Evaluation of Technical Routes and Commercialization Prospects of China's Four Leading GPU Enterprises

• Orientation: User context mentions it follows the “full-functionality route imitating NVIDIA”, with business layout covering AI intelligent computing, graphics rendering, and smart cockpit.
• Market and Capital Performance: Tool information shows it has been listed, with a strong debut performance; current market value is approximately RMB 359.5 billion; the company is in an expansionary loss period and is expected to achieve profitability as early as 2027 [1].
• Evaluation Points (Key Information to Verify): Product matrix (e.g., specific models and parameters for training/inference/graphics rendering), software stack compatibility (e.g., CUDA compatibility path), ecosystem maturity (frameworks, drivers, toolchains), mass production and customer structure (proportion of government/operators/internet/industry clients). Before obtaining more detailed performance and order data, the evaluation of its “full-functionality path” needs to be cautious.

• Orientation: User context states it focuses on “government-enterprise, medical, and other vertical domain toB businesses”; tool information shows three co-founders are from AMD, with “AMD genes” [2].
• Product System: Tool information shows core products cover intelligent computing inference (Xisi N Series), training-inference integration/general computing (Xiyun C Series), graphics rendering (Xicai G Series), among which Xiyun C600 achieves “full-process domestic supply chain closed loop” and has outstanding comprehensive performance [1].
• Market and Capital Performance: Listed on the STAR Market, raising approximately RMB 4.197 billion; online issuance received nearly 3,000 times subscription; strategic placement includes the National Artificial Intelligence Industry Investment Fund; debuted with a significant rise, with a total market value of approximately RMB 332 billion at one point; not yet profitable, but production and sales rate exceeded 100% starting in 2025, expected to achieve break-even as early as 2026 [2].
• Evaluation Points: Quality of orders and contracts in vertical toB scenarios (number of government-enterprise/medical clients and single-client value), gross margin path of chips and boards, stability and cost control of supply chain (manufacturing, packaging and testing). Current commercialization progress is relatively advanced among the four leading players.

• Orientation: User context states it “focuses on high-end training, targeting 100-billion parameter model training scenarios”, with a greater emphasis on training end and computing power density optimization.
• Market and Capital Performance: Tool information shows it has obtained备案 from China Securities Regulatory Commission, plans to list on Hong Kong Stock Exchange, and simultaneously promotes “full circulation” of unlisted domestic shares; current valuation is approximately RMB 15.5 billion [1].
• Evaluation Points: Performance and energy efficiency in large-scale training scenarios (e.g., model throughput, communication topology, video memory and interconnection), software stack and framework adaptation (e.g., support for mainstream large model frameworks), mass production delivery and benchmark clients (whether it has been deployed in large intelligent computing centers/internet clients).

• Orientation: User context states it “binds Tencent Cloud services, deep dives into inference scenarios”, focusing on the inference side and cloud service ecosystem.
• Market and Capital Performance: Tool information shows it has initiated the listing process [2].
• Evaluation Points: Depth of cooperation with cloud vendors (integration level, joint product release, conversion path from POC to commercial use), inference throughput and cost advantages, support for mainstream inference frameworks and model ecosystems.

• GPU (General-Purpose): Strong programmability, mature ecosystem, wide application scope (training + inference + graphics rendering), but energy efficiency and cost are not advantageous compared to DSA.
• GPGPU: Extension of general-purpose GPU to computing acceleration, focusing more on AI computing tasks while retaining good programmability.
• ASIC/TPU/DSA (Domain-Specific): More focused on specific loads (e.g., matrix multiplication, convolution, sparsification), with obvious advantages in energy efficiency and cost; however, long development cycle, high threshold for ecosystem construction, and relatively fixed application scenarios.
• Trend Judgment: User context points out that “NVIDIA will remain the mainstream of computing cards, but DSA market share will gradually increase”, which is consistent with the industry consensus that “general-purpose remains the base, and domain-specific is the growth pole”.

• Advantages: Covers multiple scenarios such as training/inference/rendering, with a wider target market; easy to establish extensive cooperation with cloud vendors and industry clients.
• Challenges: Directly competes with NVIDIA in ecosystem and performance, requiring huge R&D and ecosystem construction investment; mass production and delivery capabilities are hard constraints; large loss pressure, with financing and cash flow management being key.
• Observation Points: Large-scale commercial deployment of benchmark clients (government-enterprise, operators, leading internet), software stack maturity (e.g., CUDA compatibility and migration path), delivery stability on the manufacturing side (yield rate, production capacity, cost).

