Biren Technology
Last reviewed
Sources
15 citations
Review status
Source-backed
Revision
v2 · 2,254 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
15 citations
Review status
Source-backed
Revision
v2 · 2,254 words
Add missing citations, update stale details, or suggest a clearer explanation.
Biren Technology (Chinese: 壁仞科技; pinyin: Birèn Keji) is a Chinese fabless semiconductor company, based in Shanghai and founded in 2019, that designs general-purpose GPUs and AI accelerators for data-center training and inference. Its flagship products are the BR100 and BR104, 7nm chips built by TSMC and unveiled in 2022, after which the company was cut off from that foundry by United States export controls and added to the US Entity List in October 2023. In January 2026 Biren became the first mainland Chinese GPU start-up to list on the Hong Kong Stock Exchange, where its shares closed 76 percent above the offer price on debut, making it one of the most closely watched test cases for China's effort to build domestic alternatives to NVIDIA hardware. [1][2][11]
The company unveiled the BR100, a 7nm GPU it claimed could rival contemporary NVIDIA data-center parts on some metrics, at the Hot Chips 34 conference in August 2022. Between that launch and its 2026 listing, US export controls suspended its access to TSMC and the Commerce Department added a group of Biren entities to the Entity List, alongside Moore Threads. [1][2][8]
Biren was founded in 2019 by Lingjie Xu (Xu Lingjie), a former NVIDIA and Alibaba engineer, together with co-founders previously employed at NVIDIA, Alibaba and STMicroelectronics. The chief executive is Zhang Wen (also known as Michael Zhang Wen), who previously served as president of SenseTime and holds a doctorate in juridical science from Harvard. Other early leaders included chief technology officer Mike Hong, who co-presented the BR100 at Hot Chips, and Jiao Guofang, a former Huawei engineer who ran the GPU product line. The company is registered through a group of entities, with Shanghai Biren Intelligent Technology Co., Ltd. as the principal operating company. [1][3][7]
The start-up raised money rapidly during China's domestic-chip investment boom, attracting roughly 991 million US dollars across multiple rounds by 2025. By 2022 its backers included Qiming Venture Partners, IDG Capital and Hillhouse, and reporting that year valued it at roughly 2 billion US dollars, making it one of the country's most valuable chip unicorns; a 2023 valuation of about 2.2 billion US dollars was widely cited ahead of the IPO. After it lost access to its foundry in late 2022, fundraising shifted toward state-linked investors: a roughly 280 million US dollar pledge from Guangzhou government-backed funds was reported in late 2023, followed by investment from Shanghai State-owned Capital Investment in early 2025 and a 1.5 billion yuan round in mid-2025 that put its pre-IPO valuation near 14 billion yuan. [1][3]
| Milestone | Date | Detail |
|---|---|---|
| Founded | 2019 | Shanghai; founders from NVIDIA, Alibaba, STMicroelectronics [1][3] |
| ~2 billion USD valuation | 2022 | Backers include Qiming, IDG Capital, Hillhouse [3] |
| BR100 unveiled | Aug 2022 | Hot Chips 34; 7nm, 77 billion transistors [4][5] |
| US export controls tighten | Oct 2022 | 600 GB/s interconnect cap; TSMC suspends Biren chips [1][8] |
| Added to US Entity List | Oct 17-19, 2023 | With Moore Threads; "advanced computing" rationale [2][9] |
| 1.5 billion yuan round | Mid-2025 | Pre-IPO valuation near 14 billion yuan [1] |
| Hong Kong IPO | 2 Jan 2026 | Stock code 6082; closed +76 percent on debut [11][12] |
The Biren BR100 is a 7nm general-purpose GPU for data-center AI training and inference, detailed at the Hot Chips 34 conference in August 2022. It is a chiplet design built on TSMC's 7nm process and assembled with TSMC's 2.5D CoWoS packaging. The full part combines two compute dies into a package of about 1,074 mm2 carrying around 77 billion transistors, a die area and transistor budget broadly comparable to NVIDIA's contemporary H100. Each compute die contains 16 Streaming Processing Clusters, and the chip carries 256 MB of on-die cache and 64 GB of HBM2e memory. The two dies are joined by an 896 GB/s die-to-die link, and the package targets a 1 GHz clock and a 550 W thermal design power. The BR100 supports FP32, BF16, FP16, INT32 and INT16, but does not support FP64. [4][5][6]
For interconnect between GPUs, Biren designed a proprietary link called BLink and supported PCIe Gen5 and CXL. The company introduced a custom data format, "TF32+," alongside standard precisions. Biren marketed two products from the design: the BR100 in the OCP Accelerator Module (OAM) form factor, and the BR104, a single-die PCIe card offering roughly half the performance at a lower power envelope. The BR104 line was positioned for mainstream inference and training where energy efficiency matters more than peak throughput. [4][6]
The headline performance numbers below are Biren's own figures from its 2022 disclosures; the parts have not been independently benchmarked at scale against NVIDIA hardware, and the company's comparisons should be read as vendor claims.
