Price per Million Tokens Drops to Just a Few Yuan: The Rise of the "Computing Power Supermarket" Model
0xBroomberg
Large-model inference now costs just a few yuan per million tokens and is still falling, driven by a shared GPU-scheduling model dubbed the 'compute supermarket' that is reshaping cost structures for smaller AI firms across China.
A few yuan per million tokens — how cheap is that?
Jing Rongwei, deputy GM of Zhejiang Computing Power Technology (浙江算力科技), told CCTV Finance that large-model products priced per million tokens now cost roughly a few yuan — and the trend is still downward.
In plain terms = running a large model used to be a serious line item; now the same volume of calls costs about as much as a cup of coffee.
This means → for small and mid-sized firms, the "can't afford to use a large model" barrier is vanishing fast.
What exactly is the 'compute supermarket' model?
The model pools GPU capacity from multiple sources into a unified scheduling platform; users draw what they need without building their own data centers.
Billing can go as granular as a single GPU card per hour, or settle on a token-based annual or monthly package — pay-as-you-go, like a utility.
This means → companies no longer need a large upfront capital outlay on servers; compute becomes a flexible operating expense.
Who is using it, and what real problems does it solve?
A Hangzhou-based AI film-production company peaks at roughly ¥150,000 per day in compute consumption and demands high stability; the platform lets it procure on demand, avoiding heavy upfront investment and idle capacity.
Industrial-robotics firm Lingxi Robot (灵西机器人) used the same platform to overcome insufficient local VRAM and limits on parallel multi-tasking.
In plain terms = one firm needed massive, stable throughput; the other needed on-demand scalability — the shared scheduling platform handled both.
SMEs make up 80% of users — what does that signal?
The platform currently serves a user base that is 80% small and mid-sized enterprises, spanning education, tech, culture, e-commerce, AI, robotics, and embodied intelligence.
This reflects a rapid diffusion of real large-model demand from big tech players down to the SME tier.
This means → falling token prices combined with shared compute infrastructure are actively reshaping SME cost structures — whether this can drive broader commercial deployment across more use cases is the key question for the next phase.
Content is for reference only, not financial advice.