Wubu Infra, a Tsinghua University spin-off, has closed a C round of RMB 2.3 billion, bringing cumulative financing to RMB 4.3 billion and setting a national record for China’s AI Infra category. Foreign media report the company is preparing an IPO.
From lab to heterogeneous compute
Wubu grew out of the high-efficiency computing group led by Tsinghua professor Wang Yu. In 2024 it completed the world’s first 70-billion-parameter heterogeneous mix-training across four domestic chips (Ascend, Iluvatar CoreX, MetaX, Moore Threads) and two import chips (AMD, Nvidia) at 97.6 per cent utilisation. It now adapts a new chip in two working days and cuts domestic-chip adaptation cost by over 70 per cent, and in trillion-parameter MoE training it lifts throughput 41 to 75 per cent over mainstream open frameworks while cutting inference cost tenfold in a year.
The three-part layout
At WAIC 2026 the company unveiled a front-shop, back-factory, one-centre structure. The compute hub aggregates over 37,000P of heterogeneous compute across 16 chip types, including a 4,000 km cross-region training run between Hami and Shenzhen that lifted joint performance 165 per cent. The token factory reported daily token calls up 40 times since December, with a new disaggregated inference architecture cutting first-token latency 51.5 per cent and per-token cost 37.5 per cent. The productivity store turns tokens into deliverables in gaming, law and energy.
A light asset
Wubu owns no data centre. It earns on efficiency as a neutral layer connecting models, chips and customers. Backers include Tencent, Baidu, Lenovo, Sequoia China and Hongshan. The contrast with US Neo Clouds is sharp: Fireworks AI is valued at US$17.5 billion, Baseten at US$13 billion and Together AI at US$8.3 billion, all riding Nvidia. Wubu must solve chip heterogeneity that is a Chinese necessity, not a Western one.
Editor’s note: This is an adapted translation of the original Zhidx report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.zhidx.com/p/599642.html.