Washington tried to place a technological ceiling over China’s AI ambitions. Instead, it may have helped create one of the most extraordinary engineering races of the modern semiconductor era.

For years, the conventional wisdom was straightforward: artificial intelligence runs on advanced chips, and the world’s most advanced chips depend on an extraordinarily concentrated global supply chain involving American chip design and software, Dutch lithography equipment, Taiwanese manufacturing, South Korean memory and a constellation of specialised technologies across U.S.-aligned economies. Cut China off from enough of that ecosystem, the theory went, and its ability to compete at the frontier of artificial intelligence would slow dramatically. That theory was not foolish, and in several important respects the restrictions have worked. But something else happened as well. China began redesigning the problem itself.

The U.S. began imposing sweeping advanced-computing and semiconductor-manufacturing controls on China in October 2022, later tightening those restrictions and expanding them to semiconductor equipment and high-bandwidth memory, or HBM, the ultra-fast memory essential to modern AI accelerators. The intention was clear: make it considerably harder for Chinese companies to acquire or manufacture the hardware required for cutting-edge AI. China could not simply reproduce the existing Nvidia-TSMC model overnight, so its engineers increasingly pursued another route. Rather than asking, “How do we immediately manufacture the world’s best individual GPU?”, the question became: “How much computing power can we extract from the technology we can manufacture?” That distinction is enormously important, and Huawei’s answer has been particularly striking.

One of the clearest demonstrations arrived with Huawei’s CloudMatrix 384, a system that connects 384 Ascend 910C AI processors into a tightly integrated computing architecture. Individually, those processors are not equivalent to Nvidia’s most advanced accelerators, so Huawei attacked the problem at the system level: more chips, extremely fast interconnection, large-scale clustering, software optimisation and architectural engineering designed to make the entire machine perform as something greater than its individual processors. When Huawei publicly demonstrated CloudMatrix 384 in Shanghai in 2025, analysts cited by Reuters said the system could outperform Nvidia’s GB200 NVL72 on certain metrics, despite the individual Huawei processors being less powerful. That is perhaps the most fascinating feature of China’s response to semiconductor restrictions. It is not simply attempting to copy the American technology stack component by component. It is increasingly trying to engineer around its weaknesses. If transistor density is difficult, improve packaging. If one processor is weaker, connect hundreds. If manufacturing efficiency is lower, optimise the whole computing system. If CUDA is a barrier, build alternative software ecosystems and compatibility layers. The contest is gradually moving from individual chips toward complete computing platforms.

Huawei’s Ascend processors have become central to this effort. The Ascend 910C emerged as China’s most prominent domestic alternative to Nvidia accelerators, even though producing it has been difficult and manufacturing yields have reportedly been a major challenge because Chinese semiconductor fabrication lacks unrestricted access to the world’s most advanced production equipment. Yet the development cycle has continued. Huawei has said its newer Ascend 950DT accelerator will become available in the fourth quarter of 2026, and the constraints surrounding it reveal both how impressive China’s progress has become and how far it still has to go. As of September 2026, Reuters reported that Huawei had raised indicated prices for the 950DT to more than 250,000 yuan, or roughly $37,000, amid soaring prices and shortages of high-bandwidth memory. That is an important reality check: China has not made the semiconductor blockade disappear. It has learned to keep advancing while carrying it.

An even bigger development is taking place beyond Huawei. China is developing an increasingly broad collection of AI-chip companies, including Cambricon, MetaX, Moore Threads, Biren, Iluvatar CoreX and Tencent-backed Enflame. In September 2026, Enflame prepared to list in Shanghai after raising roughly $912 million, valuing the company at around $9.1 billion. The significance is larger than any one company. The strategic consequence of export controls may not simply be Huawei producing a Chinese alternative to an Nvidia processor; it may be the creation of an entire domestic semiconductor ecosystem that might otherwise have taken much longer to emerge. Customers who once had little reason to abandon mature Western hardware suddenly have an enormous reason to test Chinese alternatives. Chinese cloud companies have a reason to optimise software for domestic chips, investors have a reason to finance them, universities have a reason to train engineers around them, memory manufacturers have a reason to accelerate development, and Beijing has a powerful reason to subsidise practically every layer of the supply chain. Restrictions changed the economics of technological independence.

