New DeepSeek model marks AI milestone. Is China closing the gap with the US?

BEIJING – China’s AI darling DeepSeek has launched its new flagship model that uses Huawei’s chips, marking a milestone in Beijing’s efforts to seek technological self-reliance.

Amid the fanfare, however, China’s leading artificial intelligence models remain mere months behind those from America, and the latest release does not appear to have altered this paradigm significantly.

On April 24, DeepSeek launched its v4 model, a much-awaited update that keeps pace with its American counterparts, just hours after the release of OpenAI’s GPT-5.5.

DeepSeek also announced that v4 was designed to use Huawei’s chips, as well as its software, called CANN.

This marks the first time DeepSeek has officially announced such a deep software-hardware collaboration with the Chinese telecommunications giant, which said on WeChat that its Ascend line of processors now offers “full-stack support” for v4.

An AI “stack” refers to the host of technologies and tools needed to build, maintain and use AI models.

A report by benchmark firm Artificial Analysis showed that DeepSeek’s open-source v4 Pro ranked just behind US closed-source models such as OpenAI’s GPT-5.5, Anthropic’s Claude Opus and Google’s Gemini 3.1 Pro.

But Mr Poe Zhao, founder of the Hello China Tech newsletter that analyses Chinese tech developments, said the v4 model is less about benchmarks and more about systems.

“It shows how Chinese AI labs are learning to turn hardware constraints into a design problem, but it is still far from complete AI stack self-reliance,” he told The Straits Times.

Can Huawei chips replace Nvidia?

Mr Zhao said the compatibility with Huawei’s Ascend 950 series of supercomputers is significant if it translates into stable deployment at scale.

“It does not mean Huawei chips can immediately replace Nvidia. The more important signal is model-hardware co-optimisation. Chinese model developers and hardware vendors are trying to build a domestic deployment path under chip constraints,” he said.

AI is a critical battleground between China and the US. As Washington has sought to throttle China’s development in the semiconductor supply chain with export controls, Beijing has invested heavily in its own suppliers to address this perceived weakness.

A Reuters report in February cited an unnamed US official as saying that DeepSeek’s latest model was trained on Nvidia chips called Blackwell, which are banned in China, indicating that China may still lack the vast number of powerful, domestically made chips needed.

Chinese media outlet Caixin reported on April 21 that leading AI companies, including MiniMax, Kimi and Z.ai, have experienced recent service shortages amid limited computing power and surging demand, prompting many users to “try again later”.

DeepSeek was catapulted into mainstream conversations about China’s AI development in January 2025 when its R1 reasoning model was released. It showed that a Chinese AI model can produce comparable performance to leading US models at a fraction of the cost.

Since then, the Hangzhou-based start-up has mostly kept a low profile, although its latest release has sparked excitement in China about the potential boost it would give to the domestic supply chain, at the expense of the US.

Many cited comments by Nvidia chief executive Jensen Huang, who said in an interview on April 15: “The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation (the US).”

Mr Huang has lobbied Washington for less powerful Nvidia chips to be allowed to be sold to China, so that Chinese AI firms continue to rely on American hardware.

Associate Professor Marina Zhang, who specialises in technology and geopolitics at the University of Technology Sydney, said DeepSeek v4 does not prove that Chinese AI models have overtaken the best US closed models across the board.

“But it does narrow the gap and reinforces China’s different competitive logic: near-frontier performance at much lower cost, with stronger open-source diffusion and domestic hardware adaptation,” she told ST.

As the competition to advance frontier models heats up, American firms such as Anthropic and OpenAI have accused Chinese AI firms of distillation, a technique that essentially allows newer models to learn from the answers of a more powerful model.

Mr Chris McGuire, a senior fellow for China and emerging technologies at the US think-tank Council on Foreign Relations, said DeepSeek’s new model does not challenge the consensus on the state of the AI competition between the US and China.

US models still lead by about seven months and leading Chinese models remain dependent on US tech, he wrote on X, calling it “largely a status quo release”.

“Like all other leading Chinese models, v4 was trained using US chips and on data illicitly distilled from frontier US models. If China fully lost access to US chips and models, not to mention US and allied chipmaking tools, DeepSeek and others would likely fall much farther behind,” he added.

Prof Zhang said that distillation and access to restricted chips may have played some role in parts of China’s AI ecosystem, but added that DeepSeek’s progress also reflects real engineering capability.

While DeepSeek v4’s Huawei optimisation is a significant step towards a domestic AI stack, Nvidia still leads in areas such as chip performance and its CUDA ecosystem, she said, referring to Nvidia’s software that allows its chips to be optimised for AI workloads, which Huawei’s CANN is seeking to replace.

China can increasingly support competitive AI deployment for many commercial and industrial uses, but a complete break from the US AI stack remains years away, said Prof Zhang.

“The likely outcome is not clean decoupling but two increasingly parallel and partially interoperable ecosystems.”