China’s New Chip Strategy Could Change the Rules of the AI Race

Washington tried to hold China back by denying it the world’s most advanced chipmaking equipment. Instead of surrendering, Chinese engineers are redesigning chips, software and entire computing systems around the technologies they can still obtain.

For much of the past decade, the semiconductor race has been described as a contest measured in nanometres. The country capable of building the smallest transistor was assumed to hold the future of computing. That assumption shaped American policy as much as it shaped the technology industry.

Beyond the nanometre

Beginning in 2019, Washington imposed increasingly severe export controls intended to prevent China from acquiring the most advanced semiconductor manufacturing equipment. The logic appeared straightforward. Without the latest lithography machines, China could not manufacture the processors required for advanced artificial intelligence. Without those processors, it would struggle to compete with Nvidia, Intel and TSMC.

Seven years later, those controls have unquestionably slowed China. They have raised manufacturing costs, reduced production yields and denied Chinese companies access to some of the most sophisticated industrial machinery ever built.

But they have also produced an unintended consequence.

Instead of trying only to recreate the Western semiconductor industry, Chinese engineers have begun asking a different question: if they cannot compete by making ever-smaller transistors, can they compete by redesigning the computer itself?

The blockade worked but not quite as intended

It is tempting to describe America’s semiconductor strategy as either a complete success or a complete failure. Neither description is accurate.

Chinese manufacturers remain unable to buy the most advanced extreme ultraviolet lithography machines made by the Dutch company ASML. These machines use exceptionally short-wavelength light to etch microscopic circuits onto silicon wafers and are indispensable to the most efficient production of leading-edge processors.

Without them, Chinese foundries must rely heavily on older deep ultraviolet equipment. Advanced chips can sometimes be reproduced through repeated exposures and additional manufacturing stages, but every added step increases complexity, lowers yields and raises costs.

A Chinese factory may therefore manufacture a highly advanced processor, but it will generally produce fewer working chips from each wafer than a comparable Taiwanese or Western production line.

The controls have made advanced semiconductor production substantially harder. Yet if China cannot compete on manufacturing precision alone, it must compete elsewhere.

Increasingly, that means architecture.

Rethinking what makes a computer fast

For decades, semiconductor progress followed a simple rule. Smaller transistors allowed engineers to place more of them onto a chip, producing greater computing power. This became known as Moore’s Law and drove much of the digital revolution.

Modern processors, however, are no longer limited only by the speed of individual transistors.

Imagine a city filled with exceptionally fast cars but hopelessly congested roads. Building faster cars achieves little if every vehicle still spends most of its journey trapped in traffic.

Modern chips face a similar problem. Increasingly, performance is constrained by the time and energy required to move information between processors, memory and different sections of the same circuit.

Huawei argues that this movement of information has become one of computing’s central bottlenecks. Its proposed Tau Scaling Law shifts attention away from transistor size and towards the time required for information to travel through an entire system.

Instead of concentrating only on smaller transistors, engineers would shorten signal paths, redesign circuit layouts, improve communication between processors and optimise hardware and software together.

Whether Tau Scaling becomes as influential as Huawei hopes remains uncertain. But the principle behind it is significant. China is no longer merely trying to follow the technological roadmap drawn in Silicon Valley. It is attempting to construct another one around the tools and industrial advantages it possesses.

Folding the chip instead of shrinking it

One example is a technique Huawei calls LogicFolding.

Imagine writing a sentence across a sheet of paper. Fold the paper and words that were previously far apart suddenly sit beside one another.

Huawei’s engineers are attempting something broadly similar with electronic circuits. Instead of spreading every component across a flat silicon surface, sections of the circuit are arranged in vertically stacked layers connected by microscopic links. Signals that once travelled long horizontal distances can move across much shorter vertical ones.

Huawei says the approach reduces delay, improves energy efficiency and increases effective circuit density without requiring an entirely new lithography process.

Western engineers remain cautious. Three dimensional integration is not new. Intel, AMD, TSMC and other companies have spent years developing chiplet designs, hybrid bonding and advanced packaging.

The important question is therefore not whether stacking works, but whether Huawei’s implementation represents a major architectural advance or a sophisticated adaptation of techniques already used elsewhere.

Independent examination of future commercial processors will be necessary before the company’s more ambitious claims can be judged. Even so, the direction of travel is unmistakable. The competition is no longer confined to making a better transistor. It increasingly concerns the design of the entire computing system.

The nanometre illusion

A recent laboratory teardown of a Huawei processor illustrated how misleading simple manufacturing comparisons have become.

Analysts examining the chip’s microscopic metal wiring found some dimensions tighter than those measured in one of Intel’s newer production processes.

Taken alone, that sounds extraordinary. It does not mean that Huawei or its manufacturing partners have overtaken Intel, TSMC or Samsung.

