and more “which chip can be trusted to remain inside the system?” That gives domestic hardware a political advantage long before it reaches technical parity.

For mid-tier training and R&D, H200s could still matter. If they enter China in meaningful numbers, they may support fine-tuning, experimentation, benchmarking, hybrid infrastructure, and workloads where domestic alternatives remain awkward or supply-constrained. Chinese firms may want them for practical reasons. They could help teams iterate faster, compare domestic accelerators against Nvidia hardware, and bridge the gap while local supply chains mature.

But buying some H200s would not mean returning to dependence. It may simply mean using the foreign chip as a tool inside a broader exit strategy.

For frontier training, the geography changes. If the best chips cannot safely be imported, Chinese firms can try to move some workloads to where the chips are. Offshore data centers in Southeast Asia and other jurisdictions become part of the adaptation map. The issue is no longer only whether China can import the chip. It is whether Chinese firms can move the workload to the chip.

That does not make offshore training easy, unlimited, or risk-free. It can raise legal, regulatory, data-security, cost, and policy risks. U.S. officials can notice loopholes. Host countries can face pressure. Data cannot always move freely. But the strategy matters because it shows that the chip ban does not merely create a domestic substitution problem. It creates a compute-geography problem.

The result is a layered response rather than a single pivot. Domestic chips can handle more inference and sovereignty-sensitive workloads. Imported H200s, if allowed, can support mid-tier training, benchmarking, and temporary bridging. Offshore Nvidia clusters can support frontier runs where possible. Domestic chipmakers continue to improve. Chinese software teams adapt models, compilers, and workflows around the hardware they can reliably access.

This is not clean decoupling. It is not simple dependence either.

It is segmentation under pressure.

That is why the H200 question matters less than it first appears. China may buy some chips and still reduce dependence. It may reject some chips and still use offshore compute. It may rely on domestic chips for ordinary workloads while using foreign hardware for specific bottlenecks. It may tolerate inefficiency in one layer to gain sovereignty in another.

China domesticates what must be sovereign, offshores what must be frontier, imports what is useful, and treats every dependency as temporary.

The Industrial Mobilization Beneath the Chip War

The chip ban does not merely deny access. It creates a national mission.

That mission is often misunderstood. It is not necessarily to beat Nvidia immediately. That would be a convenient straw man, because it lets critics dismiss Chinese substitution efforts as inferior, inefficient, late, or technically immature. Some of those criticisms may be true. Chinese domestic accelerators may lag. Yields may disappoint. Software stacks may be clumsy. Hardware may be less efficient. Developers may prefer Nvidia. None of that makes the mission irrelevant.

The mission is not to surpass Nvidia tomorrow.

The mission is to make China less hostage to Nvidia over time.

That difference matters because a national substitution campaign does not need to produce a perfect replacement at the beginning. It needs to create a protected path of improvement. It needs customers who are told to adapt, capital that can tolerate failure, procurement rules that favor domestic supply, engineers who learn by building, and institutions willing to accept inefficiency in exchange for sovereignty. A worse chip can still become strategically useful if the alternative is dependence on a chip that may not arrive.

But a domestic AI chip ecosystem is not made out of patriotic speeches. It is made out of fabs, packaging lines, memory systems, interconnects, data centers, cooling equipment, power infrastructure, substations, transformers, industrial parks, machine tools, cutting tools, chemical inputs, cleanroom systems, compilers, software stacks, developer ecosystems, maintenance teams, and procurement mandates.

That is where the chip ban begins to escape the chip layer.

Once China commits to reducing dependence on foreign AI hardware, it has to build the world around domestic hardware. That means more than designing accelerators. It means building the physical and organizational environment in which those accelerators can be manufactured, deployed, powered, cooled, maintained, programmed, benchmarked, and improved. The chip is the visible object. The system around the chip is the actual mobilization.

Critics may say this buildout will be inefficient.

