On October 8, domestic GPU maker Biren Technology disclosed a significant milestone: its next-generation AI chip, the BR20X, has been successfully lit up, completing boot testing and formally entering post-tape-out validation, ahead of the company's earlier schedule. Biren says it is accelerating customer sampling and commercialization preparations, with sampling and mass production planned over the next several quarters, though validation and certification work remains.

The same day, Biren announced a placement plan: 130 million new H shares at HK$31.08 apiece, for expected gross proceeds of roughly HK$4.04 billion and net proceeds of about HK$4.02 billion.

What "Lighting Up" Means

BR20X Technical Specifications

In the chip industry, "lighting up" marks a specific checkpoint: after a design is finalized and sent to the foundry for tape-out, the first silicon comes back packaged, gets powered on for the first time, and runs basic functions. It is the first gate a chip passes on the way from blueprint to physical product, but only the first.

That line is drawn explicitly in the announcement itself: successfully lighting up a product is not the same as completing performance validation or reaching mass production.

The BR20X is built on Biren's in-house instruction set, in-house compute cores and a chiplet architecture. Chiplet design abandons the approach of packing every function into one large die, instead splitting the design into multiple smaller dies linked together through advanced packaging, which allows architecture and packaging to compensate for performance when access to advanced process nodes is constrained.

Compared with the previous generation, the chip focuses its upgrades on four areas: compute performance, memory capacity, memory bandwidth and chip-to-chip interconnect, with added support for FP8 and FP4 low-precision computing. Lower bit widths matter because they raise throughput per unit of computing power and cut memory footprint, and quantized models carry noticeably smaller weights, one of the main routes currently used to bring down large-model inference costs.

On interconnect, the BR20X carries Biren's in-house BLink 2.0 technology, designed to support scale-up clusters of up to 1,024 cards, targeting large-model training and inference as well as the market for super-node computing power. Scale-up here means wiring many accelerator cards into a single high-speed interconnect domain so they behave as one large machine, as opposed to scaling out across data halls; a super node can be understood as a large pool of accelerator cards organized through high-speed interconnect into one unified compute unit.

HK$4.04 Billion Placement, Roughly 70% Headed to the Supply Chain

The terms: 130 million new H shares at HK$31.08 each, with expected gross proceeds of about HK$4.04 billion and net proceeds of about HK$4.02 billion.

About 70% is earmarked for supply chain procurement, capacity preparation and commercialization of the new generation of products, with the remainder going to R&D, software ecosystem upgrades and working capital. The placement still requires regulatory approvals and has not been completed.

From Single-Card Performance to System-Level Competition

Biren said that as large models, agents and AI data centers expand quickly, demand for GPUs, key components and wafer capacity keeps climbing, putting pressure on procurement costs and delivery timelines.

At the same time, customer demand is shifting from individual GPUs toward super-node and rack-level computing power solutions that integrate interconnect, networking, storage and cooling. The BR20X push signals that Biren is moving beyond a contest over chip performance and into competition over large-scale AI clusters and system-level computing power.

Three Gates Still Ahead

The BR20X remains in post-tape-out validation. Under the company's roadmap, it still has to clear customer sampling, customer certification, yield ramp in mass production and volume delivery in sequence, each carrying its own uncertainty. Competition among domestic GPU makers is entering a new and decisive phase.