중국 엔비디아 탄생 소식을 전하며, 100% 자체 개발 기술을 바탕으로 국산화 대체를 가속할 것이라고 시나닷컴이 보도했다.

For a long time, GPUs were almost exclusively the stage of American companies.

The global graphics processor market has been firmly held by Nvidia, a company founded in 1993 in Santa Clara, California.

Since the GeForce 256 debuted in 1999, the company has built its CUDA ecosystem, through successive generations of products, into what the industry calls the "Wall of Sighs."

Looking back from 2026, however, the situation in China has quietly changed.

A group of domestic GPU companies that insist on fully independent R&D across the entire technology chain has begun to emerge. Outsiders have dubbed them "China's Nvidia," and the label "100% self-developed" carries particular weight.

Behind those words is not a patchwork of overseas IP cores, nor a licensed instruction set refined in-house, but real technical capability built up from the underlying architecture.

Domestic substitution has finally moved from slogan to industrial reality, with products on shelves and data to back it up.

A few years ago, when people talked about AI chips, nine out of ten had no idea what they were, but everyone knew Nvidia was the name to beat, and that the whole world was competing to buy its products.

At that time, in China's domestic market, roughly nine and a half of every ten AI chips sold were Nvidia's, and no one could seriously compete for its business.

Within just a few years, the picture changed. According to the research firm Bernstein, Nvidia's share of China's AI chip market fell from 95% in 2022 to 55% in 2025, and is estimated to be only 8% by 2026.

TrendForce, another research organization, goes further, projecting that by 2026 domestic chips will account for nearly 90% of China's high-end AI chip market.

Isn't that a fast change? Some people must be wondering whether domestic chips really are this capable, or whether this is just hype.

To be honest, domestic chips were not impressive a few years ago, and even running a large model was a struggle. But the past two years have been different. It's no longer one or two companies working quietly; a large group of firms has caught up.

On September 11, 2026, Enflame (燧原科技) listed on the STAR Market, with shares priced at 142.18 yuan, raising 6 billion yuan in a single offering.

With Enflame's listing, the four companies known as the "Four Little Dragons" of domestic GPUs (Moore Threads, MetaX, Biren Technology, and Enflame) had all gone public.

Three years ago, who would have imagined that China could produce four high-end GPU companies all listed at once?

Of the four, Enflame is the most distinctive. While the others still reference Nvidia's approach to some extent, Enflame has opted not to follow that path at all.

It uses a DSA architecture, which, put simply, means chips designed specifically for AI computing, without the graphics rendering features that go beyond that purpose.

Most importantly, its instruction set, operators, and drivers were all written line by line in-house, and it is deliberately not compatible with Nvidia's CUDA ecosystem.

Enflame chairman Zhao Lidong has said that constantly following closely behind others might save effort in the short term, but over time it leaves you forever vulnerable at your weak point.

Incompatibility with CUDA is certainly a hassle. The software ecosystem has to be built from scratch, and customers must readjust to using it.

But the benefits are equally clear: core technology stays in your own hands, so no one can cut off your supply just by declaring a cutoff.

Another particularly interesting company is Dongfang Suanxin.

On July 13, 2026, the company held a launch event in Shanghai, unveiling an AI chip called the DF1000.

What makes this chip unusual is its 14-nanometer process technology. Anyone familiar with the field knows that the most advanced chips today are made at 3 or 4 nanometers, so 14 nanometers appears several generations behind.

Yet Dongfang Suanxin has achieved high computing power on 14-nanometer technology, with a single card capable of 520 trillion operations per second. At the exhibition, the company even presented the claim that 14 nanometers is roughly equivalent to 4 nanometers.

How did it do this? By focusing on architecture: software-defined chip design combined with 3D stacked near-memory computing, which stacks the compute and storage units on top of each other to shorten data transfer distances dramatically and boost efficiency.

Dongfang Suanxin's founder, Wei Shaojun, is a professor at Tsinghua University who has spent two full decades researching reconfigurable computing technology.

Every stage of the DF1000, from design through manufacturing to packaging and testing, was carried out by domestic companies, with no reliance on foreign leading-edge processes or high-end memory components.

At launch, the chip had already completed stable operation in a 128-card large-scale cluster. It was not a lab sample on display but a product ready for real workloads.

The DF1000 later won the SAIL Leading Track Award at the 2026 World Artificial Intelligence Conference.

Simply put, this company has chosen a path different from everyone else's. While others race to pursue more advanced process nodes, it is using mature process technology to deliver computing power.

If this path proves viable, China's chip industry will no longer have to fixate on a single bottleneck such as lithography machines.

Beyond these two companies, there is also T-Head. T-Head is Alibaba's chip subsidiary, and in January 2026 it launched a self-developed AI chip called the Zhenwu 810E, with performance not far from that of the H20, the most powerful chip Nvidia was permitted to sell to China.

The Zhenwu 810E is already deployed at scale in Alibaba Cloud's data centers, where it is used to train the Tongyi Qianwen (Qwen) large model and serves more than 400 enterprises in China.

In May, T-Head followed up with the Zhenwu M890, a new-generation chip designed for both training and inference.

