---
title: "Biren: China's NVDA—How Capable Is It?"
type: "Topics"
locale: "en"
url: "https://longbridge.com/en/topics/42001604.md"
description: "On Jan 2, 2026, $BIREN TECH(06082.HK) debuted on HKEX, with the IPO priced at the top end at HK$19.60. The stock opened up 82.14%; intraday mkt cap briefly topped HK$100bn, and it closed at about HK$82.5bn.The retail tranche was oversubscribed 2,347x, highlighting a strong retail bid. Sentiment was buoyed by its scarcity as the first GPU company listed in Hong Kong.The 'domestic $NVIDIA(NVDA.US)' label further stoked buying as investors chased a China Nvidia proxy. But setting aside the aura..."
datetime: "2026-06-18T14:03:31.000Z"
locales:
  - [en](https://longbridge.com/en/topics/42001604.md)
  - [zh-CN](https://longbridge.com/zh-CN/topics/42001604.md)
  - [zh-HK](https://longbridge.com/zh-HK/topics/42001604.md)
author: "[Dolphin Research](https://longbridge.com/en/news/dolphin.md)"
---

# Biren: China's NVDA—How Capable Is It?

On Jan 2, 2026, $BIREN TECH(06082.HK) debuted on HKEX at the top-end issue price of HKD 19.60. The stock opened +82.14%, with intraday mkt cap briefly topping HKD 100 bn and closing around HKD 82.5 bn. The retail tranche was oversubscribed by 2,347x, underscoring frenzied demand.

The rally reflects its scarcity as the first GPU listing in Hong Kong, and the 'China’s $NVIDIA(NVDA.US)' label supercharged capital interest. Stripping away the halo, the key question is whether Biren is a high-quality company.

**1\. Journey: A company forged under the Entity List**

To understand Biren, you have to trace its path, which mirrors the trials of China’s domestic chip industry. The journey itself reads like a tough history of local semis.

**1) 2019–2022: VC backs domestic chip substitution.** The company was founded in 2019, the year the U.S. put Huawei on the Entity List and HiSilicon was forced to stop supplies. Primary markets pivoted to the 'domestic chip substitution' theme, with China’s GPU gap demanding startups that could carry the 'domestic NVIDIA' narrative while protecting investors’ capital.

Against that backdrop, founder Zhang Wen stepped in with a highly eclectic resume. He studied engineering at SJTU and later law and biz overseas, started as a lawyer, then turned investor, and only entered semis in 2011, assembling a top-tier GPU team from AMD, Qualcomm and Huawei HiSilicon.

With narrative, talent, vision and scarcity (neither $Moore Threads(688795.SH) nor $MetaX(688802.SH) existed then), capital flocked in. From 2019–2021, the company completed seven financing rounds totaling RMB 4.7 bn, with post-money valuation over RMB 11 bn at Series B.

**2) 2022–2024: From splashy product debut to forced line 'upgrades'.** In Aug 2022, Biren unveiled BR100/104, with BR100 as the flagship. The company pitched it as '3x+ A100 compute and near the unreleased H100' and rolled out the self-developed BIRENSUPA stack, fueling a China H100 narrative.

But the specs told a story: $Taiwan Semiconductor(TSM.US) 7nm, CoWoS-S and HBM2e. Savvy investors could anticipate the next shoe to drop, and two months later BIS introduced 3A090 export controls, directly capturing BR100 parameters.

Four months on, in Jan 2023, Biren abruptly mass-produced BR106. Where did it come from? Dolphin Research believes BR106 was a 'de-rated' iteration of BR104 to navigate the new rules, likely still made at TSMC initially, then shifted onshore after it was added to the Entity List.

Note: _In Oct 2023 Biren was targeted for blocking, as BIS added it and 12 other Chinese entities to the Entity List with footnote 4, requiring foundries to obtain licenses before deliveries. **TSMC foundry access was shut**_**.**

From sanctions to domestic line ramp, Biren moved reasonably fast. Yet across the two rounds of sanctions, co-founders Jiao Guofang and Xu Lingjie left, and with a high-valuation 'domestic NVIDIA' that suffered a flagship abort, spec downgrades, forced line shifts, and team departures, the market halo faded.

**3) 2024–present: Commercialization starts**

The company regained cadence only in 2024. BR166, a dual-die high-compute config within the 100-series and currently the revenue driver, only entered mass production in Aug 2025.

There are three products under BR100: single-die BR106, dual-die BR166 (revenue driver) and IoT chip BR110. In essence, all three trace back to designs from 2022.

