Walking into the offices of Silicon Valley's prominent venture capital firm Andreessen Horowitz (a16z), partner Martin Casado discovered that the artificial intelligence (AI) models used by the company's portfolio startups were likely all sourced from China.

"I would say there's an 80% probability that they're using Chinese open-source models," he added.

The transformation began in January this year when Chinese startup DeepSeek made headlines by freely releasing an advanced AI model developed at low cost, creating a sensation that shook global stock markets.

The "DeepSeek aftershocks" continued, with Chinese AI models from tech giants like Alibaba and other companies quietly gaining increasing attention overseas.

The Economist analyzed on the 21st that unlike US giants competing to spend vast sums trying to crack competitors' proprietary models, Chinese companies are waging a fundamentally different war. In the words of Stanford AI expert and "father of Google Brain" Andrew Ng, this represents a "Darwinian struggle" among Chinese large language model (LLM) developers over who can be more open.

The Economist noted that in various intelligence tests released this year, Chinese open-source models outperformed similar models from US companies like Meta, with capabilities increasingly approaching top-tier proprietary models. The publication argues that the competitive enthusiasm of Chinese developers should serve as a wake-up call for the West.

Take ChatGPT developer OpenAI as an example. CEO Sam Altman recently admitted in an interview that China's "open-source battle" has put enormous pressure on OpenAI, forcing it to change its model release strategy.

In the mid-2010s, OpenAI promoted the concept of greater "openness" in AI, which was reflected in its company name. However, after 2020, to maintain profitability, OpenAI became less "open," instead selling only its proprietary large language models and refusing to fully open-source its technology.

Recently, Altman suddenly realized something was wrong—his clients' usage of open-source and open-weight models, including Chinese models, had significantly increased.

Seeking to capture a share of this market, OpenAI quickly launched two open-weight models in August: gpt-oss-120b and gpt-oss-20b, available for free and supporting developer customization.

This marked OpenAI's first release of open-weight models since GPT-2 in 2019, and the first such models to emerge since signing an exclusive cloud services agreement with Microsoft six years ago.

US media at the time characterized this as a "major strategic shift" for the company that had long maintained closed management of its technology, with OpenAI betting on "improving technology accessibility" to expand its developer ecosystem and strengthen its advantage over Chinese competitors.

The detail of using lowercase letters in these model names was quite meaningful. They were not only relatively small in scale but also received mixed market reactions. Many voices criticized them for "lacking highlights" and being stripped of many core powerful features from OpenAI's commercial products.

In the same week, OpenAI also released the highly anticipated latest proprietary model GPT-5, which suffered a reputation disaster, forcing the company to revert ChatGPT's default model to the previous version due to overwhelming criticism.

Such setbacks left Altman expressing "shock." However, The Economist views this as merely proving that OpenAI's supposed "embrace" of openness was insincere, and other US companies might be similarly positioned.

Not to mention some US companies are "turning back the clock."

Just as Altman realized that OpenAI's past closed-model approach was "on the wrong side of history," Meta, which had been widely praised in the open-source community for its open AI model Llama, began to perform a reversal.

CEO Mark Zuckerberg announced that the company would be more cautious in its choices regarding open content. He also committed to building so-called "artificial general intelligence," launching a "talent acquisition" mode with compensation packages worth up to hundreds of millions of dollars to poach top researchers from OpenAI, Google, Apple, Anthropic, and other companies, drawing dissatisfaction from peers.

Scale AI founder and CEO Alexandr Wang was also recruited to join the team as head of a new laboratory.

This Chinese-American entrepreneur, who once declared that "American AI cannot lose to China," immediately began planning to abandon Meta's most powerful open-source AI model Behemoth upon taking office, instead developing a new closed-source model.

US scholar Ethan Mollick, who saw this news in the media at the time, was deeply moved. This AI expert, who has been deeply involved in open-source for over twenty years and began working with large language models early on, is considered by US media as "the go-to AI expert for American policymakers and business leaders."

Even Meta's shift toward closed-source led him to lament, "With this, the US has essentially exited the competition for frontier open-source large language models. Europe still has one competitor, but the rest of the market is almost China's domain."

Ali Farhadi from the Allen Institute for AI also pointed out to The Economist that Chinese companies are going all out, publicly releasing their highest-quality models, while US companies are keen on keeping their "shiny new things" tightly held as proprietary assets.

He stated bluntly, "As difficult as it is to accept, I do believe we are now behind (in the open-source field)."

From a business perspective, the results of different path choices between Chinese and US companies are vastly different. US proprietary models generate far more revenue than Chinese open-source models. Proprietary models are undoubtedly easier to monetize, and the profits can be reinvested in innovation and research, which is an obvious advantage.

However, compared to proprietary technology, open-source models can drive entirely different application scenarios. Percy Liang, co-founder of open-source weight model platform "Together AI," explains that enterprises, governments, and researchers can more easily adapt open-source models to various "niche scenarios." These models can also help users run AI tools locally without relying on cloud services. Additionally, developers can still achieve profitability by providing customized support and other auxiliary services.

In other words, while US laboratories are heavily betting on earning huge profits by pushing the frontiers of intelligence, their Chinese competitors are more focused on encouraging the widespread adoption of artificial intelligence.

This sounds an alarm for The Economist: "If they (Chinese manufacturers) succeed, then the impact brought by DeepSeek may be just the beginning."