MagicShape's token factory lands a second round in three months
Chinese AI-infrastructure startup MagicShape (魔形智能) has closed its A-round just three months after its Pre-A, a sign of how fast capital is consolidating around the "token factory" model that powers much of the current open-model boom.
"Token super-factory" MagicShape announced an A-round led by Yida Capital, with several strategic backers joining and existing shareholders adding over-subscribed follow-on investment — barely three months after its multi-hundred-million-yuan Pre-A round in May. The company, founded by CEO Xu Lingjie and CTO Jin Chen, now sells trillions of tokens per day to internet giants and large-model labs, posts revenue in the hundreds of millions of yuan, and says some models are already at breakeven. Its next target is pushing daily token output toward ten trillion.
The story behind the rounds is the industry's quiet pivot from owning models to producing tokens. With DeepSeek, Kimi, GLM and MiniMax opening up weights, anyone can download a model — but as Xu told tech outlet Zhidx, "the model is open source; an efficient way to deploy it is not." Whether tokens actually sell rests on holding down latency, throughput, success rate and cost under big concurrent load. MagicShape's edge is a self-developed inference engine — splitting prefill and decode, KV-cache-aware scheduling, multi-hardware tuning — plus work on "supernode" hardware that interconnects accelerators at high bandwidth, a platform it argues is essential for agent workloads chasing thousands of tokens per second.
The take: the funding validates a thesis that has reshaped the 2026 AI landscape from chips to clouds — that the moat is no longer the model but the economics of running it. Just as US firms race to cut inference cost, Chinese capital is now pouring into companies that can manufacture usable tokens per yuan and per watt, betting that the winner will be the operator, not the intellect. MagicShape's founder, a Biren co-founder and former Alibaba Cloud AI-infrastructure director, has spent two decades in exactly this hardware-software-vein, which helps explain why the two funds landed so close together. Two rounds in three months is rare in any environment; in this one, it reads as a signal that token production is where the next generation of AI infrastructure fortunes will be made.
Which will matter more for AI's economies — owning the frontier model or controlling the cheapest way to run it? Tell us in the comments.
