Compute Rent and the Real Constraints of the Four Little Dragons
Using compute rent as its framework, the article analyzes the four response paths chosen by Kimi, Zhipu, MiniMax, and StepFun in 2026 when facing the same cost constraint. Kimi relies on financing to purchase GPU compute, Zhipu builds its own compute infrastructure and uses domestic chips at scale, MiniMax shifts its revenue focus from the consumer side to B-side token calls, and StepFun bets on mobile phone terminals. The author argues that the narrative of the Four Little Dragons is borrowed from the CV era and can no longer describe the current competitive landscape between big tech companies and startups.
Why read this
AI score 63
The article uses the framework of compute rent to place the different predicaments of the four companies under the same cost constraint, making it easier to compare the costs of each path.
Topics
Body · AI translation
Kimi is searching everywhere for GPUs and compute, Zhipu has run into real-world constraints on large-parameter training, MiniMax has concentrated its resources at this stage on M3 Pro, and StepFun has made mobile its direction that it can double down on at any time. The current predicament of China's "four AI tigers" is often summed up as "a complete mess." That phrasing carries emotion, but it also points to a fact: all four companies are simultaneously showing signs of strain. This strain is not an accidental operational misstep, but the different shapes squeezed out of different strategic paths by the same structural force—compute, in the form of a "new type of land rent," has redrawn the cost curve and the survival threshold for large-model competition.
Only by understanding this can one understand why these difficulties have emerged in concentrated fashion.
Compute: absolute rent in the large-model era
In classical political economy, "land rent" refers to the portion of surplus that landowners, by virtue of their monopoly over limited natural factors, extract from producers. In the large-model industry of 2026, high-end compute is playing a similar role.
Kimi's case is the most intuitive. In July 2026, less than 48 hours after the K3 model launched, global request volume maxed out its GPU clusters, forcing the company to suspend new user subscriptions. More critical is the unit economics: a consumer user conducting two deep long-text conversations per day could consume GPU compute costs exceeding the 59-yuan monthly subscription fee. Kimi's ARR surpassed $300 million in June 2026, with API revenue accounting for more than 70%, commercialization figures that are already impressive among independent model companies, but compute costs are compressing the gross margin space brought by growth at almost the same speed. Moonshot completed intensive financing from Series C through Series F in 2026, with the F round reaching $3.5 billion and valuation climbing all the way from $4.3 billion to $35 billion, but the financing flowed in a highly clear direction: expanding graphics cards, locking in capacity, and training a larger K4. This is not a cash-flow crisis for one company, but the inevitable strain of a business model under rigidly rising compute costs—the more users, the heavier the inference load, and the more rigid the "rent" paid to the compute supply chain.
Zhipu's choice is another kind of response. It did not try to race against costs on the track of "renting Nvidia," but instead staked its chips on the direction of "becoming a landlord": All-in Infra, building its own compute infrastructure, and incorporating domestic chips into its main inference compute, achieving large-scale inference on domestic chips at the hundred-thousand level, with unit token inference cost down 80% from the start of the year. In the first half of 2026, MaaS revenue was 825 million yuan, up more than 27-fold year-on-year, proving that this path is not a castle in the air commercially. But the price of being a "landlord" is equally heavy: adjusted net loss in the first half was about 1.964 billion yuan, R&D expenses were 2.131 billion yuan, and building clusters of ten thousand or even a hundred thousand cards requires locking in high-end capacity and network equipment with large proportional prepayments. Zhipu's "stuckness" is not that it does not know what to do, but that the path of "building its own compute foundation" itself requires it to complete a round of extremely capital-intensive fixed-asset investment before profitability, and the payback period of this investment is longer than the time window the capital market is usually willing to wait. This is a mismatch between path choice and time window, not recklessness about the direction itself.
Four "solutions," one kind of squeeze
The divergence among the four tigers in 2026 is essentially four different payment methods chosen when facing the same "compute land rent."
Kimi pays the "rent premium"—through round after round of financing, it buys scarce GPU compute from the market and cooperates with cloud vendors, trading capital for time and time for model capability. The risk is that rent is rigid and floating, while the leading window for model capability is fleeting. K3's performance proved 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 financing round, and an IPO therefore shifts from an option to a stage-necessary step.
Zhipu pays the "infrastructure sunk cost"—using massive capital expenditure in exchange for a long-term decline in unit costs and supply-chain autonomy. If the ecosystem migration cost of domestic chips can be diluted by scale within two to three years, the endgame of this path will be more solid; but if code rewriting and operator optimization from CUDA to the domestic ecosystem continue to consume engineering resources, then "self-building" will create long-term pressure on cash flow. What Zhipu is betting on is time: whether the verification speed of the economics of domestic compute can outrun the exhaustion speed of capital patience.
