On 23 September, the Qinghai project information network published the bid result for phase one of the Kunlun Lingyi intelligent computing centre project. Shanghai GPU leader Biren Technology won package one with a bid of RMB 174 million.
The phase-one project issued its bid invitation on 16 September with a procurement budget of RMB 626 million, split into package one at RMB 174 million and package two at RMB 452 million. Package two was voided after fewer than three suppliers submitted bids.
At last week’s close, Biren’s Hong Kong shares traded at HK$38.16 (about RMB 32.68), for a total market value of HK$98.9 billion (about RMB 84.7 billion).
On 17 September Biren released its 2026 interim report. For the first half of 2026 revenue reached RMB 1.236 billion, up 1,997.6 per cent year on year. Period loss was RMB 377 million, narrowed 76.4 per cent. Adjusted period loss was RMB 337 million, narrowed 38.9 per cent. Gross margin was 42.7 per cent, up 1,080 basis points. R&D spend was RMB 804 million, up 40.7 per cent.
At this year’s World Artificial Intelligence Conference the company launched a next-generation NPO (near-package optics) optical-interconnect, distributed decoupled-architecture supernode, building a three-tier product matrix from a 16-card standard server to a 128-card cabinet to a 1,024-card optical-interconnect supernode, completing the shift from electrical to NPO interconnect ahead of domestic peers.
Biren’s Bili training and inference products already serve top internet firms, large AI model developers, national AI compute platforms, AI data centres, telecom operators, and government and enterprise clients across manufacturing, aerospace, electronics, energy, utilities, finance and education. With partners it built a Beijing-Tianjin-Hebei benchmark domestic compute cluster, and with another partner released a commercial distributed optical-interconnect optical-switch GPU supernode already deployed at several-thousand-card scale, with measured end-to-end training performance up over 30 per cent on typical mixture-of-experts models.
In training, Biren worked deeply with several large-model firms to complete pretraining and reinforcement learning of multimodal models on a thousand-card Bili cluster, achieving domestic migration and commercial operation of the full train-finetune-inference pipeline, with precision fully aligned to international products.
The next-generation product series, on schedule for release and production, is built on Biren’s own instruction set, compute cores and key IP, plus a Chiplet architecture, with comprehensive upgrades in architecture, compute, memory capacity and bandwidth, and interconnect. It supports low-precision FP8 and FP4 compute and the in-house BLink 2.0 interconnect protocol scaling up to 1,024 cards.
At 30 June 2026 the company had 910 employees and 1,917 patent applications, of which 1,823 were invention patents, with 1,006 granted patents including 927 inventions, placing it at the front of China’s GPU field. Going forward Biren will focus on three strategic directions: accelerating next-generation product production and delivery; pushing chip-software-system co-design around PD separation, attention-feedforward separation and large-scale expert parallelism plus trillion-parameter MoE training; and deepening ties with top clients and ecosystem partners to lower migration barriers.
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/597604.html.