Nvidia Photo: VCG

US chipmaker Nvidia is paying $6 billion to license the model-building technology of artificial intelligence (AI) start-up Poolside and hiring more than 100 of its engineers, to build an open-weight AI model able to compete with Chinese rivals such as DeepSeek and Moonshot AI's Kimi K3, The Wall Street Journal reported on Saturday. Chinese analysts said on Monday that the move marks a shift in a US industry that long bet on closed development, and one that China welcomes.

Nvidia will also invest $1 billion in Poolside separately at a $12 billion pre-money valuation, and the start-up's co-founders will stay on, the report said.

CEO Jensen Huang has framed the effort in national terms, arguing in a letter titled "Open Weights and American AI Leadership" that the AI race will be judged not by a single frontier model but by whether the US builds an open ecosystem that spreads into every sector, the report said. 

Nvidia's spending is better read as a strategic adjustment under competitive pressure than as a fundamental change of view in the US industry, Chen Jing, a vice president of the Technology and Strategy Research Institute, told the Global Times on Monday. Nvidia's logic as a chipmaker is straightforward: the more widely models spread, the greater the demand for training and inference, and the more chips it sells. Capable open-weight models lower the barrier to AI adoption and directly enlarge the market for its graphics processing units, he said. 

But the core of US model development has not turned, Chen said. OpenAI, Anthropic and Google, the real giants of the model layer, still treat architecture and weights as their moat. Nvidia is using capital and computing power to back a Western open-weight standard-bearer against Chinese models, a defensive move whose priority could fall away once competitive pressure eases. A genuine change of route would need the model-layer companies to embrace open source collectively, Chen noted. 

"I wouldn't say Chinese companies forced Nvidia into this entirely, but it is clear they have put considerable pressure on it, and Nvidia is now facing significant anxiety of its own," Xiang Ligang, director-general of the Zhongguancun Modern Information Consumer Application Industry Technology Alliance, told the Global Times on Monday.

That anxiety is structural, he said. Nvidia's revenue rests on sustaining an industry-wide expectation of continuous chip purchases and data-center construction. That expectation has been strained this year as talk of an AI bubble grows louder, while the largest US AI firms build their own data centers and keep their models closed.

The gap that prompted the deal widened despite years of US export controls on advanced chips. Chinese open-weight models took roughly 41 percent of downloads on Hugging Face over the past year, the largest national share, up from low single digits in late 2024.

The approach simply works, Xiang said: opening the weights wins support and users fast, while the developer gains capability and improves the model rapidly. That has overturned an earlier assumption, he added: "We used to think China might not be able to do models well, that it would necessarily require Nvidia chips for training. In fact, that has turned out not to be the case. China has found its own path, and that path has been quite successful."

Bloomberg reported on August 20 that with aggressive pricing, US companies such as Airbnb, food-delivery giant DoorDash and top crypto exchange Coinbase Global have all adopted Chinese models hosted on local servers. This month, Alibaba disclosed its open-weight models had accumulated more than 3 billion downloads in six months, eclipsing Meta Platforms, Alphabet and domestic peers to become the world's No.1 AI model family.

Chinese open models have meanwhile entered national-level deployments in Southeast Asia. Singapore's government-backed AI Singapore program selected Alibaba's Qwen for its latest regional model, and Malaysia's Communications Ministry hosts DeepSeek's open-source model on locally sited servers, Bernama reported.

China's first-mover advantage will inevitably be diluted as well-funded US firms borrow the architectures and algorithms Chinese companies have already published, Chen said. China will welcome the trend anyway. A larger open ecosystem lowers the technology threshold worldwide and extends the reach of Chinese models, while more US firms turning to open weights erodes the dominance of closed-source giants such as OpenAI and Anthropic. 

"What is diluted is a short-term lead; what is gained is long-term leadership of the ecosystem," said Chen.

Western discussion of AI habitually runs in terms of competition and winners, but open source is knowledge-sharing by nature and resists zero-sum logic, Chen said. DeepSeek's code has been reused directly by Silicon Valley engineers, and Llama's architecture helped Chinese firms learn to build large-language models. Differences over data governance, chip controls and standards may still push the two apart at the application and compliance layers.

"It will not split into two parallel worlds," Chen said. "Core algorithms and architecture standards will stay globally interoperable — nobody turns down an efficient architecture. More companies building open models pushes the technology forward and lets more people benefit. From that perspective, such attempts are welcome."