Biren's next-generation chip is currently in the critical pre-production phase, akin to a restaurant's soft opening, where final validation and customer onboarding occur before full-scale delivery.
Much like a new restaurant requires a soft opening before its grand debut, a data center chip must clear three major hurdles between design freeze and large-scale delivery: tape-out validation, system integration, and customer adaptation. Each of these stages is more time-consuming than a product launch. On October 3, industry media in Taipei, China, reported that Biren Technology's next-generation data center GPU, the BR20X, is preparing for mass production, with a roadmap focused on wider low-precision support and locally accessible manufacturing. This article synthesizes public reports and brokerage research; actual mass production and commercialization timelines are subject to the company's future announcements.
Organizing known information into three layers—product, delivery, and funding—clarifies the position of this seven-year-old general-purpose GPU company.
Product Layer: FP8/FP4 Is the Core Focus of This Generation
According to the company's previous product roadmap and multiple brokerage reports, the BR20X is Biren's flagship data center chip based on its second-generation proprietary architecture, initiated in 2024 for cloud training and inference. Compared to the previous generation, its positioning includes a specific tilt: while maintaining advantages in training, it offers specialized optimization for the exponentially growing compute demands of the inference era. Upgrades focus on four areas: larger memory capacity and bandwidth, faster interconnects, a redesigned super-node solution, and native support for FP8 and FP4 low-precision formats. Low-precision support is becoming increasingly critical in the inference era—as model providers deliver quantized models as their primary offering, a chip's support for FP8/FP4 directly determines the task volume processed per unit of power. On the interconnect side, the series will feature a super-node solution with the proprietary Blink2.0 interconnect protocol, supporting scale-up configurations up to the thousand-card level, targeting large-scale parallel computing.
Delivery Layer: The Previous Generation Has Already Paved the Way
Assessing the mass-production readiness of the BR20X, the delivery track record of the previous generation serves as the most concrete benchmark. In 2025, Biren completed full-form-factor mass production and scaled delivery of the BR106 and BR166, serving customers including national AI computing centers, telecom carriers, commercial AI data centers, and model developers. In collaboration with ecosystem partners, the company launched the 'LightSphere X' supernode solution, leveraging optical interconnect and distributed optical switching technology to achieve a 2,048-card optical interconnect cluster. This solution has been deployed on national-level computing platforms and won the SAIL Award at the 2025 World Artificial Intelligence Conference. On the software side, the in-house platform supports mainstream large model series such as DeepSeek, Qwen, Zhipu GLM, and Tencent Hunyuan, and is compatible with major frameworks like PyTorch and vLLM. From individual units to clusters to the software stack, the company's delivery pipeline is fully operational, allowing the BR20X to plug directly into an established production line.
The shift in customer composition is also noteworthy. According to brokerage research, Biren's top five customers accounted for approximately 70% of revenue in 2025, all of which were newly acquired clients that year. This indicates that the revenue structure, previously reliant on landmark projects, is loosening, with income sources expanding from a few pilot projects to a broader customer base. For the BR20X, a new product targeting volume scaling, this is more critical than any single large order: the first batch of orders following the chip's launch will likely come from these customers who have already navigated the previous generation's delivery process, incurring the lowest migration costs.
2,048 cards
Optical interconnect cluster scale
2 types
New native low-precision formats
Thousand-card scale
Blink2.0 supernode upper limit
The three figures above represent, respectively: the scale of the optical interconnect cluster delivered by the previous generation product and ecosystem partners; the two new native low-precision formats, FP8 and FP4, supported by the BR20X; and the upper limit of the scale-up capability supported by the supernode solution based on a proprietary interconnect protocol, according to brokerage reports.
Funding Layer: Capital for the Commercialization Window Is Secured
Financially, Biren reported revenue of 1.035 billion yuan in 2025, a year-on-year increase of approximately 207%, driven primarily by intelligent computing solutions. A significant portion of the reported net loss stems from the accounting treatment of pre-IPO investor redemption rights, while R&D investment remained high at approximately 1.476 billion yuan. In July 2026, the company completed a placement of approximately 7 billion HKD, with proceeds allocated to the commercialization of next-generation products, R&D, and supply chain development. Between the window for domestic substitution and the continuous cash burn of R&D, the commercialization pace of the BR20X serves as the cash-flow switch for this funding chain.
My assessment: If the BR20X enters commercialization as scheduled in 2026, the real test will come in 2027. In the second wave of procurement orders from AI computing centers and cloud providers, its market share relative to competitors will be determined by the depth of software adaptation and delivery stability, with chip specifications being secondary. Falsification condition: If the BR20X series does not appear on the public procurement lists of any major cloud provider or AI computing center by the end of 2027, it will indicate that volume ramp-up has fallen short of expectations, and this assessment will be invalidated.
The competition for next-generation chips increasingly resembles a contest over who has the smoother kitchen workflow. Which hurdle do you think is the hardest to clear in this BR20X soft launch?