**Including Biren, the value reset for domestic chip startups** came in H1 2025 when H20 was banned. Even after subsequent relaxations and H200 flows, Chinese cloud budgets were rigidly allocated to domestic chips, cementing 'national chips' as a **must-have** under substitution.

With the new narrative in force, Moore Threads and Muxi listed on STAR, while Biren and $ILUVATAR COREX(09903.HK) listed in Hong Kong, and Enflame is set to follow soon. Within this repricing, what makes Biren different, how long can it ride the tailwinds, and where are the core watchpoints? We continue.

**2\. Near- to mid-term: Shipments rule**

**For domestic chips, the near-term earnings logic is simple: supply rules.** Whoever can ship at scale captures cloud capex budgets, as $NVIDIA(NVDA.US) went from near-100% share in China’s GPU market to almost zero, and the Agent era further exploded demand. The supply-demand gap is massive.

By Dolphin Research’s rough math, the gap likely won’t close until 2028–2029. Demand becomes the main driver only after 2029, and the 2026 shortfall should be around 40–50%.

**1) Supply: Advanced capacity is scarce**

It’s consensus that domestic GPU single-die performance lags, typically 1–2 generations behind NVIDIA/$AMD(AMD.US). Local players stack dies (Die-Die/GPU-GPU) and interconnect to bridge gaps, but they can’t erase process disadvantages vs. offshore foundries.

\- EDA: Local EDA leader $EMPYREAN(301269.SZ) holds ~15.7% share domestically, but only 2–3% globally. It has yet to offer a full digital IC suite, and the U.S. tightly controls exports of EDA for advanced nodes (≤7nm).

\- Lithography: China still relies on $ASML(ASML.US) DUV imports. Procurement accelerated in 2025 to hedge tighter controls, while EUV export bans remain the critical choke point.

\- Foundry: Equipment limits mean domestic advanced nodes (N+2: 7nm; N+3: 5nm) are scarce capacity. $SMIC(00981.HK) holds most of the domestic advanced capability, with many local GPU designers queuing for SMIC capacity. Hua Hong Semi offers small-scale production, but yield and capacity need validation.

Given capacity constraints, some GPU vendors have to use N+1 lines (first-gen 7nm). Performance is tightly tied to the process, so this is an unavoidable trade-off under scarcity.

Dolphin Research estimates supply via 'wafer capacity → AI allocation → manufacturing yield → dies per wafer → packaging yield → total die supply'. It’s a rough framework to size deliveries.

**2) Demand: AI agent demand is surging**

On demand, imports of AI logic chips are materially constrained, making domestic substitution a necessity. As most local products still struggle for training, we focus mainly on inference demand for now.

a) Internet majors such as Alibaba, ByteDance and Tencent.

b) Foundation model players like DeepSeek, MiniMax and Zhipu. They consume tokens non-stop, have outsized single-point demand, and can bypass CSPs to request directly.

These cohorts benefit from structurally higher inference loads. Leading models are already consuming over 10 tn tokens per day, making them the primary demand drivers.

c) Telcos, SOEs and local governments as sovereign AI players. Demand stems from national AI infra, data sovereignty, compute centers and public sector applications.

We derive demand via domestic AI capex plans, chaining 'client AI/cloud capex → minus CSP intl spend → ×server mix ×AI server mix ×accelerator mix → AI GPU TAM → /ASP/compute to back-solve cards → GPU die count → CPU:GPU ratio → CPU die count → total die demand'.

Overall, 2026 requires ~4.2 mn AI chips, while supply is only ~2.6 mn. The gap is visible to the naked eye.

**3\. Short term: How strong is Biren’s capacity-locking?**

To gauge Biren’s near-term revenue burst potential, the essence is whether it can lock enough capacity to ship at scale. Our checks indicate foundry capacity is rationed via a quota-like system in today’s tight environment.

Capacity is tilted toward firms with growth potential, avoiding extreme concentration at a single head provider. Under this planned allocation, Dolphin Research believes Biren, as a relatively second-tier chip vendor (we explain why below), may actually benefit.

**1) Baseline to get quota: Pass model readiness**

Quota allocation prioritizes Day 0 support the moment models release, with a strong emphasis on software capability. Domestic AI chip firms work closely with local model developers to achieve Day 0/deep adaptation.

Most leading vendors claim Day 0 support at release. Actual runtime smoothness is another matter, but Biren’s compatibility with mainstream models meets the pass baseline.

Next, the 'soft' condition for locking capacity: funding and relationships. Sovereign capital matters.

