Realistic Constraints of Computing Power Land Rent and the Four Little Dragons

Kimi is scouring for GPUs and computing power, Zhipu is facing practical constraints in large-parameter training, MiniMax is concentrating phased resources on M3 Pro, while Stepfun is positioning the mobile end as a direction that can be scaled up at any time. The current situation of China's "Four Little Dragons" of artificial intelligence is often summed up as "a total mess". This statement carries emotional overtones, but it also points to a fact: all four companies are showing signs of strain at the same time. This tension is not an accidental operational mistake, but a different shape squeezed out by the same structural force on different strategic paths: computing power, in the form of a "new type of land rent", has redrawn the cost curve and survival threshold for large model competition.

Only by understanding this point can we understand why these difficulties are emerging in a concentrated manner.

Computing Power: The Absolute Land Rent in the Large Model Era

The "land rent" in classical political economics refers to the part of surplus that landowners extract from producers by virtue of their monopoly on limited natural factors. In the large model industry in 2026, high-end computing power is playing a similar role.

Kimi's case is the most intuitive. In July 2026, less than 48 hours after the launch of the K3 model, the global request volume fully saturated the GPU cluster, forcing the company to suspend new user subscriptions. More critically, the unit economic model shows that if a C-end user conducts two in-depth long-text conversations per day, the GPU computing power cost consumed may exceed the monthly subscription fee of 59 yuan. Kimi's ARR exceeded 300 million US dollars in June 2026, with API revenue accounting for more than 70%. Its commercialization data is already remarkable among independent model companies, but the computing power cost is compressing the gross profit margin brought by growth at an almost equal speed. Moonshot AI completed intensive financing from Series C to Series F in 2026, with the Series F financing amount reaching 3.5 billion US dollars and the valuation rising all the way from 4.3 billion US dollars to 35 billion US dollars, but the flow of financing is highly clear: expanding graphics cards, locking production capacity, and training the larger K4. This is not a cash flow crisis of one company, but an inevitable tension of a business model under the rigid rise of computing power costs — the more users there are, the heavier the inference load, and the more rigid the "rent" paid to the computing power supply chain.

Zhipu's choice is another response. Instead of trying to race against costs on the track of "renting NVIDIA GPUs", it has placed its bets on the direction of "becoming a landowner": going all-in on Infra, building self-owned computing power infrastructure, and integrating domestic chips into the main inference computing power, realizing large-scale inference of hundreds of thousands of domestic chips, with the unit token inference cost dropping by 80% compared with the beginning of the year. In the first half of 2026, MaaS revenue reached 825 million yuan, a year-on-year increase of more than 27 times, proving that this path is not a castle in the air commercially. But the cost of being a "landowner" is equally heavy: the adjusted net loss in the first half of the year was about 1.964 billion yuan, and R&D expenditure was 2.131 billion yuan. The construction of 10,000-card or even 100,000-card clusters requires locking high-end production capacity and network equipment with a large proportion of advance payment. Zhipu's "stuck" situation does not lie in that it does not know what to do, but in that the path of "building a self-owned computing power base" itself requires it to complete a round of highly capital-intensive fixed asset investment before making profits, and the payback period of this investment is longer than the time window that the capital market is usually willing to wait for. This is a mismatch between path selection and time window, rather than a reckless direction itself.

Four "Solutions", the Same Squeeze

The divergence of the Four Little Dragons in 2026 essentially means that when facing the same "computing power land rent", they have chosen four different payment methods.

Kimi pays the "rent premium" — through round after round of financing, it purchases scarce GPU computing power from the market and cooperates with cloud vendors, exchanging capital for time, and time for model capabilities. The risk is that the rent is rigid and floating, while the leading window of model capability is perishable. The performance of K3 proves the effectiveness of this path, but the paradox that "the stronger the model, the more money it burns" means that every technological leap is accompanied by the rigid demand for the next round of financing, so IPO has changed from an option to a phased necessity.

Zhipu pays the "infrastructure sunk cost" — using huge capital expenditure in exchange for long-term unit cost reduction and supply chain autonomy. If the ecological migration cost of domestic chips can be diluted by scale within two to three years, the final outcome of this path will be more stable; but if the code rewriting and operator optimization from CUDA to the domestic ecology continue to consume engineering resources, then "self-construction" will form long-term pressure on cash flow. What Zhipu is betting on is time: whether the verification speed of the economy of domestic computing power can outrun the depletion speed of capital patience.