| Specification | BR100 | BR104 |
|---|---|---|
| Form factor | OAM module | PCIe card |
| Process | TSMC 7nm | TSMC 7nm |
| Architecture | Dual-die chiplet | Single die |
| Transistors | ~77 billion (package) | N/A |
| Memory | 64 GB HBM2e | 32 GB HBM2e |
| Memory interface | 4096-bit (~1.64 TB/s) | 2048-bit (~819 GB/s) |
| FP32 (claimed) | up to 256 TFLOPS | up to 128 TFLOPS |
| BF16 (claimed) | up to 1,024 TFLOPS (1 PFLOPS) | up to 512 TFLOPS |
| INT8 (claimed) | up to 2 POPS | up to 1 POPS |
| TDP | 550 W | ~300 W |
Biren said the BR100's 16-bit floating-point (BF16) throughput reached about 1,024 TFLOPS, meeting its target of "one petaflops horsepower" in a single package, and that the design was, on some AI workloads, several times faster than NVIDIA's A100 generation. "We had to raise the bar from hundreds of teraflops to 1 petaflops," co-founder and president Lingjie Xu told the conference, describing 1 petaflops of mixed-precision 16-bit tensor performance as the explicit design goal. These claims positioned the BR100 against the A100 and, in selected cases, the H100, though without standardized third-party results the comparison remains the company's. [4][5][7]
In October 2022 the US Bureau of Industry and Security tightened rules on the export of advanced computing chips and manufacturing equipment to China. The new controls restricted, among other things, integrated circuits with an aggregate bidirectional interconnect bandwidth of 600 GB/s or more. Because Biren relied on TSMC in Taiwan to manufacture the BR100, the rules put its supply at risk, and TSMC suspended production of Biren's chips while the parts were reviewed against the thresholds. [1][8]
According to analysis published at the time, the original BR100 combined 512 GB/s of BLink with 128 GB/s of CXL 2.0, for 640 GB/s, just above the cap. Biren reportedly cut the number of BLink connections from eight to seven, lowering the combined interconnect bandwidth to 576 GB/s and bringing the design under the 600 GB/s line. The single-die BR104 sat below the threshold by construction. These adjustments reduced data-movement capability while preserving most of the per-chip compute, an approach echoed across the industry that year (NVIDIA, for instance, produced the bandwidth-limited A800 and H800 for the China market). [8][9]
Yes. The reprieve from the 2022 review was temporary. The US Commerce Department's Bureau of Industry and Security announced on October 17, 2023, and published in the Federal Register on October 19, 2023, the addition of Biren and Moore Threads, a total of 13 entities, to the Entity List, naming a group of Biren affiliates (including the Shanghai, Guangzhou, Hangzhou, Zhuhai and Beijing entities) as "involved in the development of advanced computing integrated circuits." The designation barred US suppliers from selling to Biren without a license, with applications reviewed under a presumption of denial, and reinforced its loss of leading-edge foreign foundry access. The disruption was costly: co-founder Jiao Guofang departed in early 2023, and co-founder and president Lingjie Xu resigned, with his departure reported in early 2024. [2][9]
Biren's software platform is called BIRENSUPA. It bundles a hardware abstraction layer, a SUPA programming model, the BRCC compiler, deep-learning and general-compute acceleration libraries, toolchains, inference engines and application SDKs. The stack supports mainstream frameworks including PyTorch, TensorFlow and PaddlePaddle, and Biren positions the SUPA model as familiar to developers coming from NVIDIA's CUDA. As with other Chinese accelerator vendors, the maturity of this software ecosystem, rather than raw silicon, is widely seen as the harder problem in displacing NVIDIA, whose CUDA platform underpins most existing AI code. [6][10]