T-Head's approach differs from Enflame's. Backed by Alibaba Cloud, with its own large models and application scenarios, its chips can be put to use as soon as they are produced, so it has no worries about finding buyers.

Cambricon is another company, one that entered AI chips early. Its Siyuan 590 has already been sold at scale, and the next-generation Siyuan 690 has begun volume production and delivery.

Hygon Information has taken the route of CUDA compatibility and has developed self-designed computing interconnects that scale from ten thousand cards to one hundred thousand cards.

MetaX is even more notable. In the second quarter of 2026, it achieved profitability for a single quarter, the first of the four Little Dragons to make money.

With so many companies working simultaneously, each with its own strategies and specialties, there is real reason for confidence.

Some may ask how domestic chips suddenly emerged over the past two years. The reasons are not complicated.

First, they were forced into existence. Starting in 2022, the United States steadily tightened chip export controls, first barring the sale of the most advanced chips to China and later restricting the H20, a second-tier product.

Nvidia CEO Jensen Huang initially believed the Chinese market could not function without his company. But as restrictions piled on, Chinese customers were pushed toward domestic chips.

In 2026, a US official acknowledged that shipments of products such as the H200 to China had been very small, and that the Chinese market was no longer buying.

Put simply, domestic chips were not especially strong at the start. When supply was cut off, companies had no choice but to find their own solutions, and in the process, they actually succeeded.

Second, China's market is enormous. It is the largest AI chip consumer market in the world. In 2026, the domestic AI chip market is expected to reach 250 billion yuan, and by 2029, it is projected to reach 1.34 trillion yuan.

With a market of this size, as long as products work, there will be buyers.

China also has more internet companies, more large-model startups, and more manufacturing enterprises than any other country, and all of them are potential AI chip customers.

When customers are willing to give domestic chips a chance and work alongside suppliers on debugging and optimization, the chips improve with use.

This creates a virtuous cycle: the more people use them, the faster they improve, and the faster they improve, the more people use them.

A third factor is often overlooked: China's industrial chain is comprehensive.

Making chips is not just about design. Manufacturing, packaging, testing, supporting software, and complete server systems are all required, and the absence of any one link is a problem.

Dongfang Suanxin's ability to build a fully domestic supply chain is not due to the company alone. It is supported by an entire industrial chain behind it.

From wafer fabrication to packaging and testing, and from PCB boards to complete servers, China now has competitive companies at every step.

This full-chain advantage is one that few other countries in the world can match.

Of course, we should not only highlight the positives. Compared with Nvidia, domestic chips still have clear gaps.

The most advanced process technologies remain restricted, high-end HBM memory chips still have to be imported, the software ecosystem lags far behind CUDA, and the number of developers is nowhere near comparable.

Single-card performance has caught up to some degree, but the stability and scheduling efficiency of large-scale clusters still require more time to validate.

Many domestic chip companies are still losing money. Enflame alone posted a net loss of nearly 1.2 billion yuan in 2025. Whether these companies can keep burning cash until they become profitable is uncertain.

These problems cannot be ignored. But the gap being large does not mean there is no hope.

Three years ago, who would have imagined Nvidia's share in China falling to 8%? Five years ago, who would have expected four high-end GPU companies in China to list at the same time? Ten years ago, who would have believed Chinese companies could achieve computing power comparable to 4-nanometer chips using 14-nanometer technology? Such outcomes were unthinkable before, and now they are reality.

Technology breakthroughs are a matter of when, not if, as long as the right path is chosen, money is invested, and people are willing to put in sustained effort.

There is another point I'd like to raise. Many people assume that domestic substitution means producing an exact copy of Nvidia's product and taking over its market. That view is too narrow.

True substitution is not imitation; it means forging your own path. Enflame's DSA architecture, Dongfang Suanxin's software-defined design with 3D stacking, and T-Head's integration of chips and models are not copies of Nvidia. They are explorations of new technical routes.

If Chinese chips simply keep chasing Nvidia from behind, they will never catch up. Only by taking a different path can they truly surpass it.

These companies are already doing exactly that, and that is what makes the future most worth watching.

For ordinary people, AI chips may sound abstract, but they affect everyone.

Voice assistants on phones, recommendation algorithms for online shopping, AI-assisted medical diagnosis, and driver-assistance systems in cars all depend on computing power.

Lowering the cost of computing power will make these services cheaper and more useful. As domestic chips grow stronger, the price of computing power should fall, and ultimately, the people who benefit most will be ordinary users of phones and the internet.

Moreover, when core technology stays in your own hands, there is less risk of being cut off one day, and data security is better protected.

These are not empty platitudes but tangible benefits.

Looking ahead, the road for China's AI chip industry is still long, and it will certainly face many difficulties.

US restrictions are unlikely to ease, technical challenges will persist, and market competition will only intensify.

But the direction is now clear, and the momentum has taken shape. This is not something anyone can stop.

In four years, Nvidia lost the Chinese market, falling from 95% to 8%, while domestic chips built a foothold from nothing in the same period.

The next four years will be even more interesting.

Friends, when do you think Chinese chips will truly catch up with Nvidia? Have you used domestic AI products? Share your views in the comments and let's discuss.