China’s independent AI model firms are finding that strong demand does not guarantee healthy economics: scarce, costly computing power is squeezing the business from several directions. A new industry analysis describes four different responses, from building infrastructure to moving AI onto phones, and shows why the cost of running models now matters as much as the models themselves.
Watch Desk analysis
What happened
The analysis focuses on Moonshot AI, Zhipu, MiniMax and StepFun. It argues that each is trying a different way to cope with the cost of compute: Moonshot is raising capital and securing capacity; Zhipu is investing in its own computing infrastructure; MiniMax has shifted towards enterprise services; and StepFun is betting on phones and on-device AI.
The figures in the account illustrate the trade-offs. It says MiniMax’s first-half gross margin fell to 17.9%, while its enterprise and open-platform revenue grew sharply. It reports that Zhipu’s inference costs per token fell by 80% from the start of the year as it scaled domestic chips, but that the company also recorded a large adjusted net loss. For StepFun, the article describes its phone strategy as an attempt to rely less on cloud inference, while noting that sales have yet to prove the case. BigGo Finance says its data is compiled from public reporting and company disclosures.
Why it matters
AI’s economics do not stop at training a model or attracting users. Every conversation and agent task needs computing capacity, and the bill can rise alongside usage. A strategy that wins customers may still struggle if inference costs eat into what those customers pay.
That makes the companies’ choices more than a contest of product launches. Owning infrastructure may reduce unit costs but demands heavy upfront investment. Enterprise services can bring different revenue, while on-device AI could shift some work away from the cloud. Each path has its own bill, and none is a shortcut around proving that customers will pay enough.
Our read
The useful lens here is not which company has the most dramatic growth figure, but whether it can turn that growth into durable margins. The analysis offers a clear comparison of four strategies, though its detailed figures should be read as reported claims, not a like-for-like audited scorecard. Compute is the industry’s rent; the question is who can afford the lease without passing the whole bill to customers.
What to watch
- Whether Moonshot’s fundraising and capacity expansion translate into better margins as usage grows.
- Whether Zhipu can earn back the capital tied up in its infrastructure.
- Whether MiniMax’s enterprise growth improves margins rather than simply bringing more inference costs.
- Whether StepFun’s devices find enough buyers to make on-device AI a viable alternative to cloud use.
Discussion spark: Which strategy looks most likely to make an independent AI company sustainable: owning compute, focusing on enterprise sales, or shifting more AI onto devices?
Sources and evidence
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.