Company Overview
Founded in 2018 and headquartered in Zhangjiang, Shanghai, Tianshu Zhixin is one of the earliest AI chip design companies in China to independently develop general-purpose GPUs. Its core team members come from NVIDIA, AMD, Huawei HiSilicon, and other top-tier chip companies, with full-stack capabilities from architecture design to mass production and delivery. The company has so far mass-produced two flagship products: Tiangai 100 (a 7nm general-purpose training GPU) and Zhikai 100 (an inference accelerator card), and has achieved large-scale commercial deployment in key industries such as telecom operators, supercomputing centers, and internet companies. In early 2025, the company took the lead in fully adapting DeepSeek, the domestically developed open-source large model, marking a new stage of moving from "having usable chips" to "model-chip synergy."
Tianshu Zhixin's business centers on general-purpose GPU chips, with current revenue primarily derived from two major product lines:
Tian Gai 100 is the company's first 7nm general-purpose GPU, based on a self-developed computing architecture, supporting FP32, FP16, INT8 and other precision modes. Its single-card FP32 computing power reaches 14 TFLOPS (double precision), and FP16 computing power exceeds 100 TFLOPS. It is mainly used in AI training scenarios, such as large model pre-training, recommendation systems, scientific computing, etc. Key customers include the three major telecom operators (China Mobile, China Telecom, China Unicom), national supercomputing centers, and some major internet companies.
Zhi Kai 100 is an accelerator card optimized for large model inference scenarios, adopting a 7nm process, supporting INT4/INT8 quantized inference, with single-card INT8 computing power up to 200 TOPS+ and inference latency below 10ms. It is mainly applied in cloud inference, edge AI, intelligent customer service, and other scenarios. In 2024, with the explosive growth of DeepSeek inference demand, shipments of Zhi Kai 100 doubled year-over-year.
| Product Line | Revenue Share (2024E) | Core Customers | Gross Margin (Estimate) |
|---|---|---|---|
| Tian Gai 100 (Training) | ~70% | Telecom operators, supercomputing, internet | ~45% |
| Zhi Kai 100 (Inference) | ~25% | Telecom operators, internet, government | ~40% |
| Others (IP licensing, customization) | ~5% | Research institutions | ~60% |
Technical Moat
Moat 1: Self-developed General-Purpose GPU Architecture Balancing Performance and Versatility
Iluvatar CoreX employs a fully self-developed general-purpose GPU architecture, the "Tianqe Architecture," with no dependence on any overseas licensing. The architecture benchmarks against NVIDIA A100-class levels in SIMT execution model, memory hierarchy, and on-chip interconnect, and supports rapid CUDA code migration through a compatibility layer. Currently, the Tianqe 100 achieves 85%-90% of NVIDIA V100's training speed on mainstream AI models, delivering significant cost-performance advantages in domestic substitution scenarios.
Moat 2: Deep Integration into the Domestic Large Model Ecosystem with Software-Hardware Synergy
The company has fully adapted to leading domestic large models including DeepSeek, Qwen, Baichuan, and ERNIE Bot, providing a complete software stack spanning underlying drivers, operator libraries (with 300+ accumulated high-performance operators), and inference acceleration toolchains. In particular, after announcing DeepSeek adaptation in early 2025, the company became the only vendor-independent domestic GPU enterprise simultaneously supporting the full DeepSeek-V3/R1 series, creating a "model lock-in" effect.
Moat 3: Government and Enterprise Customer Relationship Barrier with High Entry Threshold
The company's products have been included in the centralized procurement lists of the three major telecom operators (securing approximately 30,000 training card orders from China Mobile in 2024) and have repeatedly won bids for national intelligent computing center projects. Government and enterprise customers impose extremely high requirements on chip stability, service responsiveness, and Xinchuang (IT application innovation) adaptation, with enormous replacement costs that create natural customer stickiness.
| Dimension | Data |
|---|---|
| 2024 China-made AI training chip market share | ~8% (third-party consulting estimate) |
| 2024 China-made AI inference chip market share | ~6% |
| Industry ranking (domestic GPU) | 4th (top 3: Huawei Ascend, Hygon, Cambricon) |
| Main competitors | Huawei Ascend (~50% share), Hygon Information (~20%), Cambricon (~12%), Moore Threads (~8%) |
| Downstream customer distribution | Telecom operators (40%), supercomputing/AI computing centers (30%), internet companies (20%), government & enterprise (10%) |
Competitive landscape analysis: The domestic GPU market is currently in a "one dominant player with multiple strong contenders" phase, with Huawei Ascend leading the market through its self-developed Da Vinci architecture and MindSpore ecosystem. Tianshu Zhixin's advantage lies in its higher compatibility with the CUDA ecosystem (non-proprietary programming framework), as well as superior power efficiency and cost-effectiveness in inference scenarios compared to competitors of the same generation. Following DeepSeek adaptation in 2025, it is expected to rapidly penetrate the internet and small-to-medium model customer segments.
Financials & Growth
| Metric | Data |
|---|---|
| Revenue (2024E) | Approximately RMB 1.5–2.0 billion (shipments driven by IT innovation/domestic substitution procurements) |
| Gross Margin (Blended) | ~43% (higher for training cards, slightly lower for inference cards) |
| Net Margin | ~5% (heavy R&D investment, accounting for ~30% of total revenue) |
| Core Growth Logic | 1) Accelerated domestic substitution, with surging order volumes from telecom operators and government IT innovation projects; 2) Explosive inference demand from open-source large models such as DeepSeek; 3) Performance leap driven by mass production of the Tiangai 200 (5nm) in 2025 |
Key Growth Drivers: In 2024, the company's shipment volume grew over 100% year-over-year, primarily fueled by the construction of intelligent computing centers by telecom operators. In 2025, catalyzed by the DeepSeek ecosystem, inference card shipments are expected to achieve 200% growth. The company has launched a new round of Pre-IPO financing at a valuation of approximately RMB 15 billion, with plans to list on the STAR Market in 2026.
Main Risks:
- Technology Iteration Risk: Next-generation products such as NVIDIA's B200 hold a substantial performance lead. If the company's Tiangai 200 architecture is delayed or underperforms expectations, the gap with overseas products could widen further.
- Ecosystem Dependency Risk: The company is currently deeply adapted to DeepSeek. If the model ecosystem shifts to other frameworks (e.g., MindSpore), the company will need to continuously invest in adaptation costs. Additionally, the CUDA compatibility layer still incurs a 10%~15% performance loss in complex scenarios.
- Supply Chain Risk: TSMC's foundry services are subject to geopolitical influence. If geopolitical conflicts escalate, 7nm/5nm capacity could be constrained, affecting shipment and delivery.
- Intensified Market Competition: New rivals such as Hygon, Moore Threads, and SpacemiT are accelerating their expansion, and price wars may compress gross margins.
Core Investment Logic / Industry Value Summary:
Iluvatar CoreX is one of the few domestic general-purpose GPU vendors with full-scenario adaptation capabilities for large models. In the short term, it enjoys the dual dividends of Xinchuang (domestic IT innovation) and DeepSeek; in the medium to long term, the competitiveness of the Tiangai 200 remains to be validated. During the transition of domestic computing power from "usable" to "easy to use," the company is well-positioned to maintain its presence by leveraging its government and enterprise customer moat and first-mover advantages in the open-source model ecosystem.