\- Sovereign: Hygon/Cambricon have strong central state backing, likely to benefit first from SMIC’s advanced lines (Huawei still top priority). Shanghai also supports Biren/Muxi/Enflame locally, so Biren is not the sole favorite.

\- Industry: Before IPO, Enflame/Cambricon had Tencent/Alibaba on board, giving stronger demand-side assurance vs. Biren. Biren currently lacks a tightly bound mega-client relationship.

**Despite no unique 'deep ties', it should still secure quota.** For local gov., all vendors are stakeholders and hard to favor one over another, which helps second-tier firms get better-than-market capacity shares.

**Second,** $HUA HONG GRACE(01347.HK) **(Shanghai SASAC 51.59%, Shanghai Guosheng 18.36%, Shanghai Intl Group 18.36%, Shanghai Instrument 11.69%)** appeared as a 'close associate of existing minority shareholders' in Biren’s placement list (albeit at just 0.13%). In short, a wafer foundry is a shareholder.

Our checks also show Biren is engaging foundries beyond SMIC and has prepared two product paths. As Hua Hong’s 7nm ramps, **Biren should gradually secure wafer supply alongside Hua Hong’s capacity release.**

**2) Why might yield constrain?**

Even with the above positives, SMIC’s advanced capacity still won’t prioritize Biren. We expect most supply to come from Hua Hong, whose advanced nodes are early in ramp, with capacity to climb over time.

These nodes rely on DUV multi-patterning rather than EUV. That typically means lower yields and energy efficiency, which constrain shipments and form the most critical near-term risk.

**4\. Long term: Can Biren outperform?**

Over the long run, competition is about product strength. It spans three dimensions.

a. Single-die capability: core specs and iteration speed. b. Hardware-software integration and ecosystem. c. System-level delivery.

Among domestic players, only HiSilicon has strong system-level delivery so far. Most are still focused on the chip alone, and for Biren specifically: 1) GPGPU-based hardware systems; 2) BIRENSUPA software platform.

**1) GPGPU-based hardware systems**

In AI accelerators, ASIC, FPGA and GPGPU are the mainstream tracks. Biren designs GPGPU chips, accelerator cards and servers, operating fabless and outsourcing wafer manufacturing and packaging/test.

Biren has mass-produced BR106/BR110/BR166, delivered via PCIe cards and OAM modules. The core specs have been summarized as follows.

After BR100/104 were aborted, the company stopped disclosing fully comparable specs. Everbright Securities notes single-die (BR100/104/106) performance is strong among domestic peers, supporting a +273% YoY surge in total chip shipments.

Those are historical products, while Dolphin Research focuses on the BR166 flagship and the coming BR20X series. They matter more for forward competitiveness.

**a) Is BR166 competitive?**

BR166 is essentially a dual-die BR106 with doubled compute and unchanged architecture, and BR106 traces back to the 2022 BR100 architecture. Per the prospectus: '_we use chiplet tech to integrate two BR106 dies and four DRAMs in one package… D2D bidirectional bandwidth between the two BR106 dies reaches up to 896 GB/s._'

The architecture (diagram below shows BR100; BR166 follows it with DRAM memory disclosure) was innovative for AI workloads then. It paired high compute with a very large on-die L2 cache (H100 50 MB vs. BR100 256 MB).

At the micro-architecture level, NVIDIA’s SM is the base scheduling unit, while Biren’s SPC breaks down into smaller EUs. It supports dynamic grouping across 4/8/16 EUs for finer-grained resource reuse based on workloads.

On matrix throughput, each EU has one T-Core (one SPC has 16 parallel T-Cores). NVIDIA’s Hopper uses four Tensor Cores per SM, and Biren’s approach suits chiplet architectures under domestic yield realities.

**Lack of native FP8 & FP4 support is the biggest flaw.** Without low-precision formats, efficiency in large-model training and inference is structurally disadvantaged, and these formats are now widely adopted.

Among local peers, since BR100 dates back to 2022, single-die compute remains decent, but Biren has been outpaced in interconnect, memory bandwidth/capacity and precision. Ascend 910C essentially matches H100, and Muxi/Enflame products sit between H100–H200, leaving BR166 less compelling.

Downstream partner mapping (yellow means delivered) corroborates this. **Biren mainly ships to telcos/intelligent compute centers and sovereign AI projects, without large CSP orders.**

A major reason is CSPs’ inference needs in the Agent era, which demand big memory and strong interconnect. Biren’s current lineup is relatively weak there, making BR166 a less competitive all-rounder.