The processor still trails leading competitors in overall transistor density, manufacturing efficiency and production yield. Its compact wiring was achieved through repeated manufacturing steps that are slower, more expensive and harder to execute consistently.

The result nevertheless demonstrates a larger point. Modern computing performance is no longer captured by one number printed on a brochure. Architecture, packaging, memory, networking and software increasingly matter almost as much as the nominal size of the transistor.

The computer becomes the chip

Bigger, not necessarily better but perhaps good enough

If Huawei’s chip strategy challenges conventional thinking about semiconductor design, its approach to artificial intelligence infrastructure challenges another assumption: that the fastest computer must be built around the fastest individual processor.

Nvidia has dominated artificial intelligence by producing extraordinarily powerful graphics processors. China has struggled to match that performance chip for chip.

Huawei appears to be responding by changing the scale of the contest.

Rather than waiting until one Chinese processor can equal Nvidia’s best, it is building systems that combine thousands of less powerful processors into something approaching a single machine.

Imagine two construction companies. One owns the world’s largest crane. The other owns thousands of smaller cranes capable of working together. The first machine is more efficient, but the second company may still complete the project if it can coordinate its equipment effectively.

Huawei’s Atlas 950 SuperPod follows that philosophy. It links more than 8,000 domestically produced AI processors through a high-speed optical network. The completed installation occupies roughly the floor space of two basketball courts and contains long rows of cabinets connected by fibre-optic links.

To Western engineers, such a system can appear extravagant. It uses more processors, more networking, more floor space, more electricity and more cooling than a comparable system built around more advanced chips.

But China possesses several resources in abundance: industrial land, large power systems, construction capacity, state-backed finance and government customers willing to purchase domestic equipment.

That changes the economics. A less efficient system may still be strategically attractive if it can be built at scale, financed domestically and operated without dependence on foreign suppliers.

The question is therefore no longer whether one Huawei processor is better than one Nvidia processor. It is whether thousands of Chinese processors can together perform enough useful work to sustain China’s domestic AI industry.

Early evidence suggests the answer may increasingly be yes for at least some workloads: not as efficiently, not as elegantly, but perhaps sufficiently.

Hardware is only half the battle

Even if China produced a processor equal to Nvidia’s tomorrow, another obstacle would remain.

Software.

Much of modern artificial intelligence rests on Nvidia’s CUDA platform. Developed nearly twenty years ago, CUDA allows programmers to use Nvidia graphics processors for scientific computing and machine learning.

It has since become far more than a programming tool. Universities teach it. Researchers publish code built around it. Companies have spent years developing applications and internal systems that depend upon it.

CUDA’s strength does not arise because replacing it is technically impossible. It arises because almost everyone already uses it.

That creates a powerful network effect. Moving to another platform can require years of code to be rewritten, tested and maintained. It is rather like asking an entire country to begin driving on the opposite side of the road. The vehicles may still work, but nearly every supporting system must change.

Chinese researchers increasingly identify software compatibility, rather than chip manufacturing alone, as one of the country’s principal weaknesses.

That is why China is investing in AI-assisted software translation. Instead of rewriting every program manually, new tools attempt to convert applications built for Nvidia hardware into software capable of running on Chinese processors.

These systems remain imperfect. Complex scientific applications still require substantial human intervention, and no serious analyst believes CUDA’s dominance is about to vanish.

But technological independence does not require CUDA to disappear. It requires the cost of leaving it to fall.

If software migration can eventually be measured in days rather than months, more Chinese institutions and businesses will consider domestic hardware where previously they would not.

The forgotten bottleneck

Processors receive most of the attention, but advanced artificial intelligence also depends on enormous quantities of specialised high-bandwidth memory, or HBM.

HBM moves data between processors at extraordinary speed. Without it, even a powerful AI chip may spend much of its time waiting for information.

This has been one of China’s greatest vulnerabilities. Most leading HBM is produced by companies in South Korea and the United States, giving Washington another point of leverage.

China’s response has centred partly on ChangXin Memory Technologies, better known as CXMT. The company has expanded rapidly in conventional DRAM and is spending heavily to develop advanced memory for artificial intelligence.

The challenge remains formidable. HBM is produced by stacking multiple memory layers, connecting them through thousands of vertical pathways and manufacturing the completed module with exceptional precision.

Chinese production yields are still believed to remain below those of established competitors. Each successful module is therefore more expensive, and large-scale reliability remains difficult.

This is a reminder that China’s semiconductor story is not one of effortless breakthroughs. Progress is real, but so are the penalties: higher costs, lower yields, greater power consumption and increased engineering complexity.

The contest is not simply between success and failure. It is between efficiency and independence.

China appears increasingly willing to sacrifice some of the former to obtain more of the latter.

