Two years ago, the United States was assumed to hold a commanding, multi-year lead in frontier AI. That assumption has collapsed. The 2026 Stanford AI Index put the top US model only 2.7 percent ahead of China’s best in March, and by mid-year analysts were measuring the gap in months rather than generations. The reason is a strategy rather than a single breakthrough, and the strategy is open weights. While the leading American labs keep their strongest models closed and metered, China’s biggest players, including Alibaba’s Qwen, Moonshot’s Kimi, DeepSeek, Z.ai and MiniMax, release downloadable weights under permissive licences and let the world build on them for nothing. Moonshot’s Kimi K3 briefly became the largest open model in the world at 2.8 trillion parameters, and Alibaba published the weights of its 2.4-trillion-parameter Qwen 3.8-Max. Hugging Face now reports that Chinese models have overtaken American ones in both monthly and all-time downloads.
New York University professor Scott Galloway has a blunter word for it. “I think China is beginning to engage in what I’ll call AI dumping,” he argues, comparing it to flooding a foreign market with below-cost goods until the domestic industry collapses, the way cheap imports once hollowed out Detroit. China does not need to build a better model than OpenAI, on this view. It only needs to make the American one uneconomic. That is where the United States looks exposed, because so much of its market now rests on so few companies. SpaceX, OpenAI and Anthropic carry a combined valuation of roughly $5.2 trillion, more than the first-day value of all 3,365 US technology companies that went public between 1980 and 2025 put together, which came to about $4.1 trillion in data compiled by University of Florida professor Jay Ritter. If free Chinese models undercut the business case for paid American ones, the exposure runs well beyond a handful of firms to a large slice of US market value tied to them.
Dumping, not charity
Galloway’s point is that giving the models away is the business model, not the absence of one. On The Prof G Pod he has called Chinese open-weight AI “the biggest story in business that very few people are talking about,” noting that these models now power a majority of the AI tokens consumed inside the United States at roughly 70 percent below the cost of American equivalents. Dumping is the economist’s term for selling below cost to capture a market and drive competitors out, and it does not require the dumped product to be the best on the shelf. It only has to be good enough and dramatically cheaper. Plenty of market leaders have been killed by exactly that combination.
For a Chinese state that treats AI as strategic infrastructure, there is a logic to it that has nothing to do with quarterly profit. Cheap, ubiquitous Chinese models become the default layer the rest of the world builds on, which spreads Chinese technical standards, seeds dependence, and denies American labs the high-margin licensing revenue their valuations assume. The models themselves need not turn a profit for the strategy to pay off.
Why the US is unusually exposed
The concentration is the vulnerability. A combined $5.2 trillion across three companies against $4.1 trillion for four and a half decades of tech listings is a startling ratio, though it comes with a caveat worth stating plainly: SpaceX’s roughly $2 trillion is a live private-market valuation, Anthropic’s $2 trillion is an IPO target, and OpenAI’s $1.2 trillion is a proposed private round. These are expectations as much as prices. The businesses underneath them are also cash-hungry. OpenAI has projected around $278 billion of negative free cash flow between 2026 and 2030, and Anthropic is not expected to reach positive cash flow until 2028. Those plans assume years of paid demand at healthy prices. A flood of free, capable alternatives is precisely the thing that could break that assumption, and because so much market value is stacked on so few names, the damage would not stay contained to them.
The case against the panic
The counterargument deserves a fair hearing. American labs still hold the top of the frontier on the hardest benchmarks, backed by a compute advantage China cannot yet match, and the most demanding enterprise work continues to flow to closed models with support, guarantees and liability behind them. Open weights are cheap to run but expensive to fine-tune, host and secure at scale, which favours firms that can pay for expertise. Many Western buyers also remain wary of routing sensitive data through Chinese-origin models on trust and compliance grounds alone. Cheaper has toppled better many times in business history, but not every time, and the US retains real structural advantages. The honest position is that China has turned a disadvantage in raw compute into a genuinely threatening distribution strategy, and whether that is enough to win is still an open question rather than a settled outcome.
What is no longer arguable is the shape of the contest. The race is not only about who trains the most capable model. It is also about who sets the default, who captures the developers, and whose economics survive contact with a competitor willing to charge nothing. On that scoreboard, China is doing considerably better than the headline benchmark numbers suggest.