Beyond product strength, Biren’s origin story underscores **geopolitical supply-chain uncertainty**. That makes it hard to convince internet customers of long-term, stable, high-volume supply capability.

That said, with demand running hot, those issues are less acute near term. Financials show visibility via orders, prepayments and inventory build.

\- Backlog: Binding orders under framework and sales contracts total RMB 822 mn, supporting future revenue. - Contract liabilities: As of end-2025, contract liabilities were RMB 77 mn, indicating prepayments that lock orders.

\- Inventory: End-2025 inventory net was RMB 949 mn, up over 500% YoY, with raw materials at RMB 386 mn and WIP at RMB 431 mn totaling ~85%. That suggests inventory expansion is driven by confirmed orders.

**b) What about BR20X?**

BR166’s overall competitiveness is not standout, but at the 2026 juncture a 3–4 year major iteration is due. The new BR20X flagship is imminent.

(1) BR20X series is in physical design and tape-out validation, with **commercial launch planned for Q4 2026**. Biren will boost single-die capability and accelerate ultra-node systems.

(2) As the next-gen flagship fully on domestic supply chains and free of export controls, the evolution aligns with our expectations. - FP8 & FP4 will be supported in BR20X to speed large-model training/inference.

\- Compute will be stronger. - **Memory will be larger and faster,** interconnect bandwidth higher, and ultra-node systems designed for scale.

We estimate BR20X will benchmark NVIDIA’s H200. **There is no public spec yet; Dolphin Research** derives an estimate from peer benchmarks and checks.

The new line mainly plugs memory capacity gaps, while interconnect likely still needs improvement. Overall, fixing those weaknesses should secure mega-client orders on product merit.

The key constraint remains **yield ramp on new lines**. That will determine volume and delivery reliability.

**2) BIRENSUPA software platform**

On software, CUDA is the wall no one can bypass. Its moat comes from first-mover accumulation, deep HW/SW binding for performance and compatibility, plus rich toolchains.

Local responses split into two paths. One is CUDA compatibility, like Moore Threads’ MUSA and Muxi’s MACA, which they sell as commercialization enablers, and the other is full-stack self-developed ecosystems like Huawei’s CANN and Cambricon’s Neuware.

The former binds sovereign orders via years of operator accumulation, while the latter goes deep in recommendation systems and wins ByteDance and other CSPs. Biren initially aimed for a self-developed ecosystem too.

BIRENSUPA’s **open-source depth and stack richness are clear shortfalls**, and the ecosystem is early. Engineering friction exists, with IR edits or higher compile error rates in complex scenarios, and some comm libs need tuning for sync latency in 1,000-card clusters.

To address gaps, Biren plans to invest 40% of IPO proceeds in the software ecosystem. It is adding comprehensive support across PyTorch, vLLM, SGLang and other mainstream frameworks to tap the CUDA-built installed base and lower customer migration costs.

Near term, **pivoting from pure self-developed to compatibility makes sense**. Biren’s scale, client mix and commercialization maturity can’t support a fully proprietary path, and customers ultimately judge solution price/performance, a HW/SW combination.

Longer term, **compatibility itself won’t create a moat**. Even under optimistic assumptions, Biren could be a better middleware than peers, but without differentiation in ease of use, stability and third-party ecosystem, it risks homogenized competition.

**3) Ultra-nodes: Is there a turnkey plan?**

Biren delivers via GPU clusters/ultra-nodes, but its value-add is mostly in the GPU. Interconnect, switching and rack-level integration are largely handled by partners such as Xizhi and ZTE.

Outside Huawei, most domestic players are early in interconnect self-development (often limited to card-level protocols). Switch chips (NVSwitch analogues) and rack system integration are typically joint efforts with upstream/downstream partners.

R&D today concentrates on compute chips and software stacks, leaving interconnect undifferentiated. Yet in AI, interconnect value is magnified, as scale-up/out/across determines model sizes per node.

We believe firms that can deliver 'turnkey' first will better build long-term ecosystem moats. Biren is not fundamentally different from other startups in this respect.

**5\. Valuation: Is there still upside?**

Net-net, Biren’s differentiation lies mainly in single-die capability. But for inference chips, the focus is shifting away from single-die compute toward memory and interconnect at the module level.

It has no edge in system-level ultra-node delivery, which is still early across the board, and in critical software ecosystems it’s not among the strongest. The positives are the quota model and continued policy support, and listing in Hong Kong provides an official endorsement.

