AI, China, and the struggle over the infrastructure of technological power

I am not writing this as a specialist in International Relations, economics, or China. My background is in Computer Engineering, and my interest comes precisely from that position: someone who learned to look at technology through the software layer, but started to suspect that some of the most important disputes of this decade are taking place far below it.

For a long time, technology was narrated as a software story. Applications, platforms, social networks, cloud, data, and, more recently, artificial intelligence seemed to be at the center of everything. For people who work or study computing, it is easy to get used to this abstract layer: code, APIs, containers, models, services, dashboards, deployments, and that YAML nobody knows exactly why it works but nobody dares to touch.

Hardware was always there, of course. But it often appeared as backstage infrastructure: important, yet invisible. The world seemed to be heading toward a simple conclusion: whoever controlled software would control the future.

The 2020s began to dismantle that perception.

The global semiconductor shortage, the race for GPUs, the rise of generative AI, Taiwan’s centrality, the almost invisible role of companies such as ASML and TSMC, and the technology sanctions against China revealed something that may have been hidden behind the word “cloud”: software does not run in a vacuum.

The cloud has land, a legal entity, an electricity bill, and a cooling system.

Artificial intelligence has factories. And computing, before it looks magical inside a beautiful interface, depends on energy, silicon, machines, data centers, water, logistics, currency, industrial policy, and geopolitical power.

This text comes from that unease.

It is not an attempt to explain all of China’s geopolitics or predict the future of artificial intelligence. It is an attempt to organize a perception: cutting-edge technology is not only software. It is a struggle over the physical, industrial, and political infrastructure that makes software possible.

1. Software seemed sufficient

For years, the technological imagination was dominated by the idea that software was the main force transforming the world.

After all, the most influential companies seemed to be software companies. The biggest changes in everyday life came from applications, social networks, recommendation systems, streaming platforms, marketplaces, digital banks, cloud services, SaaS, automation, and, later, artificial intelligence.

The phrase “software is eating the world” became almost a summary of an era. Everything seemed to be moving toward abstraction: fewer atoms, more bits; fewer factories, more platforms; less heavy industry, more digital products.

For a while, it seemed that the world would be conquered by whoever had the best app, algorithm, platform, or blue-button onboarding screen.

Even when hardware mattered, it stayed in the background. Few people outside specialized fields wanted to know how a chip was manufactured, who made lithography machines, where factories were located, which countries controlled particular materials, or which companies dominated critical parts of the supply chain.

For the end user, it was enough for the phone to be faster, the laptop to have better battery life, the game to run more smoothly, and the cloud to be available. The physical infrastructure behind all of it almost disappeared from public perception.

In a sense, this invisibility was a sign of success. The better infrastructure worked, the less we had to think about it.

It was as if the physical world had become merely an implementation detail.

But that abstraction had a limit.

And that limit began to appear forcefully after 2020.

2. The decade when hardware returned to the center

The pandemic did not create the world’s dependence on semiconductors, but it exposed its scale.

Suddenly, chips stopped being an electronics-engineering topic and became economic news. Automakers delayed production. Electronics became more expensive. Consoles, graphics cards, laptops, automotive components, and industrial equipment began competing for a strained supply chain.

That revealed something uncomfortable: a huge part of the modern economy depended on small, complex components that were difficult to manufacture at scale.

At the same time, artificial intelligence began changing its position in the public imagination.

AI was obviously not born in 2020. The ideas that form the field are much older. Neural networks, machine learning, statistics, optimization, pattern recognition, and mathematical models had already been studied for decades. It is also wrong to reduce AI to chatbots, image generators, or synthetic videos. AI appears in demand forecasting, risk analysis, computer vision, recommendation, logistics, medicine, finance, industry, and many other places.

But the popularization of large language models made the infrastructure behind AI much more visible.

ChatGPT, Gemini, Claude, DeepSeek, Copilot, Grok, and many other names turned AI into a daily topic. Along with them, a previously neglected layer came into view: training and running models of this size requires more than good algorithms. It requires massive computing capacity.

