The "2026 Zhitong Finance Summer Roadshow Conference," hosted by Zhitong Finance, was successfully held in Qianhai, Shenzhen, on May 28.
This event brought together over 40 popular Hong Kong-listed companies, covering core sectors such as AI, hard tech, new energy, biopharma, consumer goods, and smart manufacturing.
During this roadshow, He Feihong, CFO of Haizhi Tech (02706), engaged in in-depth discussions with investors on the company's core technologies, product portfolio, industry applications, financial performance, and future strategy, centering on the theme "Graph-Model Fusion Technology Drives the Development of Industrial-Grade AI Agents."
The following is the transcript of the Haizhi Tech roadshow session.
**Haizhi Tech Group (02706)**
Haizhi Tech is the first company in China to apply graph-model fusion technology to industrial-grade AI solutions, listing on the main board of the Hong Kong Stock Exchange on February 13, 2026, and dubbed by the capital markets as the "first AI Hallucination-Reduction/Harness stock."
The company's core positioning is an industrial-grade Agent and AI operating system centered on graph-model fusion technology, addressing the core pain points of large language model deployment in industrial settings—hallucinations and lack of control—enabling a leap from "probabilistic intelligence" to "deterministic execution."
Reflecting on the company's 13-year development, it has undergone three key technological transitions.
From 2013 to 2018, the technology pioneering phase: as a leader in enterprise knowledge graphs in China, it refined the BDP data visualization platform and the Atlas knowledge graph platform, focusing on high-frequency, high-value scenarios in government affairs and finance.
From 2019 to 2022, the self-reliance phase: to overcome performance bottlenecks of foreign graph databases, the company independently developed the distributed cloud-native graph database AtlasGraph, which broke a world record in the 2023 LDBC international benchmark test with a 45% advantage, becoming one of the world's fastest graph databases.
From 2023 to the present, the agent implementation phase: with the rise of large language models, the company deeply integrated its decade-long graph technology with LLMs, launching the Atlas industrial-grade agent, establishing itself as a leader in graph-model fusion technology in China.
To date, the company has cumulatively served over 400 top-tier enterprise clients, covering five core industries: finance, energy, government affairs, telecommunications, and high-end manufacturing.
From 2023 to 2025, the Atlas agent business revenue achieved explosive growth of 1637% and achieved positive adjusted net profit in 2024.
What is graph-model fusion?
Why does the company persist with the graph technology path in the era of large language models?
First, an industry consensus: Agent = Model + Harness.
A useful industrial-grade agent requires not only powerful LLM capabilities but also a comprehensive harnessing system.
LLMs are inherently probabilistic, leading to issues like hallucinations, lack of control, and lack of auditability, making them unsuitable for direct deployment in industrial scenarios demanding extremely high reliability.
Over the past three years, AI engineering has undergone three iterations.
Prompt engineering: relies on writing prompts to elicit model capabilities, but results are unstable and cannot handle complex tasks.
Context engineering: uses RAG (Retrieval-Augmented Generation) to feed data to models, but it only provides static knowledge and cannot understand business logic and relationships.
Harness engineering: treats LLMs as uncontrollable engines, using external systems to establish rules, constraints, processes, and safety guardrails, which is precisely Haizhi's core competency.
The company's Harness system is built on Ontology 2.0.
Unlike the static ontologies commonly used in the industry, Haizhi's ontology is a dynamic execution kernel: it is not just a "map" of knowledge but an "autonomous driving system" for AI, supporting read-write-execute full-chain operations, capable of directly driving tool calls and business processes.
It achieves automatic extraction and logical self-evolution through graph-model fusion technology, significantly reducing construction costs.
Simply put, traditional RAG gives AI a manual, while Haizhi's ontology gives AI a master with ten years of industry experience, telling it what can be done, what cannot, and how to do it correctly.
For example, in the power grid industry, a standard LLM analyzing transformer failure causes might answer: "Over the past two months, 110kV transformer failures increased by 18%; average repair time increased by 6 hours; spare parts procurement costs rose by 12%."
Haizhi's agent answers: "This cost increase is not due to isolated failures but a spreading 'family defect.' The seven recently failed transformers all belong to the same manufacturer, same batch, same winding process family, and are ontologically linked to defective equipment in the Northeast region last year.
Current spare parts inventory only covers the next two weeks of demand, with the core supplier lead time at 45 days. It is recommended to immediately initiate a 'family defect equipment linkage warning' and simultaneously trigger material procurement, operation inspection, and inventory transfer processes."