• Advantages: Scenario focus, easy to form deep customer relationships and stickiness; customized demands are conducive to differentiated competition; production and sales rate improvement means enhanced order fulfillment capability [1].
• Challenges: Market size and ceiling may be lower than general-purpose path; customer concentration risk (e.g., high proportion of single government-enterprise/medical large client); gross margin is restricted by project-based delivery and customized costs.
• Observation Points: Growth of order and contract value in vertical industries, cross-industry reuse capability, marginal cost improvement of software-hardware integrated solutions.

• Advantages: Aligns with current large model training demands, with high bargaining space and customer stickiness; high fit with computing power centers, internet AI labs, and operator intelligent computing projects.
• Challenges: High R&D threshold and capital consumption, relying on collaboration of manufacturing and interconnection processes (packaging, bandwidth, etc.); if scale delivery is not formed, valuation support is unstable.
• Observation Points: Measured benchmark of training throughput and interconnection topology, ability to stably deliver and support 100-billion parameter model training, software stack and model ecosystem adaptation progress.

• Advantages: Deep binding with cloud service providers, conducive to rapid large-scale deployment of products; inference side has higher energy efficiency and cost sensitivity, with strong domestic substitution demand.
• Challenges: Risk of over-reliance on a single cloud ecosystem; many competitors in inference (including general-purpose GPU and other DSA solutions), requiring continuous proof of TCO advantages.
• Observation Points: Joint product form and commercialization rhythm with Tencent Cloud, comparison of inference throughput and unit cost, support breadth for mainstream models and frameworks.

• Listed Company Examples: Moore Threads and Muxi have seen significant market value increases after listing, indicating that the market has given high expectations and premiums to the “domestic substitution + AI computing power” track [1, 2].
• Industry Comparison and Implied PS: If the market uses Cambricon (which has achieved profitability and has a market value of over RMB 560 billion) as a partial reference [1], valuations for new players often imply higher revenue growth expectations and “track + scarcity” premiums.
• Risk Reminder: High valuations combined with current general unprofitability or late profit inflection points mean that “valuation support” depends more on sustained high revenue growth and fulfillment quality, not just themes.

• Policy and Capital Support: National-level promotion of technological self-reliance and chip industry capital support (e.g., media reports consider new capital support scale may reach RMB 500 billion level), forming continuous policy and capital endorsement for domestic GPU [2].
• Supply Chain Security and Compliance: Geopolitical and technical restrictions increase the demand for independent controllability, especially in key infrastructure, intelligent computing centers, and government-enterprise informatization.
• Scenario Demand: Domestic large models and industry AI applications continue to expand, and training and inference demands form a “just-needed window” for domestic computing power supply.
• Ecosystem and Customer Base: Whether to form stable large clients and ecosystem partners (cloud vendors, operators, vertical industry leading clients) is a key variable for substitution implementation.

• Performance and Ecosystem: Compared with overseas leaders, there are still gaps in peak performance, interconnection bandwidth, software stack maturity, and model support breadth, which need to be continuously narrowed.
• Supply Chain and Manufacturing: Access to advanced manufacturing nodes, yield rate, and delivery rhythm are still hard constraints.
• Competitive Landscape: Multi-line competition from overseas and domestic players may compress profit margins through price and performance wars.
• Commercialization Rhythm: There are obvious fulfillment lags and uncertainties between orders and revenue, and between revenue and profit.

• Revenue and Cash Flow: Current revenue scale, growth rate and quality (whether it is sustainable product revenue), whether the loss convergence path is clear (e.g., Muxi’s production and sales rate exceeds 100% and expects break-even in 2026 [2], Moore Threads expects profitability in 2027 [1]).
• Customer Structure: Whether the number and single-client value of government-enterprise, operators, cloud vendors, leading internet clients, and vertical industry clients continue to increase.
• Product Capability and Ecosystem: Measured performance and cost in training/inference/rendering scenarios, software stack and framework adaptation progress (e.g., compatibility and support for mainstream models and inference frameworks).
• Supply Chain and Delivery: Mass production delivery capability, yield rate, availability of process technology, and cost optimization path.
• Technical Route Adaptability: Whether the choice of general-purpose and domain-specific scenarios matches its target market and customer needs (e.g., DSA penetration in specific scenarios, breadth and depth balance of general-purpose GPU).
• Competition and Pattern: Market share and positioning in respective segmented tracks, and competitive situation with peers and overseas competitors.