What MiniMax paid was "the price of a valuation track switch." It did something uncommon among Chinese large-model companies: in 2026 it switched its revenue engine from C-end products with a gross margin of only 4.7% to B-end token calls with a gross margin of 69.4%. ARR was pushed from less than $200 million at the start of the year to more than $800 million, and enterprise customers surpassed 2 million. This was a clear improvement in the unit economics model. But the feedback from the capital market was direct: its market capitalization once stood at HK$410 billion, and the stock price has retraced nearly 88% from its high, falling below the issue price. The reason is not that MiniMax did anything wrong, but that the story of a "platform-type AI company" requires it to prove two things that pull against each other at the same time: revenue can expand with token consumption, while gross margin will not be devoured in reverse by inference costs. First-half gross margin was 17.9%, actually down 7.5 percentage points from 25.4% for all of 2025—agents pushed up token consumption and also pushed up GPU bills proportionally. The pressure is not about technical capability, but about whether the market is willing to give it enough time for scale effects to outrun the cost curve.
What StepFun paid was "a concentrated bet on a path switch." While the other three were battling it out at the cloud model layer, StepFun chose a physically "heavier" path: making phones. Step AOS, the personal agent Amoo, and STEPX Neo, with full-stack in-house development from operating system to terminal hardware, attempt to use the on-device entrance to bypass the cost structure and distribution difficulties of pure cloud inference. This is a logically self-consistent strategy: if the ultimate form of agents requires cross-app scheduling and system-level permissions, then model companies at the pure App layer will always be constrained by the walled gardens of operating systems and super apps. But the risks are equally clear: according to industry reporting, 2025 revenue was less than RMB 500 million, with about RMB 1.2 billion expected in 2026, cumulative financing exceeding RMB 20 billion, while the regular-configuration cost-effectiveness of its first phone is not outstanding in a highly competitive phone market, and whether sales can support the closed loop of "terminals defining models" remains an unresolved question. StepFun's "always ready to go all out into mobile" is less an offensive posture than an attempt, under the condition that the "land rent" at the cloud model layer is too expensive, to open up an enclave on the terminal side that is not governed by the same set of cost rules.
The "Four Little Dragons" narrative needs to be re-evaluated
The deeper change occurs in the meaning of the label "Four Little Dragons" itself.
This term is borrowed from the CV era's "SenseTime, Megvii, CloudWalk, Yitu," and itself implies an assumption: that the large-model track will, like computer vision, accommodate several independent, comparable-sized startups competing in parallel. But the industrial reality of 2026 is falsifying this assumption. DeepSeek proved with extreme engineering efficiency (2,048 H800s completing V3 pretraining) that model capability is not necessarily linearly related to compute stacking; major-company models such as Tencent Hunyuan and ByteDance Doubao are in the first tier in overall tokens usage; if measured by the token usage of the four leading models, two 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 face is a net woven together by major companies' compute reserves, cloud ecosystem lock-in, and price-war capability. Zhipu's bell-ringing in January 2026 repriced the entire track, MiniMax immediately followed with the fastest IPO record, and Moonshot AI's IPO timetable was therefore rewritten. Going public is no longer a sign of "success," but a stage-necessary means of obtaining sustained financing capability in the compute arms race.
At the same time, the shift in major companies' strategic focus on independent AI products is also obvious. ByteDance's Coze Enterprise Edition has entered a countdown to shutdown, with the team merged in full into Doubao; according to incomplete statistics, in 2026 major companies and startups have together retired or merged more than 25 independent AI products. This is not an isolated product adjustment, but a clear signal: when compute cost becomes the core variable of competition, the chain of packaging model capabilities into independent products and then recouping compute investment through independent products is being replaced by the vertical integration logic of "model as infrastructure, entrance as super app."
Conclusion
"A mess" is real, but what it describes is not the incompetence of the four companies, but a structural mismatch at a stage of the industry. Compute, in the form of "land rent," has pushed the fixed costs of large-model competition up to a level that only sustained large-scale financing or deep proprietary cash flow can bear. The Four Little Dragons respond to this constraint in four completely different ways: Kimi uses capital to exchange for compute, Zhipu uses capital to exchange for infrastructure, MiniMax uses revenue structure to exchange for gross-margin space, and StepFun uses the terminal entrance to exchange for a new line in the cost structure. Each path has its internal logic, and each path also has its internal fragility.
The biggest misleading aspect of the "Four Little Dragons" analogy is that it makes people think this is still a race among four startups. In fact, each of them is wrestling with the same structural opponent: in a market where compute supply is constrained by geopolitics and the capital return cycle is being continuously lengthened, can independent model companies establish a survival closed loop that does not depend on continuous external blood transfusions before the "land rent" exhausts their patience? The answer to this question will not be revealed at any press conference of any of the four companies. It will be written silently in the cash flow statements and in the opening and closing of IPO windows over the next twelve to eighteen months. For the four companies, the real test is not who is better at telling stories, but who can walk their own choice to the end under constraints.
This article comes from the WeChat public account "Tech Business," authors: Zisen and Dingyi, published by 36Kr with authorization.