That capacity depends on GPUs, specialized accelerators, memory, high-speed networks, data centers, energy, cooling, and continuous access to advanced chips.

AI can look immaterial when it appears in a chat window. But behind a response generated in seconds lies enormous physical infrastructure.

On the front end, it looks like a text box.

At the back end of civilization, it looks like a race for chips, energy, cooling, and money.

This was when hardware stopped being backstage infrastructure and returned to the center of the technology dispute.

3. AI’s hunger for computing

The race for artificial intelligence made one thing obvious: computing became a strategic resource.

Training larger models, running inference for millions of users, reducing latency, maintaining global availability, and competing on performance requires an absurd amount of hardware. It is no accident that NVIDIA became one of the world’s most important companies. In the new gold rush, the people selling the shovel get rich too.

For decades, many people associated NVIDIA mainly with gaming. Graphics cards were wanted by people who wanted better gaming performance, worked with computer graphics, or needed to accelerate specific tasks. But GPU architectures, designed to handle parallel processing well, became essential to modern AI workloads.

AI models depend on many mathematical operations running in parallel. That made GPUs and specialized accelerators central pieces of contemporary technology infrastructure.

The result was an explosion in demand.

Companies that want to compete in AI need chips. Anyone who wants to train larger models needs chips. Anyone who wants to offer AI as a product needs chips. Anyone who wants to host AI applications needs chips. Anyone who wants technological sovereignty in AI also needs chips.

NVIDIA became a symbol of this shift. In fiscal 2026, the company reported annual revenue of US$215.9 billion, up 65% from the previous year; fourth-quarter revenue reached US$68.1 billion. In more recent results, data-center revenue became the largest part of the company’s business, showing how AI infrastructure became its economic center.

Generative AI looks like magic until you remember that behind the chatbot’s friendly answer, an indecent amount of GPU is sweating somewhere in a data center.

And then the problem appears: advanced chips are not software.

You cannot simply copy a file, open a repository, deploy more containers, or create a new cloud account. Somewhere in the world, someone has to physically manufacture the components that make this computing possible.

Manufacturing those components is one of the most complex industrial tasks humanity has ever undertaken.

4. Chips do not scale like code

People who work with software get used to a certain elasticity in the digital world.

If an application needs to serve more users, we can optimize code, scale horizontally, use caches, distribute services, automate deployments, hire more cloud capacity, partition databases, use queues, replicate instances, create pipelines, and pretend everything is under control now.

Sometimes it works.

Sometimes we merely move the problem somewhere else and give it a beautiful name.

But, in general, software scales through copying, automation, and networks.

Hardware does not.

Hardware scales through factories, capital, energy, machines, materials, logistics, specialized labor, and time.

A chip factory does not appear with a git clone. A data center does not appear with a Terraform script. A lithography machine is not installed as a project dependency. Yield does not improve through political will alone. And a highly specialized supplier network cannot be reproduced overnight.

Unfortunately, nobody has invented docker compose up asml yet.

If hardware scaled like software, someone would already have deployed a TSMC on Kubernetes.

This contrast is central.

OpenAI, Google, Microsoft, Meta, Amazon, and other giants may have the best researchers, engineers, and a great deal of money. Still, if the physical capacity for computing is limited, there is a bottleneck.

And that bottleneck is not only in the final chip. It is in available energy, the power grid, cooling, data-center construction, memory supply, logistics, foundry capacity, semiconductor manufacturing equipment, and the availability of highly specialized professionals.

Modern data centers consume enormous amounts of energy. They also require constant cooling. In some cases, they depend on large volumes of water or complex cooling systems. This means the AI race pressures not only the chip market, but also energy infrastructure, urban planning, water resources, and industrial supply chains.

Artificial intelligence, which often appears to users as pure abstraction, is also a struggle over matter, energy, and space.

This is where the title of this text begins to make sense.

Whoever has no chip hunts with software.

