Hundreds of Chinese AI labs are undercutting each other on price and poaching each other’s researchers in a domestic fight that produces engineering output but almost no profit, according to venture investor David Cheng, who spent a week meeting model labs, robotics companies, and academics in Beijing. Cheng, writing on his Substack newsletter Earned Intuition, frames the dynamic through a Chinese term, 内卷 (involution): effort that keeps escalating without anyone actually getting ahead.
The stakes are commercial, not just competitive. Cheng argues the only move that converts that grinding into anything of value is 出海, going overseas, and he quotes a line he heard repeatedly on the trip: “if you don’t go overseas, all that grinding was for nothing.” That framing, he says, explains behavior American observers tend to read as aggression: open-weight releases, collapsing prices, and what he calls a “desperate courtship” of American developers.
Cheng’s own estimate, gathered from conversations on the trip rather than published benchmarks, puts US effective compute at more than ten times all Chinese labs combined. He reports a senior researcher at a major Chinese lab telling him their post-training gains exist specifically “because we don’t have access to high performing chips,” and cites Moonshot gating access to its Kimi K3 model over serving capacity as evidence the constraint is real, not strategic.
That scarcity, in Cheng’s account, also explains China’s open-source strategy. He quotes an unnamed university researcher’s formulation: “When you trail the frontier, openness maximizes reputation per unit of capability. The moment you lead, you close.” Cheng notes ByteDance keeps its Doubao super app, which he says has more than 300 million users, closed precisely because it already owns distribution, while Alibaba keeps its largest Qwen models closed and API-only.
The revenue picture backs the “export or bust” thesis with real numbers. Cheng cites MiniMax’s first-half 2026 filing showing international revenue above 60 percent of a $116.6 million total, with enterprise and platform sales making up 63.4 percent of that mix. Qwen, by his account, has more than a billion downloads overseas but earns nothing directly from most of them, since its models ship under an open Apache license; Alibaba instead monetizes through cloud pull-through.
Cheng’s most pointed argument concerns why Chinese models reportedly lag on agentic tasks: several researchers he spoke with put current models at 10 to 20 percent completion rates on long-horizon, multi-step work. His preferred explanation is not talent or compute but the absence of an interoperable domestic tech ecosystem for agents to train against, since Chinese companies rarely open APIs to each other for fear of being “eaten” by a rival.
He also cites the reversal of Meta’s roughly $2 billion acquisition of the agent startup Manus. In April, on national security grounds, China’s National Development and Reform Commission forced that deal apart, and Manus returned to operating independently by August. Cheng reads the episode as a warning shot: Beijing wants the standing option to claw back a transaction years later, however far along it is, whenever Chinese money sat anywhere in its history.
None of this is independently verified reporting; it is one investor’s synthesis of conversations conducted on a single trip, and Cheng is explicit that his own estimates and quoted sources are unnamed beyond title and lab. The piece is an argument, not a data release, and its numbers on filings and downloads are the only load-bearing facts that come from outside his own notebook.
For operators evaluating Chinese open-weight models, the practical read is in the incentive structure Cheng lays out: labs currently giving away frontier-adjacent weights are doing so because they trail, and his sourcing suggests that calculus flips the moment compute constraints ease. Anyone building a product roadmap on continued Chinese open-weight releases should treat that access as a temporary condition of the current price war, not a permanent feature of the market.
Based on an essay by David Cheng published on Earned Intuition on September 18, 2026.