They may be right.

But inefficiency does not make the material demand disappear. It can make the material demand larger. A less efficient domestic stack may require more chips, more data centers, more power, more cooling, more redundant capacity, more replacement parts, more grid hardware, more machine tools, more factories, more spare inventory, and more capital sunk into parallel systems. Wasteful mobilization is still mobilization. It still consumes steel, copper, specialty metals, rare earth magnets, tungsten carbide tooling, power electronics, cooling systems, transformers, chemicals, land, water, and labor.

This is one of the stranger asymmetries in the chip war. The West may look at Chinese substitution and see waste. But if the wasted effort consumes materials the West also needs, the waste is not harmless. An inefficient Chinese buildout can still tighten global supply chains. It can still absorb industrial capacity. It can still push foreign buyers into longer queues.

A bad substitute can still be a good sponge.

This is why H200 access does not necessarily reverse the domestic buildout. Even if some H200s enter China, they do not erase the strategic lesson. H200 access remains politically mediated. It remains below the true frontier. It can be revised by future U.S. rules. It may be unacceptable for sensitive workloads. It cannot provide full strategic autonomy. It may help Chinese firms bridge certain gaps, but a bridge is not a home.

So the domestic path continues.

That path may be slower than China wants. It may be expensive. It may duplicate capacity. It may produce dead ends. It may create ugly overbuilds and bureaucratic waste. But it also creates factories, suppliers, trained workers, software adaptations, procurement habits, and a domestic customer base that learns to live without the ideal foreign input.

The chip war is not only fought at the chip layer. It is fought in the mines, refineries, chemical plants, machine shops, grids, transformers, cooling systems, waste streams, and maintenance crews beneath the chip layer.

The Queue Behind the Chip

This is where the quieter danger enters.

The obvious fear is that China might weaponize exports. The story is familiar: the U.S. restricts chips, China retaliates with rare earths, tungsten, gallium, germanium, graphite, or other strategic materials, and the West scrambles for substitutes. That version is visible. It has villains, announcements, countermeasures, and headlines.

But the more durable problem may be less theatrical.

China may not need to formally cut anyone off. It may simply consume more of its own strategic material base.

That is the difference between weaponized scarcity and absorbed scarcity. Weaponized scarcity says: you cannot have this because we refuse to sell it. Absorbed scarcity says: we need this ourselves.

The second version is harder to notice and harder to negotiate with. A ban is a policy. A queue is a condition. A policy can sometimes be reversed by a deal, a concession, a summit, or a change in leadership. Capacity cannot be negotiated into existence overnight. If the material has been consumed, if the refinery slot has been allocated, if the magnet plant is booked, if the tungsten carbide is already going into domestic tooling, there is no diplomatic phrase that makes it instantly available again.

This is why absorbed scarcity is so dangerous. It does not announce itself. It accumulates.

For a long time, it looks like boring operational noise. Lead times move from six months to nine months. Prices rise, settle, then rise again. One supplier says capacity is tight. Another says delivery is delayed. Export licenses take longer. Spot supply gets thinner. Contracts include worse terms. Buyers quietly over-order because they no longer trust the market. Financiers adjust risk models. Small firms get squeezed before large firms notice. Official trade data confirms the pattern only after procurement departments have already been living inside it.

None of this screams geopolitical crisis. It looks like procurement friction. Spreadsheet weather.

Then one day the delay is no longer treated as temporary. It is the new structure.

The easiest way to understand this is not to begin with rare earths. It is to look at the transformer bottleneck already hitting Western grid expansion. The problem is not that transformers do not exist. The problem is that data centers, grid modernization, electrification, factories, renewable projects, and replacement cycles all need the same slow-to-build equipment at the same time.

Reuters reported in May 2026 that U.S. demand for generator step-up transformers had risen 274% since 2019, while substation transformer demand had risen 116%. Prices were up roughly 80% over five years, and lead times for large units could stretch to four years. Developers were trying to secure imports and factory slots, while manufacturers were expanding capacity but still faced a demand surge they could not instantly absorb.