That sentence is a provocation. Software remains fundamental. But the 2020s showed that software alone cannot sustain the race for cutting-edge technology.

5. ASML, Taiwan, and the invisible bottlenecks

It would be easy to look at NVIDIA and think that everything begins and ends there.

But NVIDIA does not manufacture the most advanced chips it sells by itself. It designs, integrates, builds a software ecosystem, and sits at the center of AI demand. The physical manufacturing of those chips, however, involves other companies and an extremely specialized global chain.

That is where names less familiar to the general public become absolutely central to contemporary technology.

TSMC, headquartered in Taiwan, is one of the world’s most important manufacturers of advanced semiconductors. Many of the most sophisticated chips used by American, European, and Asian companies depend on Taiwanese manufacturing capacity. This makes Taiwan not only a geopolitically sensitive island, but also a critical point in global technology infrastructure.

Another essential name is ASML.

ASML is a Dutch company that manufactures lithography machines used to produce semiconductors. It does not manufacture the chips themselves. It makes some of the machines without which the most advanced chips could not be produced at the required scale and precision.

That difference matters.

In the semiconductor world, power does not only come from selling the final product. Power also comes from controlling the instruments required for the product to exist.

Calling an EUV machine a “printer” is technically unfair, but hard to resist as an explanation.

Think of a printer. Now forget the printer, because it probably jammed, asked you to replace the toner, and would not help anyway.

An EUV machine operates with precision optics, lasers, vacuum, special materials, software, metrology, and a global chain of specialized suppliers. Instead of printing an invoice, it helps print the future of computing.

And it costs hundreds of millions of dollars.

ASML describes its EUV systems as equipment that uses extreme ultraviolet light to enable mass production of the world’s most advanced microchips. The company also explains lithography as a projection system: a pattern passes through a mask, is reduced, and is projected onto a photosensitive silicon wafer.

ASML itself depends on other critical companies. Extremely precise mirrors, optical components, lasers, vacuum systems, chemical materials, sensors, mechanical parts, and specialized software come from different countries and suppliers. The result is a machine that no company or country can easily reproduce in isolation.

This is important: the semiconductor supply chain is global, but its bottlenecks are not distributed equally.

The United States has enormous strength in design, architecture, EDA, IP, fabless companies, and computing platforms. Taiwan plays a central role in advanced manufacturing. The Netherlands has ASML. Japan dominates important parts of materials and equipment. South Korea is extremely strong in memory and manufacturing. China has enormous industrial scale, a huge domestic market, aggressive state policy, and technological ambition, but still faces bottlenecks in critical parts of the chain.

This structure creates a peculiar kind of power.

Where there is a bottleneck, there is influence.

And where there is dependence, there is vulnerability.

6. China beyond the copying cliché

There is a political component to this debate that is not the whole subject of the article, but is impossible to ignore.

The debate about China often seems divided between two characters: the uncle who says “the Chinese only copy” and the cousin who says “China will dominate everything by Thursday.”

Between these two extremes lies a little-explored region called reality.

Many people still look at China through old lenses: the China of cheap copies, low-cost manufacturing, disposable products, the factory of the world that merely reproduced Western ideas with inexpensive labor.

That image may have explained part of reality at some point. But it is insufficient for understanding China today.

China now competes in electric cars, batteries, solar energy, drones, telecommunications, artificial intelligence, supercomputing, and semiconductors. Chinese companies stopped being mere imitators in several markets and began competing on price, scale, speed, vertical integration, and, in some cases, innovation.

This does not mean romanticizing China, ignoring internal problems, erasing labor, political, or authoritarian issues, or assuming the country will inevitably dominate everything. It simply means recognizing that the idea of a China that “only copies” has become too poor to explain the present.

Semiconductors show this complexity clearly.

On one hand, China has enormous industrial capacity, state planning, a gigantic domestic market, universities, engineers, researchers, ambitious companies, and a government willing to invest heavily to reduce external dependencies.