Looking beyond the transistor

China is also investing in technologies that may shape computing beyond conventional silicon.

Chinese universities have become important centres of photonic computing research, which uses light rather than electricity to perform certain specialised calculations.

Laboratory processors have demonstrated impressive speed and energy efficiency in tasks involving the large mathematical operations used in neural networks and image processing.

These machines are not replacements for general-purpose computers. They cannot yet run operating systems, databases or the full range of applications required by modern computing, and they remain far from mass commercial deployment.

Their importance is strategic rather than immediate. China is not investing only in catching up with today’s semiconductor technology. It is also exploring technologies that could matter in the generation after next.

The semiconductor contest is therefore being fought across chip architecture, advanced packaging, optical networking, software, memory and entire computing systems. It is a much broader engineering struggle than the race to manufacture a smaller transistor.

Two computing worlds

It is easy to become lost in the language of semiconductor manufacturing.

Metal pitch. Hybrid bonding. Optical interconnects. High-bandwidth memory.

To engineers, these are precise technical terms. To everyone else, they may sound like matters for specialist journals. Together, however, they describe a far larger change.

For more than thirty years, semiconductors were perhaps the most globalised industry on Earth.

An American company designed a processor. Dutch engineers built the machines that manufactured it. Taiwanese factories etched it into silicon. Japanese companies supplied specialist chemicals and materials. South Korean firms produced the memory. American software made the system useful.

No single country controlled the entire chain because no single country needed to.

That model is now fragmenting.

American export controls were intended to preserve Western technological leadership by denying China access to critical parts of that system.

In one respect they have succeeded. China still cannot manufacture advanced chips as efficiently as Taiwan, South Korea or the United States. It relies heavily on older production equipment. Its manufacturing costs are higher and its yields lower. Nvidia, TSMC and ASML remain at the technological frontier.

But technological leadership is no longer the only question.

The more important question may be whether China still requires access to every part of the Western ecosystem.

Beijing increasingly appears to be answering with a qualified no.

Instead of duplicating every Western technology perfectly, China is assembling an ecosystem that may be sufficiently capable to stand on its own.

Its processors may remain less efficient. Its memory may cost more. Its software may depend on translation rather than native compatibility. Its data centres may occupy more space and consume more electricity.

Yet if the complete system performs well enough for Chinese industry, universities and government, strategic dependence on foreign technology begins to decline.

This approach reflects an engineering philosophy that values resilience as well as optimisation. Rather than seeking the most elegant individual component, China is building systems intended to remain useful despite weaker parts.

In commercial markets, efficiency is often decisive. In strategic competition, independence may be worth paying for.

A technological Iron Curtain

During the Cold War, the world divided into competing political and military systems. The twenty first century may be witnessing a comparable division in technology.

One computing ecosystem will remain centred on Nvidia, AMD, Intel, TSMC, ASML and the established Western software platforms.

Another may gradually form around Huawei, SMIC, CXMT, Chinese software frameworks and domestic computing infrastructure.

The two systems will not be identical, nor equally efficient. China may remain technically behind the Western frontier for years.

But strategic competition is not always decided by the producer of the single best component. Countries often succeed by building technology that is sufficiently capable, sufficiently scalable and sufficiently independent to meet their own needs.

The semiconductor race increasingly resembles that kind of contest.

It is no longer only about who manufactures the smallest transistor. It is about who can construct the most resilient technological ecosystem.

Ecosystems are harder to sanction than individual companies. They are also much harder to stop.

The lesson of the sanctions

There is a final irony.

Washington’s export controls were not irrational. They delayed China’s progress at a critical moment and complicated Beijing’s ambitions. Without them, China’s semiconductor industry would almost certainly be further ahead.

But sanctions rarely leave their target unchanged.

Sometimes they suppress innovation. Sometimes they redirect it.

China’s response increasingly appears to be the latter.

Unable to buy the best lithography machines, Chinese engineers concentrated more heavily on architecture and packaging. Unable to rely indefinitely on CUDA, they invested in translation tools and domestic software frameworks. Unable to import every advanced memory technology, they poured money into local production. Unable to match Nvidia chip for chip, they built larger computing systems designed to compensate through scale and networking.

None of these measures has solved the original problem. Together, however, they have reduced China’s vulnerability.

Some of Huawei’s claims will prove exaggerated. Some technologies will disappoint. Others may never become commercially viable.

But the central development is becoming difficult to dismiss.

The race is no longer simply about who can build the smallest chip. It is about who defines the architecture of the computer that comes after it.

For most of the past decade, the West has watched China by asking whether it had caught up with Nvidia, ASML or TSMC.

That may now be the wrong question.

The more important question is whether China has stopped trying to follow the West’s path at all.

If it has, the semiconductor race is entering a different phase — one in which the world’s two largest economies are no longer competing to build the same computer, but different ones.