That is not an embargo. It is a queue. And for the project waiting on the transformer, delay can feel a lot like denial.

Absorbed scarcity works the same way. The export relationship can remain technically open while practical access deteriorates. Foreign buyers may still be able to purchase material, but later, at higher prices, under worse terms, with less predictable delivery, thinner spot-market availability, more licensing friction, and fewer refined products available when needed. The door remains open. The shelves behind it begin to empty.

This is what export starvation means. Not a sudden cutoff. Not necessarily a press conference. Not a dramatic declaration that the West may no longer buy. Export starvation occurs when the exportable surplus shrinks because domestic demand rises. China’s own industrial machine takes first call on the material. Foreign buyers remain in the system, but they move backward in the line.

The H200 episode does not negate this risk. It clarifies it.

If China hesitates to buy H200s, domestic buildout continues. If China buys some H200s, domestic buildout still continues because the access remains conditional, reversible, and politically unsafe. Either way, the path toward self-reliance remains rational. And that path does not merely require chips. It requires the material world beneath the chips: factories, grids, cooling systems, machine tools, magnets, carbide, chemicals, refiners, and workers.

By the time absorbed scarcity is visible to politicians, it may already be old news to purchasing departments.

Rare Earths and Tungsten: The Stack Beneath the Stack

This argument has to be framed carefully.

The claim is not that every AI chip directly consumes huge quantities of rare earths or tungsten. That would be too narrow, too easy to challenge, and mostly beside the point. A GPU is not a tiny dragon hoarding dysprosium and tungsten in its silicon belly.

The stronger claim is that AI independence belongs to a wider industrial mobilization. It pulls factories, grids, motors, robotics, defense systems, machine tools, cooling equipment, precision components, mining equipment, data centers, and automated manufacturing into motion. That wider system does consume rare earths, tungsten, and other strategic inputs. The chip war’s material footprint is not measured only in grams per GPU. It is measured in the industrial world that a sovereign AI stack requires.

Rare earths matter most visibly through magnets.

That sounds almost too mundane for a conversation about artificial intelligence, but the mundane layer is where the physical world keeps its veto power. Permanent magnets show up in electric motors, drones, robotics, EVs, wind turbines, industrial automation, precision equipment, aerospace systems, hard drives, actuators, factory systems, and defense applications. They are not glamorous. They do not look like the frontier. But they help make the frontier move.

The AI link is indirect, but real. A domestic AI buildout does not merely mean training models. It means expanding the industrial systems around AI: the data centers that host it, the factories that produce the hardware, the robots that deploy it, the drones and sensors that use it, the grid equipment that powers it, and the automated production lines that fold it back into manufacturing. Those systems touch the magnet economy again and again.

This is also why the rare earth question cannot be reduced to ore. A deposit is not a supply chain. A mine is not a magnet. Between the rock and the finished component sits a difficult chain of separation, refining, oxide production, metal conversion, alloying, magnet manufacturing, component integration, customer qualification, and waste management. A country can have rare earth deposits and still lack rare earth supply in the form its industries actually need.

China’s advantage is especially dangerous in these middle and downstream layers. The chokepoint is not only what comes out of the ground. It is what can be processed, purified, alloyed, manufactured, qualified, and delivered at scale.

Tungsten is less famous, which may make it more revealing.

Rare earths at least have a public mythology now. They appear in headlines. People vaguely know they matter for magnets, EVs, wind turbines, and defense. Tungsten is quieter. It lives deeper in the industrial basement, in cutting tools, tungsten carbide, machine tools, drilling equipment, mining equipment, aerospace parts, semiconductor-support equipment, high-temperature systems, wear-resistant components, and factory maintenance.