On the other hand, advanced semiconductors are not an ordinary industry. It is not enough to provide money, build buildings, and hire people. You must master decades of accumulated knowledge, critical suppliers, precision equipment, materials, design software, manufacturing processes, yield, intellectual property, and practical experience.

China understood that depending on foreign technologies in critical areas is a strategic vulnerability. Sanctions and restrictions imposed by the United States reinforced that perception. When a country realizes that its access to chips, machines, software, and suppliers can be restricted by external political decisions, the search for autonomy stops being merely an industrial project.

It becomes a question of sovereignty.

But there is a difference between wanting autonomy and fully achieving it.

Underestimating China is a mistake. Believing that scale, money, and political will solve every problem is also a mistake.

Semiconductors show exactly this: modern technological power depends as much on ambition as on ecosystem. And complex ecosystems are not built overnight.

7. The blockade as an accelerant

Before sanctions hardened, China was not exactly isolated from the global semiconductor chain.

Chinese companies had already bought lithography machines from ASML, especially DUV equipment. These machines are not the absolute top end used in the most advanced processes, such as EUV systems, but they remain extremely important for many stages of the semiconductor industry. Even in a chain dominated by cutting-edge technologies, less advanced equipment can still help produce mature chips, adapt processes, learn factory engineering, and build industrial capability.

China also had access to NVIDIA GPUs, although that access began facing increasingly strict restrictions.

The important point is that the blockade did not hit a country starting from zero. It hit an industrial power that had already accumulated machines, factories, engineers, researchers, domestic demand, and strategic ambition.

The Western effort, led mainly by Washington, is clear: make it harder for China to access the most advanced chips, the machines capable of manufacturing them, and the technologies that could accelerate military applications, supercomputing, and artificial intelligence. The U.S. Department of Commerce’s Bureau of Industry and Security says that, since October 2022, it has published rules restricting China’s ability to buy and manufacture certain advanced semiconductors critical to military applications, with later updates in 2023 and 2024.

But sanctions have a double effect.

They delay. They create bottlenecks. They increase costs. They make supplier access harder. They force Chinese companies to redesign products, use less advanced chips, seek alternative routes, or depend on domestic solutions that are still inferior.

At the same time, they accelerate the perception of dependence.

Blocking access can delay an industry.

It can also create a room full of irritated engineers, a state budget, and a national goal written in enormous letters.

When a country realizes that its access to GPUs, lithography machines, design software, and critical components can be interrupted by political decisions made outside its territory, the search for autonomy stops being merely an industrial project. It becomes a question of sovereignty.

This is where China seems to have turned semiconductors into a national priority.

Recent reports indicate that Chinese lithography projects sought to recruit former ASML engineers and scientists to accelerate the development of domestic technologies. According to Reuters, a Chinese team may have completed a functional EUV-machine prototype in early 2025 with the participation of former ASML engineers; the same report says the machine was still being tested and had not produced functional chips.

That detail matters: copying or recreating an ASML machine is not like copying an ordinary electronic product.

An EUV machine combines precision optics, light sources, lasers, chemistry, software, mechanics, metrology, sensors, vacuum systems, special materials, and a global chain of extremely specialized suppliers. No engineer knows everything. No small team masters every subsystem. ASML’s value is not only in the design of a machine, but in the ecosystem that lets it work, be maintained, evolve, and produce with industrial reliability.

In software, when something goes wrong, we blame the legacy.

In hardware, sometimes the legacy is a forty-year industrial chain that no country can recompile from scratch.

At the same time, it would be naive to imagine that China will simply accept this dependence.

A country with more than a billion people, strong STEM education, a large industrial base, state investment, and enormous numbers of engineers, physicists, chemists, and mathematicians can attack the problem from many directions at once. It can hire specialists, fund research, buy older equipment, study parts, create domestic substitutes, train new generations, and accept costs that private companies might not accept on their own.

My impression is that the question is not whether China will try to reach Western bottlenecks.

It is already trying.

The question is how long that will take, how much it will cost, and whether it can achieve not only a functional prototype, but industrial capability comparable in scale, efficiency, and reliability.