**Near term, BR20X commercialization is imminent**, giving a strong baton for the next leg of revenue release. Profitability should improve with scale as shipments ramp, and the chip-design model is well validated by markets.

In an industry tailwind phase, scale shipments find customers easily. This supply-side logic plus an upcoming new product can push valuation into a 'no-strangers' overbought zone.

Dolphin Research sizes Biren’s long-term space via steady-state market size and competitive standing. If near-term prices exceed the company’s optimistic long-term standing, stay calm.

When pricing retreats toward a more pessimistic share, beta and policy tailwinds could create opportunities. Our long-term sizing puts AI GPU TAM at $80 bn by 2030, with Huawei as the clear leader and Cambricon the strongest independent.

Alibaba-affiliated T-Head and Baidu’s Kunlun should also have paths given internal needs. The remaining six 'GPU upstarts' include Muxi, Moore Threads, Biren, Hygon, Enflame and TianShu ZhiXin.

Using overseas concentration as a reference (NVIDIA ~80%, Broadcom ~10%, AMD high single-digit), China will be more dispersed as competition is not fully market-driven. That informs share assumptions.

**Base case:** By 2030E, Huawei ~70% and Cambricon ~10%, with eight others splitting the rest. Biren’s share at ~2.5% implies ~$2.0 bn revenue vs. ~$330 mn in 2026E, a ~57% revenue CAGR.

Applying ~15x PS in 2030 (with ~40% OPM, ~38x PE equivalent) and discounting at 13% back to 2026 suggests ~$18.5 bn valuation. That is broadly in line with the current ~$19.1 bn mkt cap.

**Optimistic:** At 5% share, revenue doubles to ~$4.0 bn and valuation nears twice the base. Given Biren’s current overall strength, a move into second-tier leadership seems unlikely, so that level would be a risk flag.

**Conservative:** At 1.5% share, apply a ~40% haircut to $18.5 bn for ~$11–12 bn. With sector tailwinds maxed and a major product cycle into end-2026, prices can skew toward the optimistic case.

Thus, waiting for prices near the conservative share scenario is more appropriate to look for entries. One note: the first lock-up expiry in Jul will free ~6% of shares, small in size, so watch near-term swings for better entry.

A larger unlock wave hits in Jan 2027, shortly after BR20X commercialization. That could elevate downside risks.

<End of text\>

Risk disclosure and statement:[Dolphin Research disclaimer and general disclosure](https://support.longbridge.global/topics/misc/dolphin-disclaimer)

### Related Stocks

- [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md)
- [HUAWEI.NA](https://longbridge.com/en/quote/HUAWEI.NA.md)
- [06082.HK](https://longbridge.com/en/quote/06082.HK.md)
- [QCOM.US](https://longbridge.com/en/quote/QCOM.US.md)
- [688795.CN](https://longbridge.com/en/quote/688795.CN.md)
- [09903.HK](https://longbridge.com/en/quote/09903.HK.md)
- [TSM.US](https://longbridge.com/en/quote/TSM.US.md)
- [AMD.US](https://longbridge.com/en/quote/AMD.US.md)
- [301269.CN](https://longbridge.com/en/quote/301269.CN.md)
- [NVDL.US](https://longbridge.com/en/quote/NVDL.US.md)
- [07788.HK](https://longbridge.com/en/quote/07788.HK.md)
- [07388.HK](https://longbridge.com/en/quote/07388.HK.md)
- [NVDY.US](https://longbridge.com/en/quote/NVDY.US.md)
- [NVDD.US](https://longbridge.com/en/quote/NVDD.US.md)
- [NVDX.US](https://longbridge.com/en/quote/NVDX.US.md)
- [NVDQ.US](https://longbridge.com/en/quote/NVDQ.US.md)
- [AMDL.US](https://longbridge.com/en/quote/AMDL.US.md)
- [00981.HK](https://longbridge.com/en/quote/00981.HK.md)
- [HSMD.SG](https://longbridge.com/en/quote/HSMD.SG.md)
- [688981.CN](https://longbridge.com/en/quote/688981.CN.md)

## Comments (3)

- **无敌斩 · 2026-06-18T17:18:57.000Z**: The level of this article is not good. It says the product performance can match the h200, but then it says the company's fundamentals are average.
  - **馬上** (2026-06-19T01:08:40.000Z): No one knows how the performance will be before 20x comes out, and there are still quite a few risks.
- **蔡徐kun · 2026-06-18T15:29:45.000Z**: When is the next breakout point?Supported by HKEX