That makes tungsten a recursive bottleneck. It is not only an input into final products. It is an input into the tools that make products. A country trying to rebuild factories, mines, fabs, machine shops, weapons lines, and advanced manufacturing capacity needs the hard materials that cut, drill, shape, and maintain the machines doing the work.

You need tungsten to make the tools that make the factories that reduce your dependence on China.

That sentence is almost childish in its simplicity, which is why it matters. The industrial stack loops back on itself. If the West wants to reduce dependence on China, it needs more domestic and allied production. To build that production, it needs tools. To make and maintain those tools, it needs materials like tungsten. If those materials are also under pressure from China’s own industrial buildout, the escape route itself becomes harder to build.

This is where the argument must remain precise. China is not resource-omnipotent. It does not control every material. It imports huge quantities of raw inputs. It has energy constraints, demographic constraints, financial constraints, environmental constraints, and its own vulnerabilities across the global resource system. Treating China as a magical autarky machine would weaken the argument.

The point is narrower and stronger.

China is not resource-independent. It is choke-point dominant.

That distinction matters. A country can be vulnerable in some upstream inputs while still dominating the midstream layers that others need to turn raw material into usable industrial supply. A country can import raw resources while controlling refining, separation, processing, magnet production, carbide supply chains, and component manufacturing. It does not need to control everything to create pressure at the points where everyone else is weakest.

That is the stack beneath the stack. Beneath the AI model sits the chip. Beneath the chip sits the fab. Beneath the fab sits the toolchain. Beneath the toolchain sit minerals, chemicals, power, waste systems, skilled operators, and years of accumulated process knowledge.

The chip war only looks clean from far away.

The West Is Trying to Buy More From a Shrinking Surplus

The problem is not only that China may consume more of its own strategic materials.

The problem is that the West needs more too.

This is the demand-side collision. Western governments are not merely trying to preserve yesterday’s supply chains. They are trying to expand several industrial systems at once: AI infrastructure, semiconductor fabs, grid upgrades, EV production, batteries, wind turbines, defense production, drones, robotics, data centers, reshored manufacturing, mining equipment, power electronics, machine tools, and strategic stockpiles.

Each of these goals can make sense on its own. AI leadership matters. Chip sovereignty matters. Grid modernization matters. Defense readiness matters. Clean energy matters. Supply-chain resilience matters. Industrial reshoring matters. NATO rearmament matters. None of these ambitions is absurd in isolation.

But the physical world does not care that the policies were written in separate documents.

It sees the same demand showing up at the same door.

The AI strategy becomes a demand for power, cooling, transformers, chips, land, water, and grid equipment. The clean-energy strategy becomes a demand for magnets, batteries, copper, power electronics, transmission infrastructure, and permitting. The defense strategy becomes a demand for explosives, tungsten, rare earth magnets, aerospace components, drones, shipyard capacity, and secure electronics. The semiconductor strategy becomes a demand for tools, gases, chemicals, ultrapure water, skilled labor, machine parts, and power.

On paper, these are different policy agendas. In the material world, they overlap.

They draw from the same pools of metals, minerals, chemicals, refineries, skilled workers, equipment, power, water, logistics, permits, and capital. They compete for the same engineers, electricians, machinists, metallurgists, construction crews, port capacity, transmission queues, and financing. The spreadsheet separates them. The supply chain does not.

That is what makes the absorbed-scarcity risk sharper. The West is not trying to buy the same amount from a shrinking Chinese surplus. It is trying to buy more from a shrinking Chinese surplus.

This is also why “the West” is a useful shorthand but a misleading actor. There is no single Western buyer standing calmly in one line. There is the United States trying to rebuild defense production, build fabs, expand data centers, modernize the grid, and secure critical minerals. There is Europe trying to manage clean-energy goals, high energy costs, defense rearmament, and industrial competitiveness. There are Japan and South Korea trying to protect advanced manufacturing bases with limited domestic resource supply. There is Taiwan, industrially sophisticated but dependent on uninterrupted flows of energy, chemicals, equipment, and logistics. There are Canada and Australia with resource potential but long permitting timelines and the risk of remaining upstream suppliers without capturing enough of the midstream.