• Significant Technical Route Differences: Moore Threads emphasizes full-functionality and multi-scenario coverage; Muxi accelerates commercialization while advancing general-purpose and vertical toB; Biren targets high-end training and large model scenarios; Suiyuan binds cloud services and deep dives into inference. DSA/TPU and other domain-specific paths have better energy efficiency but higher ecosystem thresholds; GPU/GPGPU have stronger generality but fierce competition [1, 2].
• Different Commercialization Rhythms: Muxi’s production and sales rate exceeds 100% and plans to break even in 2026; Moore Threads expects profitability in 2027; Biren and Suiyuan are still in the listing process, with profit inflection points depending on orders and delivery [1, 2].
• Valuation Logic Highly Dependent on Domestic Substitution Realization: Current high valuations mainly reflect track dividends and scarcity premiums formed by “policy + capital + demand”; long-term support depends on revenue growth quality, customer structure and ecosystem construction, supply chain stability, and technical iteration capability.
• Investment Suggestion Framework:

1. Short-Term: Focus on order and revenue fulfillment rhythm, production and sales rate and gross margin changes, and progress of large clients and benchmark projects of listed/拟 listed companies.
2. Medium-Term: Focus on technical route and scenario matching degree (e.g., Biren in large model training scenarios, Muxi in vertical toB, Suiyuan in inference + cloud binding), and maturity of software stack and model ecosystem.
3. Risk Management: Diversify single technical route and single client dependence, attach importance to manufacturing and supply chain constraints, and be alert to valuation pullback risks of “theme first, performance lag”.

• Compare product specifications and measured benchmarks of the four leading players (conduct comparative analysis under available data).
• Draw a timeline and key milestones of “revenue/profit/market share” in segmented tracks.
• Establish a valuation sensitivity matrix (impact of variables such as revenue growth rate, gross margin, R&D expense rate on PS/PE/DCF).
• Combine macro policy and industry capital landing rhythm to quantify domestic substitution space and penetration path.

[1] Yahoo Hong Kong Finance - China’s Computing Power Enters Capital Harvest Period! GPU Four Leading Players Rush for IPO (https://hk.finance.yahoo.com/news/中國算力進入資本收割期-gpu四小龍衝刺上市-沐曦股份17日登科創版)
[2] Yahoo Hong Kong Finance - Muxi Shares Debut on STAR Market; AMD Halo and GPU Four Leading Players Label Expected to Open Valuation Space for the Company (https://hk.finance.yahoo.com/news/沐曦股份亮相科創板-amd光環及gpu四小龍標籤料為公司打開估值空間)
[3] Yahoo Hong Kong Finance - Reportedly, Cambricon Plans to Double Chip Production Next Year to Replace NVIDIA’s Market Share in China (https://hk.finance.yahoo.com/news/據報寒武紀計劃明年晶片產量提高兩倍-取代英偉達在華市場份額)
[4] Wall Street Journal - Biren Technology, One of China’s GPU “Four Leading Players”, Plans to Raise Up to $623 Million via Hong Kong IPO (https://cn.wsj.com/articles/壁仞科技計劃通過香港ipo籌資至多6-23億美元-4b717986)
[5] Forbes - Former Head Of Nvidia In China Is A Billionaire Thanks To Moore Threads IPO (https://www.forbes.com/sites/ywang/2025/12/05/chinese-challenger-to-nvidia-mints-a-new-billionaire-as-its-stock-soars-over-420-in-trading-debut/)
[6] Yahoo Hong Kong Finance - ASIC Scaler Global Leader Xihua Technology Rushes for Hong Kong IPO! Tsinghua Scholar Chen Xi Builds China’s Chip Dark Horse (https://hk.finance.yahoo.com/news/asic-scaler全球第-曦華科技衝刺港交所ipo-清華學霸陳曦打造中國晶片黑馬)

(Note: The above evaluation is based on user-provided context and online search results; no speculation is made on undisclosed or unvalidated technical parameters and financial details; judgments involving specific companies’ future performance and valuations should be based on the latest public disclosures and professional investment research reports.)