China's push into artificial intelligence is entering a new phase, with leading developers seeking not just to compete globally but to reshape how the technology is built, shared and monetized, a prominent AI firm founder said in an interview with China Daily during the ongoing Zhongguancun Forum on Wednesday.

Yang Zhilin, the founder of Moonshot AI, an artificial intelligence company based in Beijing, also disclosed that Moonshot AI is exploring new ways to overhaul the core architecture of large models and suggest more "possibilities" in response to Elon Musk's praise of its attention residuals.

Yang said that Chinese firms are increasingly positioning themselves as drivers of structural change in the AI ecosystem — a shift he described as both an opportunity and a defining strategy for China.

"China's willingness to openly share models and technical breakthroughs could accelerate global innovation while giving it a distinct edge over more closed ecosystems."

Yang argued that as large models approach parity in performance, competitive advantage will shift from algorithms to infrastructure — particularly the ability to generate and process vast volumes of "tokens", the basic units of AI computation.

"In the long run, the bottleneck may no longer be model capability, but how quickly you can build large-scale 'token factories'," he said, pointing to energy costs and computing infrastructure as decisive factors.

Yang predicted that if open-source and proprietary models reach similar levels of capability, open systems could ultimately dominate by unlocking broader ecosystems of developers, applications and distribution channels.

Closed models currently retain a share of the market, he said, but open-source platforms may generate greater overall "token output" — and therefore greater economic value — by enabling more participants to build on top of them.

In one of his more striking assertions, Yang suggested that AI-generated tokens could become a proxy for economic activity, potentially redefining traditional measures such as GDP.

"As productivity increasingly comes from AI agents generating tokens, those tokens could, in effect, become equivalent to GDP," he said, adding that the technology could multiply economic output by several times — or even orders of magnitude — over the long term

Earlier this month, Elon Musk commented "impressive work from Kimi" on X, highlighting a major breakthrough by Moonshot AI, Kimi's developer, on "attention residuals", a novel improvement to the Transformer architecture that enhances training efficiency and model performance at scale.

In response, Yang said the new approach introduces a different way of organizing computation models from Transformer architecture, the dominant framework behind today's large language models.

The idea, he said, is to rethink how operations are distributed across a model's internal dimensions.

"Structures that can be applied along the time dimension can also be applied along the depth dimension," Yang said, describing the conceptual shift behind "attention residuals".

By reconfiguring how information flows through layers, the approach has delivered measurable improvements in training efficiency and model performance, according to Yang, though he did not disclose detailed benchmarks.

Despite the widespread adoption of Transformers and standard optimization techniques over the past decade, Yang argued that the field is still open to disruption.

"Standards have formed, but within those standards there is still a lot that can be overturned. I think there are still many possibilities," he said.

Yang also pointed to China's education system as a foundational strength, citing decades of investment from primary schooling to doctoral training that have produced a large pool of technically skilled workers.

The scale of that pipeline, he said, allows companies to access high-level talent at relatively competitive cost, creating a feedback loop in which human expertise accelerates the development of machine intelligence.