The battle of large models has reached a point where victory no longer hinges on parameter tables—it lies in the joints of robots and on the wheels of autonomous vehicles.

01 Two Thought-Provoking Signals

In November 2026, two telling pieces of news surfaced in AI public discourse.

One came from Shanghai. MiniMax founder Yan Junjie admitted in an English-language podcast interview that China's best models have reached roughly 95% of the performance of top US models at one-tenth the cost—but "the last 5% gap may prove far harder to close than the first 95%." He summed up China-US AI as two distinct models: "The US AI ecosystem is building a closed system backed by massive capital, while China's AI ecosystem is building an efficiency-driven collaborative system."

The other came from Washington. Just after the midterm elections, members of Congress publicly called for a pause on advanced AI research and even "a ban on superintelligence." Senator Sanders had earlier proposed the AI Data Center Moratorium Act, requiring a nationwide halt to AI data center construction until Congress addresses safety and employment legislation. Even a newly elected senator declared that the US should fully pause frontier AI research and "not worry about China catching up."

Someone in the opposing camp wants to hit their own brakes—is this an opportunity for China's AI, or a challenge wrapped in sugar?

My judgment: it is an opportunity, but only for a few; for most Chinese AI companies, it is instead the whistle for a more brutal elimination round.

02 Three Technology Routes, Three Survival Philosophies

To see the next one to three years clearly, first look at what the three companies that have already broken out—DeepSeek, Kimi, and MiniMax—are each betting on.

DeepSeek is betting on "architecture is justice." Constrained by limited computing power, it used the MLA attention mechanism and an extreme MoE architecture to drive training costs down to a fraction of competitors', then fully open-sourced the model weights. The essence of this route: I don't compete with you on who has more GPUs; I compete on who squeezes every GPU's computing power more thoroughly. Open source is not sentiment; it is strategic depth—when developers worldwide build on your model, the ecosystem itself becomes the moat.

Kimi is betting on "the model is the product." Moonshot AI rose to fame with long context, experienced the frenzy of burning money on consumer traffic acquisition, and also decisively cut back on spending, refocusing on foundation model capabilities and Agent scenarios. This route is the most painful and closest to the endgame: model companies must ultimately charge for model capabilities themselves, not buy users with subsidies.

MiniMax is betting on "efficiency is business." Yan Junjie explained the company name plainly: Mini means pushing inference costs to the extreme low; Max means pushing the upper limit of intelligence to the highest. This logic has already generated real money—revenue of $116.6 million in the first half of 2026, up 283% year-on-year, with open platform and enterprise services contributing more than 60% of revenue.

The three routes seem different, but share the same underlying consensus: in China, where computing power is constrained, efficiency is not a multiple-choice question; it is a matter of life and death. A Stanford report confirms this—in 2025, US private AI investment was 23 times China's, yet the performance gap between top models was only about 2.7%. China's engineer dividend has forcibly converted the capital gap into an engineering gap, and then compressed the engineering gap to single-digit percentage points.

03 Industry View at the Meso Level: The Pause Argument Is a Performance, but It Exposes Anxiety

From a meso-level perspective, US politicians' "pause AI" proposals are unlikely to materialize—Sanders' bill still has zero co-sponsors and sits in committee. But the signal it sends is real: America's AI frenzy has begun to face a triple backlash from electricity, employment, and public opinion, with data centers voted against by residents in multiple states.

The window this gives Chinese companies is not that "opponents have stopped," but that opponents' internal friction costs are rising. In the next one to three years, three things will happen at the meso level: first, inference costs will drop by another order of magnitude, and AI will move from a "money-burning model race" to a "cash-flow accounting race"; second, model companies without self-sustaining capacity will be acquired or cleared out, and the industry will converge from a hundred flowers to a few leaders; third, Agents will truly be deployed, token consumption will explode, and platform companies selling "intelligent utilities" will capture most of the profits.

But a bucket of cold water is also needed. MiniMax peaked on its listing day, with its stock price falling from HK$1,330 to HK$156 and market value evaporating nearly 90%—the capital market has already spoken with real money: the story of narrowing the technology gap is over; from now on, only revenue and profit count. At the micro level, every company's core proposition has become the same: turn model capabilities into a renewable business and turn efficiency advantages into pricing power.

04 The Real Endgame: Step Out of the Screen and Into the Physical World

Finally, one judgment obscured by the large-model narrative must be pointed out: large language models are only the prologue; embodied intelligence—robots and autonomous driving—is the main battlefield for China's AI, and also China's true future.

The logic is clear. Competition in language models is essentially competition in data and computing power, where China is naturally in a catch-up position; but competition in embodied intelligence is essentially competition in "intelligence + manufacturing"—servo motors, reducers, LiDAR, batteries, and the supply chains for complete vehicles and machines. The world's most complete chain is in China. However strong Tesla's FSD is, it still needs the Shanghai factory to land; however beautiful American humanoid robot blueprints are, the harmonic reducers in their joints are most likely still made in China.

More importantly, the data closed loop: the physical-world data generated by robots every day in Chinese factories, streets, and homes is a scarce asset that no language model crawling the entire web's text can replace. Autonomous driving is naturally the first trillion-level scenario of embodied intelligence, and the deployment density of Robotaxi in China has already surpassed that in the US.

For DeepSeek, Kimi, and MiniMax, this means they must answer a question in the next one to three years: can your model step out of the chat box, be installed in a robot's body, and drive on real roads? Whether large-model companies become "brain sellers" or evolve into "brain + body" system companies will determine whether they are platforms or suppliers in the next decade.

05 Conclusion

US politicians wanting to press the pause button precisely shows they feel the breath of pursuers behind them. But history never rewards those who wait for opponents to make mistakes; it rewards those who turn the window of opportunity into positional advantage. The real opportunity for Chinese AI companies lies not in others' hesitation, but in their own efficiency, their own supply chains, and their own physical world.

The battle of large models has reached a point where victory no longer hinges on parameter tables—it lies in the joints of robots and on the wheels of autonomous vehicles.