Chinese AI chipmaker Biren to debut after $717m Hong Kong IPO

Shanghai Biren Technology, an AI chip designer, is set to begin trading on the Hong Kong Stock Exchange on December 2 after raising US$717 million in its IPO.

The company, which develops graphics processing units for AI models, priced its shares at HK$19.6 (US$2.52), the top of its indicative range.

Retail demand was high, with the offering subscribed over 2,300 times.

Biren is the first GPU-focused company to list in Hong Kong and comes to market amid strong investor interest in the sector.

Biren was added to a US trade restriction list in 2023, requiring exporters to obtain government licenses before shipping to the company.

The firm reported a net loss of 1.6 billion yuan (US$228.9 million) for the first half of the year.

IPO proceeds will fund further research and development.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Biren’s 2,300x retail oversubscription masks manufacturing and export control uncertainties

Materials still do not state who builds its chips after Taiwan Semiconductor Manufacturing Company (TSMC) halted output in October 2022. They also omit the process node (the manufacturing technology generation, measured in nanometers) and the status of licenses for foreign Electronic Design Automation (EDA) tools under U.S. export controls.
Biren reportedly modified the BR100 (the company’s flagship data center GPU) to process less data to bypass U.S. limits. The firm still landed on the Entity List in October 2023, a U.S. trade blacklist that requires special licenses and sharply limits access to some foreign suppliers.
The company posted 58.9 million yuan revenue in H1 2025 despite 2.1 billion yuan in pending orders (backlog), which may mean a slow ramp or contract hurdles under export controls. The 1.6 billion yuan H1 net loss exposes a gap that separates research and development (R&D) ambition from commercial traction.

Software tool providers can tap Biren’s evolving SDK ecosystem for Chinese cloud operators

BIRENSUPA (the company’s software development kit and runtime stack) supports hardware‑software tuning. The BR100 lacks FP8 (an 8‑bit floating-point format used to speed AI training and inference) support. It also gains no speedup from structured sparsity (a technique that prunes weights in predictable patterns to accelerate computation) as Nvidia does.
Independent Software Vendors (ISVs) and system integrators (firms that assemble plus optimize end-to-end solutions for enterprises) can ship compilers, kernel libraries or model‑optimization frameworks. These would serve China Mobile and China Telecom, which already use Biren GPUs in intelligent computing centers (AI-focused data centers). Export limits on advanced AI chips in China push cloud providers to seek domestic options. That creates demand for tools that port PyTorch and TensorFlow models to Biren’s architecture. Vendors can target BR100’s caching strategy and the BLink interconnect (Biren’s high‑speed chip‑to‑chip fabric) for services revenue. They can offer automated porting, performance profiling or mixed‑precision training toolkits (using multiple numeric precisions to improve speed and efficiency).