One of the hardest remaining problems is high-bandwidth memory. Modern AI computing is not simply about performing enormous numbers of calculations; processors must also be fed enormous quantities of data extremely quickly, which makes HBM critical. The United States specifically expanded controls to advanced HBM in late 2024, and in 2026 this remains a painful constraint. Reuters reported in September that Chinese AI-chip manufacturers were sharply raising prices because of HBM shortages, with Huawei and Cambricon both affected. This is evidence that export controls have teeth. But China is simultaneously pouring resources into domestic memory, with ChangXin Memory Technologies, better known as CXMT, emerging as the country’s leading DRAM manufacturer and one of the most strategically important companies in China’s attempt to build a more self-contained semiconductor supply chain. The next great contest may therefore be fought not merely over GPUs, but over memory, packaging, interconnects and semiconductor manufacturing equipment.

Perhaps the boldest idea coming from Huawei is that China’s disadvantage in semiconductor fabrication does not necessarily have to be overcome by following exactly the same path as TSMC, Nvidia and the Western semiconductor ecosystem. Huawei has publicly promoted what it calls the “Tau Scaling Law,” sometimes referred to as “Her’s Law,” emphasising system-level improvements rather than relying exclusively on ever-smaller semiconductor process nodes. Huawei has made ambitious claims about how far this approach could eventually go, and outside analysts remain understandably sceptical until more independently verified performance data appears. But the philosophy itself makes sense. There are two ways to respond when somebody blocks the road ahead of you: you can spend all your effort trying to reopen that road, or you can build another one. China appears to be doing both.

It would be tempting to declare the American strategy a failure, but that would be premature. Chinese companies still face serious disadvantages. Cutting-edge semiconductor-manufacturing equipment remains difficult to obtain. Manufacturing yields matter enormously. Energy consumption matters. HBM availability matters. Producing huge quantities of accelerators economically is very different from demonstrating an impressive prototype. Even U.S. officials who acknowledged China’s rapid progress estimated in 2025 that Huawei’s production of advanced AI processors remained limited relative to China’s enormous demand. The restrictions have therefore increased costs and slowed expansion. What they have not done is freeze Chinese innovation, and that distinction matters. There is an enormous difference between preventing a competitor from advancing and forcing that competitor to advance through a more difficult route. So far, China appears to be doing the latter.

Interestingly, U.S. policy itself has become more nuanced. In January 2026, the Commerce Department changed its licensing policy to allow applications for exports of Nvidia H200, AMD MI325X and comparable processors to approved Chinese customers to be considered on a case-by-case basis, subject to security and compliance requirements. That highlights a difficult strategic calculation for Washington. Completely excluding American companies from the Chinese market can restrict China’s access to advanced technology, but it can also create an enormous protected market for China’s domestic semiconductor champions. Every Nvidia accelerator that Chinese companies cannot purchase increases the incentive to make Ascend, Cambricon, Enflame or another domestic platform work. And once an alternative ecosystem becomes good enough, restoring access to American products does not necessarily restore American dominance. Technology markets have memory. Supply chains develop, software gets ported, engineers acquire expertise, customers become comfortable with alternatives and billions of dollars of manufacturing infrastructure gets built. Eventually, temporary necessity can become permanent industrial capability.

The Chinese AI-chip story is therefore becoming far more important than a simple Huawei-versus-Nvidia comparison. It is an extraordinary real-world experiment in technological adaptation. The world’s semiconductor ecosystem spent decades becoming globally interconnected and extraordinarily specialised. Washington then attempted to deny China access to strategically important pieces of that network, and China’s response has been to accelerate the construction of another one. It is less advanced in crucial areas, frequently less efficient and still dependent on technologies where substitutes are difficult to manufacture, but every year that ecosystem becomes deeper. Huawei’s CloudMatrix demonstrates what can happen when engineers compensate for semiconductor limitations at the system level. CXMT demonstrates how quickly strategic investment can build capacity in memory. Companies such as Cambricon, Enflame, MetaX and Moore Threads demonstrate that China’s answer is becoming an industry rather than a single national champion.

That may ultimately be the most important unintended consequence of the semiconductor confrontation. America attempted to make advanced AI computing harder for China. It succeeded. But making something harder is not the same as making it impossible. Under extraordinary pressure, China is learning how to design chips, computers, software and supply chains for a world in which access to American technology can no longer be assumed. That makes the next chapter of the AI race much harder to predict, because China is no longer simply racing to catch the leader on the same track. It is starting to build its own track.