Then there are the firms: allied defense contractors, automakers, battery makers, semiconductor companies, clean-energy developers, data-center builders, robotics firms, mining companies, utilities, machine shops, and national stockpile managers. Politically, many of them may be on the same side. Materially, they may be bidding against one another.

Friend-shoring sounds cooperative until the shortage begins.

At that point, allies can become rival bidders for offtake agreements, refinery slots, magnet supply, tungsten supply, skilled workers, processing capacity, equipment, capital, and stockpile access. One country’s secured contract can become another country’s missing input. One ally’s resilience can become another ally’s delay.

This does not mean cooperation is impossible. It means cooperation has to be built before scarcity forces improvisation. If allies wait until the queue forms, they may discover that shared values do not automatically allocate scarce capacity. Someone still gets the delivery slot. Someone still waits.

The West is not one customer waiting in one line. It is several nervous customers trying to reserve the same limited future capacity.

The Dirty Basement of Strategic Autonomy

The easiest answer is also the laziest one: the West can just build its own supply chains.

At one level, this is true. The West can mine, process, refine, recycle, stockpile, subsidize, and friend-shore. It can build new rare earth chains. It can develop non-Chinese tungsten supply. It can invest in magnet plants, carbide production, machine tools, and strategic reserves. There is no law of physics that says China must dominate these layers forever.

But supply chains are not summoned by press release.

A deposit is not supply. This is the ore fallacy. A country can have rare earth deposits, tungsten resources, lithium reserves, copper projects, or promising geological surveys and still lack usable industrial supply. Supply means material that is extracted, processed, separated, refined, converted, alloyed, manufactured, qualified, contracted, delivered on time, and available at scale. Every verb in that sentence is a chokepoint.

There is also the project fallacy: the belief that announcing a mine, refinery, magnet plant, recycling facility, or tungsten project means the bottleneck has been solved. But announcements are not capacity. Pilot plants are not capacity. Press releases are not capacity. A groundbreaking ceremony is not a supply chain. The relevant question is not whether a project exists in a ministerial speech or investor deck. The question is whether it scales enough to matter before the shortage bites.

This is where strategic autonomy meets its dirty basement.

Western publics are already uneasy about data centers. They object to power demand, water use, noise, land footprint, tax incentives, grid stress, limited permanent employment, and strain on local infrastructure. These objections are not imaginary. Data centers do impose local costs, and communities are allowed to notice when the glowing abstraction of “AI” turns into a large industrial box at the edge of town.

But data centers are the mild controversy.

Compared with rare earth separation, tungsten mining, chemical processing, tailings, waste rock, solvent extraction, wastewater treatment, radioactive residues depending on the ore body, dust control, worker-safety risks, and long-term remediation obligations, a data center is almost polite. It may be loud, thirsty, power-hungry, and unpopular. It is still cleaner than the upstream world required to make advanced technology physically possible.

This is the contradiction. The West likes the clean endpoint: the AI model, the EV, the wind turbine, the precision drone, the modernized grid, the secure semiconductor supply chain, the polished device in the box. It is much less eager to host the messy beginning. The acids, tailings, solvents, waste streams, permits, trucks, settling ponds, monitoring systems, and local hearings do not fit the aesthetic of clean technology. But the clean endpoint depends on the dirty beginning.

Every strategic project exists somewhere.

A national government may say it needs domestic critical minerals. A town may say: not next to our river. A province may say: not near our farmland. An Indigenous government may say: not without consent. Environmental groups may say: not without full review. Residents may say: not if it changes our roads, water, property values, or local risk. Many of these objections can be legitimate. A community is not irrational because it does not want to become the sacrifice zone for someone else’s resilience strategy.