MiniMax pays the "cost of valuation track switching". It has done something uncommon among Chinese large model companies: in 2026, it switched its revenue engine from C-end products with a gross profit margin of only 4.7% to B-end token calls with a gross profit margin of 69.4%. Its ARR has been pushed from less than 200 million US dollars at the beginning of the year to more than 800 million US dollars, with enterprise customers exceeding 2 million. This is a clear improvement of the unit economic model. But the feedback from the capital market is very direct: its market value once stood at 410 billion Hong Kong dollars, and its stock price retreated by nearly 88% from the high point, falling below the issue price. The reason is not that MiniMax did something wrong, but that the story of a "platform AI company" requires it to prove two mutually restrictive things at the same time: revenue can expand with token consumption, and gross profit will not be eroded in reverse by inference costs. The gross profit margin in the first half of the year was 17.9%, down 7.5 percentage points from 25.4% for the whole year of 2025 — Agents have pushed up token consumption, and also pushed up GPU bills proportionally. The pressure does not lie in technical capabilities, but in whether the market is willing to give it enough time to let scale effects outrun the cost curve.

Stepfun pays the "concentrated bet on path switching". When the other three are entangled in the cloud model layer, Stepfun has chosen a physically more "heavy" path: making mobile phones. Step AOS, personal agent Amoo, and STEPX Neo are fully self-developed from the operating system to the terminal hardware, trying to use the end-side entrance to bypass the cost structure and distribution dilemma of pure cloud inference. This is a logically self-consistent strategy: if the ultimate form of the agent requires cross-application scheduling and system-level permissions, then model companies at the pure App layer will always be subject to the fences of operating systems and super Apps. But the risk is equally clear: according to industry reports, the revenue in 2025 was less than 500 million yuan, the expected revenue in 2026 is about 1.2 billion yuan, and the cumulative financing exceeds 20 billion yuan. However, the cost-effectiveness of the conventional configuration of its first mobile phone is not prominent in the fiercely competitive mobile phone market, and whether the sales volume can support the closed loop of "terminal defining model" remains an unsolved problem. Stepfun's "always ready to fully march into the mobile end" is not so much an offensive gesture as an attempt to open up an enclave on the terminal side that is not governed by the same set of cost rules, under the circumstance that the "land rent" at the cloud model layer is too expensive.

The Narrative of the "Four Little Dragons" Needs to Be Reassessed

Deeper changes are taking place in the meaning of the label "Four Little Dragons" itself.

This title is borrowed from the "SenseTime, Megvii, CloudWalk, YITU" in the CV era, which itself implies an assumption: the large model track will, like computer vision, accommodate several independent startups of comparable size competing in parallel. But the industrial reality in 2026 is falsifying this assumption. DeepSeek has proved with extreme engineering efficiency (2048 H800 GPUs completed V3 pre-training) that model capabilities are not necessarily linearly correlated with computing power stacking; large factory models such as Tencent Hunyuan and ByteDance Doubao are in the first echelon in terms of total token usage; according to the token usage caliber of the four leading models, two models from DeepSeek, one from Tencent Hunyuan and one from Zhipu together account for 66% of China's large model token usage. What independent model companies are facing is a network woven together by the computing power reserves of large manufacturers, cloud ecosystem binding and price war capabilities. Zhipu's listing in January 2026 repriced the entire track, MiniMax followed up with the fastest IPO record immediately, and Moonshot AI's IPO schedule was thus rewritten. Listing is no longer a symbol of "success", but a phased means to obtain continuous financing capacity in the computing power arms race.

At the same time, the strategic focus adjustment of large manufacturers on independent AI products is also obvious. ByteDance's Coze Enterprise Edition is counting down to offline, and the whole team is merged into Doubao; according to incomplete statistics, in 2026, large manufacturers and startups have collectively removed or merged more than 25 independent AI products. This is not an isolated product adjustment, but a clear signal: when computing power cost becomes the core variable of competition, the chain of encapsulating model capabilities into independent products and then recovering computing power investment through independent products is being replaced by the vertical integration logic of "model as infrastructure, entrance as super application".

Conclusion

"Total mess" is real, but it does not describe the incompetence of the four companies, but the structural mismatch of an industrial stage. Computing power, in the form of land rent, has pushed the fixed cost of large model competition to a level that can only be afforded by continuous large-scale financing or deep own cash flow. The Four Little Dragons respond to this constraint in four completely different ways: Kimi exchanges capital for computing power, Zhipu exchanges capital for infrastructure, MiniMax exchanges revenue structure for gross profit margin space, and Stepfun exchanges terminal entrance for a brand new cost structure. Every path has its inherent logic, and every path also has its inherent vulnerability.

The biggest misguidance of the "Four Little Dragons" analogy is that it makes people think that this is still a race between four startups. In fact, each of them is wrestling with the same structural opponent: in a market where computing power supply is subject to geopolitics and the capital return cycle is continuously extended, can independent model companies build a survival closed loop that does not rely on continuous external blood transfusion before "land rent" exhausts patience. The answer to this question will not be revealed at the press conference of any of the four companies. It will be written silently between the cash flow statements and the opening and closing of IPO windows in the next 12 to 18 months. For the four companies, the real test is not who is better at telling stories, but who can follow through on their own choices under constraints.

This article is from the WeChat official account "Tech Business", authors: Zisen, Dingyi, authorized to be released by 36Kr.