Biren Technology wins 174M yuan AI computing center bid, delivery due in 45 days

September 24 — The results of the inquiry-based procurement for the Kunlun Lingyi AI Computing Center Project (Phase I) have been officially released. Shanghai Biren Technology Co., Ltd. won Package 1 with a bid of 174 million yuan, becoming the sole winning supplier for that package.

According to the announcement, the total procurement budget for the project is approximately 626 million yuan, divided into two packages:

• Package 1: Budget of approximately 174 million yuan
• Package 2: Budget of approximately 452 million yuan

Biren Technology secured Package 1 with a comprehensive evaluation score of 89.71 points, ranking first overall. Under the terms of the contract, Biren is required to complete delivery within 45 calendar days after signing.

Core Product: Biren Bili Series GPU

The core product in Biren's winning bid is the Bili series GPU. According to publicly available information, the Bili GPU is based on a classic GPGPU architecture, with its core computing unit—the SPC stream processor—integrating dedicated tensor computing units called T-Cores.

Biren has developed a three-generation product lineup: BR106, BR110, and BR166. The company's next-generation flagship training and inference chip, the BR20X, is targeted for release in 2026. It will support data formats such as FP8 and FP4, and will feature larger memory capacity and higher interconnect bandwidth.

Interconnect Capabilities as a Differentiator

Interconnect performance is a key competitive advantage for Biren. The company's self-developed BLink protocol enables direct GPU-to-GPU connections, with a single-channel bidirectional bandwidth of up to 64 GB/s. It can support up to an 8-card point-to-point direct interconnect configuration.

At the scale-out level, Biren employs an Optical Circuit Switching (OCS) solution. As of the end of 2025, the company has successfully delivered a 2048-card optically interconnected GPU supernode cluster, achieving a linear acceleration ratio of 95% for thousand-card clusters.

Ecosystem Compatibility

In terms of ecosystem adaptation, Biren has achieved Day 0-level compatibility with major large language models, covering leading domestic model developers including:

• Zhipu GLM
• DeepSeek
• Alibaba Qwen
• Moonshot AI (Kimi)

Its BIRENSUPA software platform supports mainstream frameworks such as PyTorch, TensorFlow, and PaddlePaddle, as well as inference frameworks like vLLM.

Package 2 Declared Void

Notably, Package 2 was declared void due to fewer than three suppliers submitting responsive bid documents. The fact that a bidding package with a budget exceeding 450 million yuan failed due to insufficient participation reflects the current concentration and scarcity of the domestic AI computing power supply chain.