Kimi, Zhipu, MiniMax, StepFun All Partner with Infinigence as It Unveils Agentic Infra Strategy at WAIC 2026

At the WAIC 2026 forum in Shanghai, two of China's leading AI model companies — MiniMax and StepFun — appeared side by side at the same AI infrastructure company's event. MiniMax signed a strategic cooperation agreement, while StepFun delivered a keynote speech. This was no coincidence: four months earlier at the Zhongguancun Forum, Zhipu CEO Zhang Peng and Kimi founder Yang Zhilin shared a roundtable stage with Infinigence's co-founder and CEO.

Infinigence (无问芯穹) is emerging as the unified infrastructure layer for China's AI industry, and this week it unveiled its full vision. The backdrop is a rapidly shifting AI compute landscape: inference has overtaken training as the primary consumer of AI compute, with inference costs dropping 280x in two years. China's daily token calls have surpassed 140 trillion, growing 40% year-over-year.

At WAIC, Infinigence announced a self-developed cross-cluster heterogeneous PD (Prefill-Decode) separation architecture. By splitting Prefill and Decode phases — which have fundamentally different hardware requirements — across clusters, the company achieved 51.5% lower time-to-first-token (TTFT) and 37.5% lower per-token cost. A three-stage PDD architecture (P, RelayDecode, MainDecode) masks inter-cluster latency by having RelayDecode serve tokens during data transfer.

The company also launched an intelligent computing cluster maintenance agent system, operating 24/7 to transform ops from "humans hunting for problems" to "problems finding humans." The system improves ops team efficiency by 5x and critical fault handling by 6x.

Infinigence's "front shop, back factory, one center" Agentic Infra strategy encompasses: a compute distribution center deploying 37,000P of computing power across 16 chip types; a Token Factory (Agentic MaaS platform) whose daily token calls grew 40x since December 2025; and an AI Productivity Store covering gaming, healthcare, legal, and energy sectors. Cross-cluster reinforcement learning training has run stably for one week straight with zero interruptions.

The company partners with Shanghai Mobile on the "Tianwen" model service portal and has launched "TokenDance" for developers, analogous to OpenRouter. Its overseas counterparts — Baseten, Together AI, and Fireworks AI — are all valued between $13 billion and $15 billion. Unlike them, Infinigence must serve a fragmented Chinese chip ecosystem rather than a single Nvidia stack, making heterogeneous compute a core competitive moat.