But the cumulative effect can be brutal. Everyone wants strategic autonomy in theory. The argument becomes harder when autonomy needs a postal code.

Even if the politics are solved, the knowledge problem remains. The West did not merely outsource production. It outsourced memory.

Industrial capacity is not just equipment sitting under a roof. A refinery is not a building. It is a practiced civilization of operators, suppliers, chemists, maintenance crews, regulators, customers, managers, safety systems, troubleshooting habits, and accumulated mistakes. A separation plant has to learn what normal looks like. A magnet producer has to qualify output for demanding customers. A tungsten carbide supplier has to meet exacting industrial needs. A maintenance crew has to know which vibration matters before the machine fails.

This is tacit knowledge. It does not return just because a government allocates money. It has to be rebuilt in workers, suppliers, institutions, machines, customers, and time.

Then there is the price problem. Scarcity does not have to deny access. It can make access too unstable to build a business on. A new Western processor or magnet plant may need predictable input costs, long-term offtake agreements, insurance, stable customers, and confidence that prices will not swing violently before it reaches scale. If material prices spike, collapse, or fluctuate unpredictably, investors hesitate. Contracts become harder to price. Subsidies need to grow. Small manufacturers get squeezed. Final investment decisions slip. The project may die before the first serious bottleneck is solved.

This is why “just pivot” is too simple. The West can build more resilient supply chains, but only if it treats the problem as physical, political, financial, environmental, and institutional at the same time. Mines are not enough. Refineries are not enough. Announcements are not enough. Recycling is useful but not magical. Friend-shoring helps but does not remove the need for processing, qualification, waste management, and local consent.

Strategic autonomy is not an app. It is a mine, a refinery, a waste stream, a grid connection, and a town-hall meeting.

The Room Was Not Just About Chips

The summit guest list is a small detail, but a revealing one.

The companies visible on the Chinese side were not simply the obvious frontier-AI players gathering around Nvidia. Reports and circulated attendee lists pointed instead to a wider industrial mix: ByteDance, Lenovo, Xiaomi, Hisense, Air China, CICC, ICBC, Haier, Fuyao Glass, Lens Technology, COFCO, Bank of China, COMAC, Wanxiang Group, and COSCO. The list included technology firms, but also aviation, banking, shipping, food, consumer electronics, manufacturing, auto components, and industrial logistics. Yicai Global identified several of those Chinese executives from CCTV footage and on-site identification, including Lenovo’s Yang Yuanqing, Xiaomi’s Lei Jun, ByteDance’s Liang Rubo, Hisense’s Jia Shaoqian, Lens Technology’s Zhou Qunfei, Wanxiang’s Lu Weiding, and Fuyao Glass’s Cao Hui.

That matters because the room did not look like China’s AI frontier begging for Nvidia.

It looked more like industrial China deciding where American technology still fits.

The absences may matter too. Huawei, Alibaba, and Tencent were not the visible center of the Chinese-side lineup. That should not be overread; summits have choreography, guest lists can be partial, and absence from one event does not mean absence from the broader policy conversation. But the optics still fit the larger pattern. Huawei’s absence kept the domestic-substitution champion offstage. Alibaba and Tencent’s absence kept the H200 question from looking like direct supplication by China’s cloud giants. ByteDance and Lenovo provided enough technology presence without making the room appear to be a narrow negotiation over frontier AI compute.

This is a subtle point, not a load-bearing proof. The essay does not need to argue that the guest list secretly explains policy. It only needs to notice what the composition suggests. The chip issue was embedded inside a broader industrial conversation: aircraft, shipping, finance, food security, consumer electronics, manufacturing, logistics, and applied technology. That is the real context in which China evaluates Nvidia access.

This supports the larger argument. China is not approaching the summit as a single buyer of American compute. It is approaching as an industrial system deciding where foreign compute still fits. The chip is one object in the room. It is not the room.