Biren listed on the Hong Kong Stock Exchange on January 2, 2026, under stock code 6082, the first mainland Chinese GPU start-up to go public there and the largest offering to date under the exchange's Chapter 18C regime for pre-revenue and early-commercialization technology companies. The company priced its offering at the top of a HK$17.00 to HK$19.60 range, selling about 284.8 million H shares at HK$19.60 each to raise roughly HK$5.58 billion (about 717 million US dollars) gross, with net proceeds reported at around HK$5.38 billion (about 692 million US dollars). The IPO drew 23 cornerstone investors, including Qiming Venture Partners, Ping An, UBS and Digital China, who committed about 372.5 million US dollars and agreed to a six-month lock-up. Institutional demand ran at nearly 26 times the shares on offer, while the retail tranche was oversubscribed by roughly 2,348 times. [11][12]
The shares surged on debut: they opened at HK$35.70, reached an intraday high of HK$42.88 and closed at HK$34.46, up about 76 percent from the offer price. Biren said most of the proceeds, around 85 percent, would fund research and development, principally its next-generation BR20X chip and software, plus commercialization and working capital. The BR20X, described as the company's next-generation flagship data-center series, is slated for launch in the second half of 2026 and is positioned as the main catalyst for reaching break-even by 2027. [11][13]
The prospectus showed a company still deep in losses but growing revenue from a tiny base as deployments expanded. Reported revenue rose from a negligible amount in 2022 to roughly RMB 336.8 million in 2024 and RMB 589 million in the first half of 2025; full-year 2025 revenue reached about RMB 1.03 billion, up 207.2 percent year on year, with gross profit of about RMB 557 million at a 53.8 percent gross margin. Net losses ran at roughly RMB 1.47 billion in 2022, RMB 1.74 billion in 2023 and RMB 1.54 billion in 2024, driven by research-and-development spending that reached around RMB 1.48 billion in 2025. Biren's first post-IPO results also reported a large one-time accounting loss for 2025 tied to the redemption of pre-IPO investor liabilities, with the adjusted (non-IFRS) net loss put at about RMB 874 million. The company reported its BR100 and BR104 GPUs being deployed in intelligent-computing centers run by customers including China Mobile, ZTE and the Shanghai AI Laboratory. [13][14][15]
Biren is one of several domestic challengers trying to fill the gap left by restrictions on NVIDIA's top accelerators in China. Its closest peers among start-ups include Moore Threads, with which it was jointly added to the Entity List, and MetaX, both of which also pursued Hong Kong listings. The dominant domestic player by volume is Huawei, whose Ascend line, including the Ascend 910C, is the most widely deployed Chinese AI accelerator, while Cambricon competes in both training and inference. Against this field Biren's distinguishing pitch is a general-purpose GPU architecture with a CUDA-like software stack, but its prospects depend heavily on securing advanced manufacturing capacity under Entity List constraints and on maturing BIRENSUPA enough to attract developers away from NVIDIA. [2][14]