DigiTimes复盘壁仞科技BR20X GPU点亮及40亿港元折价配股，核心事实均来自10月8日旧公告，无新增催化

China's Biren Technology has successfully powered on its next-generation BR20X GPU, advancing to post-tapeout validation while preparing for customer sampling and mass production. The milestone advances its AI accelerator roadmap, though US export controls and manufacturing constraints remain major obstacles.

Biren disclosed the progress in an October 8 share-placement announcement, confirming that BR20X had passed initial startup testing . The company is discussing deployments with customers and planning large-scale production, with sampling and manufacturing targeted over the coming quarters, according to DRAMeXchange and Shifair. System validation, software optimization, and customer qualification remain outstanding.
Founded in 2019, Shanghai-based Biren develops general-purpose GPUs (GPGPUs) and AI computing systems. Its portfolio has expanded from BR100 to BR106, BR110, and BR166. The BR106 and BR166 families entered full-scale production in 2025, with multiple thousand-GPU computing clusters delivered. 2025 revenue reached CNY1.035 billion (approx. US$145 million), up 207.2% year-over-year.
US export controls shape Biren's GPU roadmap
Biren has faced US export restrictions since October 2023, when the Bureau of Industry and Security (BIS) added several related entities in Beijing, Guangzhou, and Shanghai to its Entity List, restricting access to certain advanced semiconductor technologies and manufacturing services.
The company has consequently expanded its proprietary GPU architecture, software, and supply-chain capabilities. Its BIRENSUPA platform integrates chips, drivers, AI frameworks, and applications to reduce dependence on external computing ecosystems.
BR20X forms part of a broader roadmap spanning cloud AI training and inference, followed by BR30X for cloud computing and BR31X for edge inference. The latter two are targeted for commercialization around 2028, according to IT Times .
BR20X targets 1,024-GPU AI systems
Developed since 2024, BR20X uses Biren's second-generation architecture, combining proprietary instruction sets, computing cores, and chiplet technology. It supports 8-bit (FP8) and 4-bit (FP4) low-precision floating-point computation, with planned improvements in processing performance, energy efficiency, memory capacity, bandwidth, and interconnects.
Its proprietary BLink 2.0 protocol supports scale-up configurations of up to 1,024 GPUs. Biren is developing three system configurations:
● 16 GPUs: Standard AI computing servers.
● 128 GPUs: High-density rack-scale systems.
● 1,024 GPUs: Distributed, disaggregated supernodes incorporating near-package optics (NPO) interconnects.
The architecture reflects a shift from individual accelerator performance toward system-level computing efficiency, where GPU coordination and interconnect bandwidth increasingly determine large-scale AI performance.
Chinese reports have suggested BR20X could outperform Nvidia's H200, but comparable independent benchmarks remain unavailable.
HKD4.02 billion placement targets manufacturing capacity
Biren announced plans on October 8 to place 130 million new H shares at HKD31.08 (approx. US$3.96) each, raising approximately HKD4.04 billion gross and HKD4.02 billion net. CICC and UBS are placement agents, with the offer priced 9.76% below the previous closing share price.
The proposed proceeds will be allocated as follows:
● 70%: Strategic component procurement, production preparation, and commercialization.
● 20%: R&D and software ecosystem development.
● 10%: Working capital and general corporate purposes
Biren aims to secure critical components and manufacturing capacity amid shortages and longer lead times. Its existing GPU systems already serve major internet companies, AI developers, national computing platforms, and telecom operators, but BR20X's chiplet architecture and larger computing configurations will require more demanding production and integration capabilities.
Industry costs are also rising. Goldman Sachs and Morgan Stanley expect TSMC's advanced-node foundry prices to increase 5%–10% in 2027, while Goldman Sachs forecasts 0%–3% increases for mature processes. TrendForce projects conventional dynamic random-access memory (DRAM) contract prices will rise 10%–15% quarter-over-quarter in the fourth quarter of 2026.
These pressures compound Biren's restricted access to advanced manufacturing services under US export controls. The proposed share placement also remains subject to completion and applicable approvals.
Mass production remains the decisive milestone
Though Biren says BR20X development is progressing ahead of earlier expectations, the chip remains in post-tapeout verification. Customer sampling, qualification, and production scaling are still pending, with no independently verified performance or yield figures disclosed.
The company's ambitions extend beyond GPU specifications. Its 1,024-GPU architecture, proprietary interconnects, and software platform are designed to support increasingly complex AI workloads, while subsequent products will extend its reach from data centers to edge computing.
Yet successful silicon testing is only the first step. Biren's ability to secure manufacturing capacity, achieve reliable yields, and convert BR20X into sustained commercial deliveries will ultimately determine whether its expanding AI chip portfolio can compete at scale under US export restrictions.
Key considerations for Biren's BR20X rollout
● Technical readiness: BR20X has passed initial power-on testing, but performance, manufacturing yields, and customer qualification remain unverified.
● Nvidia competition: Claims of superior performance to Nvidia's H200 remain unsubstantiated by independent, comparable benchmarks.
● Funding uncertainty: Biren's proposed HKD4.02 billion placement has yet to be completed, leaving the expected proceeds subject to transaction risks.
● Memory costs: TrendForce forecasts a 10%–15% rise in conventional DRAM prices in the fourth quarter of 2026, while reports of a 15%–20% high-bandwidth memory (HBM) increase remain unverified.