Counterarguments: The Ban Still Buys Time

The strongest argument for export controls should be stated plainly: the ban may still buy time.

That is not a trivial defense. If advanced AI systems matter for military planning, cyber operations, surveillance, scientific discovery, industrial automation, and state capacity, then delaying a rival’s access to frontier compute may have real value. Export controls can force China onto inferior hardware. They can increase costs. They can complicate frontier training runs. They can slow some military-relevant applications. They can preserve some Western advantage while American and allied firms continue to develop better chips, models, infrastructure, and deployment systems.

That argument should not be dismissed as stupid or merely protectionist. Time matters.

The issue is what the West does with the time.

Buying time only matters if the time is used to build resilience. If export controls delay China but the West does not build mines, refineries, magnet plants, tungsten processing, stockpiles, machine-tool capacity, allied allocation rules, skilled workforces, permitting capacity, and full-chain redundancy, then the time-buy becomes a time-bomb. The clock keeps running, but the physical escape route does not get built.

This is the difference between a delay and a strategy. A delay says: the other side is slowed. A strategy says: while they are slowed, we are becoming harder to coerce.

The second counterargument is that China’s alternatives may remain inferior. This is also serious. Huawei is not Nvidia. Domestic chips may lag in performance, yields, energy efficiency, memory bandwidth, software support, and developer familiarity. Advanced packaging may remain difficult. Memory bottlenecks may persist. Offshore training may face greater U.S. scrutiny. Domestic software stacks may be immature. Some workloads may remain painfully dependent on Nvidia-class hardware for longer than Beijing would like.

All of that can be true.

But China does not need immediate parity for U.S. leverage to decay. It needs enough viable alternatives that Nvidia becomes preference rather than dependence. If domestic chips become good enough for inference, public-sector workloads, telecom systems, industrial AI, and a widening range of cloud tasks, then the U.S. still holds an advantage at the frontier, but less control over the base. The bargaining position changes before the benchmark crown changes hands.

Another counterargument is that China may avoid squeezing the West too hard. That is plausible. Beijing may not want to trigger a panic. A dramatic export cutoff could accelerate Western diversification, scare buyers away, justify emergency subsidies, unify U.S. allies, trigger stockpiling, and invite retaliation. China may prefer controlled dependence over a clean rupture. It may want foreign buyers to remain exposed, nervous, and partially reliant rather than fully mobilized against it.

But that possibility actually supports the absorbed-scarcity frame. The squeeze does not need to be dramatic. It may stay ambiguous. It may appear as longer lead times, tighter licensing, higher prices, domestic priority, and less reliable access rather than an obvious embargo. The absence of a dramatic cutoff does not prove the absence of pressure.

Western diversification is also real. The United States, Canada, Australia, Europe, Japan, South Korea, and allied firms are all trying to reduce exposure. There are mining projects, processing initiatives, recycling efforts, stockpile discussions, defense grants, and friend-shoring proposals. The West is not asleep.

But the question is not whether diversification exists. The question is whether it moves fast enough, at sufficient scale, across enough of the chain, with enough political durability, to matter before the next squeeze. Mining without refining is not enough. Refining without qualification is not enough. Pilot projects without scale are not enough. Friendly ore without allied processing is not enough. A resilient supply chain has to survive price volatility, permitting delays, local opposition, financing risk, environmental scrutiny, and election cycles.

The material-absorption argument should therefore be understood as a structural risk, not a prophecy. It is not a claim that China will consume all strategic materials, that Western diversification will fail, or that export controls are automatically self-defeating. It is a warning that the chip ban may produce second- and third-order consequences that are not visible if the analysis stops at the chip layer.

The ban may buy time. The danger is that China spends that time adapting while the West spends it congratulating itself for having bought time.

Conclusion: A Room Washington Can Lock

The chip ban did not merely restrict Chinese access to compute. It changed the meaning of access.