And perhaps that is exactly what worries Washington.

Sanctions may buy time for the West. But they may also accelerate China’s effort to turn dependence into autonomy. The blockade does not eliminate Chinese ambition. It only makes that ambition more urgent.

8. Dollars, sanctions, and the infrastructure of power

The semiconductor dispute does not happen only inside factories.

It also passes through currency, international trade, financing, alliances, export controls, sanctions, and jurisdiction. In advanced technology, power is not only the ability to produce. It is also the ability to decide who can buy, sell, finance, import, export, and access particular components.

This is often underestimated by people who look only at the technical side.

The global semiconductor chain depends on companies spread across many countries, but many of them are connected to markets, currencies, trade rules, and financial systems in which the United States has deep influence. The dollar remains a central part of the international economic system. According to IMF data, the dollar still represented 56.77% of allocated official foreign-exchange reserves in the fourth quarter of 2025.

That gives the United States leverage beyond its domestic industry.

When the United States restricts exports of certain advanced chips to China, or limits access to equipment and technologies used to manufacture semiconductors, it is not merely regulating trade. It is using technological control as an instrument of power.

This becomes even clearer when restrictions try to reach not only American companies, but allied suppliers as well. In April 2026, U.S. lawmakers introduced a proposal to restrict exports of chip-manufacturing equipment to China, including technologies from companies in allied countries such as ASML, seeking to close loopholes in the sale and maintenance of machines used by Chinese manufacturers. In May 2026, China criticized the proposal, which targeted key tools supplied by foreign companies such as ASML.

This shows that technological infrastructure and financial infrastructure overlap.

A country may have engineers, companies, and demand. But if it depends on suppliers subject to sanctions, machines produced in countries allied with the United States, software controlled by foreign companies, or transactions made in a financial order dominated by the dollar, its autonomy has limits.

The struggle over cutting-edge technology is not only a struggle between companies.

It is a struggle between states, supply chains, currencies, alliances, and development models.

That is why semiconductors have become so important. They sit at the intersection of industry, defense, science, artificial intelligence, telecommunications, energy, finance, and geopolitics.

Whoever controls the infrastructure of advanced computing does not control only a market. They control part of the ability to innovate, automate, simulate, monitor, defend, produce, and influence.

9. Software still matters, but it does not rule alone

“Whoever has no chip hunts with software” is a provocation.

It does not mean that software stopped mattering. Quite the opposite. Software remains one of the most powerful layers of modern technology. Through it, we interact with systems, automate processes, create products, analyze data, train models, connect services, and turn infrastructure into experience.

But the 2020s made it clear that software does not explain everything by itself.

AI needs computing. Computing needs chips. Chips need factories, machines, energy, water, materials, suppliers, logistics, capital, currency, and geopolitical stability.

The great change may be this: the digital world has revealed its physical foundation again.

For years, it seemed that the future belonged only to whoever wrote the best code, created the best application, controlled the data, or built the best platform. It is increasingly clear that the struggle is larger.

It involves who can manufacture chips.

Who can buy the machines.

Who controls the bottlenecks.

Who finances the infrastructure.

Who has enough energy.

Who has strategic alliances.

Who can impose sanctions.

Who can turn industrial capacity into computing capacity.

Artificial intelligence accelerated this perception. It made computing into a visible strategic resource. It is not enough to have good models. You need the ability to train, host, distribute, and operate them at scale.

China is central in this context not because it is invincible, nor because it is doomed to lag behind. It is central because it exposes the most important question in today’s technology dispute: what happens when an industrial power realizes that its dependence on foreign chips, machines, and software may limit its autonomy?

The West tries to preserve its advantage. China tries to reduce dependence. Companies try to protect their markets. Governments try to control bottlenecks. And through it all, AI increases the hunger for computing.

Perhaps this is the main point: cutting-edge technology is not only code. It is industry, territory, energy, currency, logistics, and power.

Software may continue to eat the world.

But someone has to manufacture the teeth.