China's token factories rake in billions: Wuwen Xinqiong closes record $310M Series C

智东西 · wechat · 2026-10-16

Deep dive into China's 2026 funding wave for "token factories" (large-scale inference services). Wuwen Xinqiong raised a 2.3B yuan Series C (4.3B total, a national AI-infra record) and is prepping an IPO; SiliconFlow has raised 2.9B yuan in 2026, with daily token throughput of 578.5B (peak 1.07T), 2025 revenue of 55.3M yuan but a 345M yuan net loss at -24% gross margin. Overseas benchmarks: Fireworks AI ($17.5B valuation), Baseten ($13B), Together AI ($8.3B), all NVIDIA-backed.

IDC projects China's public-cloud MaaS token consumption to grow 20x in 2026; Wuwen Xinqiong's daily calls grew 40x since late 2025. Funded companies share trillion-level real traffic, in-house inference engines/KVCache scheduling, and locked-in GPU/power/IDC resources.

The moat is production know-how: heterogeneous-chip training (97.6% utilization across Ascend and other domestic chips plus AMD/NVIDIA), PD disaggregation cutting first-token latency 51.5% and cost 37.5%, 90%+ cache hit rates.

Facing cloud price wars, independents bet on extreme inference efficiency, multi-model/multi-chip orchestration, and private deployment; the market should consolidate around a few scale platforms.

Original post →

More from Venture

Venture channel →