That may be the deeper effect. Export controls were designed to preserve a technological chokepoint, and in the narrow sense they may still do that. They can slow purchases, limit access to frontier chips, raise costs, complicate training runs, and force Chinese firms to adapt around constraints. None of that is imaginary. The chip layer still matters.

But the chip layer is no longer the whole story.

Once access becomes conditional, reversible, negotiated, delayed, and politically exposed, the buyer begins to think differently. The question is no longer only whether the chip is powerful. It is whether the system around the chip can be trusted. A country may still want the hardware. A company may still prefer the performance. Engineers may still know the toolchain. But if the path to that hardware runs through another state’s permission structure, then dependence itself becomes part of the cost.

That is the trap now visible through the H200 episode.

Restriction teaches substitution. Conditional easing teaches suspicion. Offshore training teaches geographic adaptation. Domestic buildout teaches material absorption. Western resilience efforts collide with China’s own industrial appetite. What began as a fight over access to chips begins to spread downward into power systems, machine tools, rare earth magnets, tungsten carbide, refineries, transformers, factories, permits, and queues.

The U.S. tried to control the chip layer. China responded by changing the meaning of dependence.

H200 approvals may open a door, but they do not restore the old assumption that American AI hardware is safe to build around. Once access becomes conditional, every path leads toward adaptation: domestic chips where sovereignty matters, offshore compute where frontier performance matters, selective imports where useful, and industrial mobilization underneath it all.

This does not mean China wins. It does not mean export controls are pointless. It does not mean the West is doomed to industrial helplessness. The ban may still buy time. The West may still use that time well. Diversification is possible. New supply chains can be built. Strategic stockpiles can be expanded. Refineries, magnet plants, tungsten processing, machine-tool capacity, and allied allocation systems can be rebuilt.

But none of that happens automatically. And none of it is clean.

The danger is that Washington treats the chip ban as a switch while Beijing treats it as a lesson. The switch can be flipped, adjusted, loosened, tightened, or defended at a press conference. The lesson is harder to undo. Once a strategic input has been politicized, the old innocence does not return just because a license is approved.

That is why the next phase of the chip war may look less dramatic than people expect. It may not begin with a slammed door. It may look like partial access, cautious buyers, domestic procurement, offshore workarounds, longer lead times, rising prices, tighter supply, and governments quietly discovering that the real constraint has moved into the layers below the visible technology.

The chip ban began as a fight over access. It may end as something stranger: a world where the door is open, the buyer is wary, and the real shortage has moved into the mines, refineries, tools, grids, transformers, and queues beneath the chip.

The H200 door may open. The deeper question is whether Beijing still wants to build its future in a room Washington can lock.

- Iarmhar

May 21, 2026

Update: China Begins Policing the Escape Hatches

Shortly after this essay was drafted, the Financial Times reported that China had added Nvidia’s RTX 5090D V2, a China-specific export-compliant gaming GPU, to a customs banned-goods list during Jensen Huang’s visit. The 5090D V2 is not an H200-class data-center chip, and it should not be treated as one. But the signal matters. If the report holds, Beijing is not merely adapting to American restrictions. It is also beginning to police its own fallback channels when those channels risk keeping Chinese developers dependent on Nvidia hardware.

That does not change the essay’s argument. It sharpens it. The chip war is no longer only about whether Washington opens or closes the door. It is also about whether Beijing wants every open door to remain usable. Foreign compute may still be useful, but usefulness alone is no longer enough to make it strategically welcome.

Follow on X to be notified when new essays are posted.

Related

The Throughput Economy
An analysis of how China’s manufacturing and platform ecosystems are compressing Western retail margins.

The Cost Singularity: How AI Breaks the Business Model of Every Major Company
A theory of how AI-driven near-zero marginal costs could break legacy business models.

When AI Meets Leverage
A warning that AI-driven cost compression could destabilize companies built around fixed debt assumptions.