Beyond the energy sector, the company has numerous mature applications in public service and safety, finance, government affairs, and other fields.
For instance, the company's agent can automatically generate case investigation DAG charts, investigate social relationships, and analyze activity trajectories, reducing investigation work that originally took two weeks to just a few hours.
In 2025, the company achieved total revenue of RMB 621 million, a year-on-year increase of 23.4%; the total number of clients reached 212, a year-on-year increase of 24%.
Regarding business structure, the company's second growth curve has taken shape: Atlas agent business revenue was RMB 145.7 million, a year-on-year increase of 68.2%, with its revenue share increasing from 2.4% in 2023 to 23.5%; traditional Atlas graph solution revenue was RMB 475.3 million, a year-on-year increase of 14.1%, maintaining steady growth.
In terms of profitability, the overall gross margin increased from 30.9% in 2022 to 43.3% in 2025; the Atlas agent business gross margin reached 53.2%, significantly higher than the traditional business.
Adjusted net profit was RMB 16.9 million in 2024, successfully turning a profit; adjusted net profit was RMB 24.1 million in 2025, with a net margin of 3.9%.
Client stickiness is one of the company's core moats.
The client repurchase rate exceeds 70%, and the cumulative procurement amount of the top 20 clients over a 4-5 year lifecycle is approximately 20-30 times their initial purchase.
State-owned enterprises, central enterprises, and the government are currently the largest payers for AI industry implementation, with clear budgets and strong policy drivers; only by serving industry leaders can the most complete and complex industry ontologies be built, forming barriers for downward penetration.
In the future, as the company expands into medium and large enterprise clients, it will gradually support usage-based payment models.
Looking ahead, the company has three core growth engines.
Deepening within industries: relying on the "graph reuse" flywheel, expanding from single departments to entire organizations, increasing the lifetime value of existing clients.
Cross-industry replication: expanding from finance and government affairs to energy, telecommunications, manufacturing, and other industries, with over 60% of the underlying architecture reusable, leading to faster deployment and lower costs in new industries.
Product standardization: packaging mature industry ontologies and agents into standardized products to reduce delivery costs and improve gross margins.
Regarding growth expectations, the company previously provided analysts with guidance of 35%-40% annual revenue growth, but actual growth this year will be significantly higher than this figure.
The explosive popularity of tools like OpenClaw has suddenly made the entire market realize that AI can not only chat but also perform real work.
The company's sales pipeline is rapidly expanding, and growth in the coming years will accelerate.
In the long term, Haizhi, whether from a business model or underlying technology perspective, is the Chinese company closest to Palantir, building an industrial-grade AI operating system.
As LLM-related technologies mature, the value of the application layer will explode, and Haizhi, as the core infrastructure connecting LLMs and industries, will fully benefit from this wave.
**Q&A Session**
**Q: Many B2B companies face client price pressure. What is the company's pricing power in the industrial chain?**
**A:** Pricing power essentially depends on your irreplaceability.
In China today, there are very few suppliers that can provide graph-model fusion and industrial-grade agent services at the company's level.
This is why the company's comprehensive gross margin can remain above 40%, with the agent business gross margin exceeding 50%.
More importantly, the company's service targets are shifting from IT departments to business departments and decision-makers.
Previously, when IT departments purchased systems, they focused on cost; now, when business departments purchase agents, they focus on what problems they can solve and how much value they can create.
For example, if the company helps a power grid discover equipment defects in advance and avoid a large-scale blackout, the value created could be millions or even tens of millions, and clients are willing to pay a premium for that value.
**Q: When will the company achieve true breakeven? What are the future R&D investment plans?**
**A:** The company achieved positive adjusted net profit in 2024, and adjusted net profit continued to grow in 2025.
The total net loss shown on the financial statements is mainly due to the fair value adjustment of convertible preferred shares; the higher the company's market capitalization, the larger this adjustment amount, which is unrelated to actual operational performance.
Regarding R&D investment, the company believes now is not the time to pursue large profits.
There are still many hard technologies to break through in areas like graph-model fusion, automated graph construction, and multi-model adaptation, and the company will continue to increase R&D investment.
In fact, the company's true gross margin is stronger than what the statements show.
This is because the company adopts a model similar to Palantir's FDE (Forward-Deployed Engineering) model.
When developing the first industry-leading client, many front-end engineering costs are accounted for as costs, but they are actually building reusable industry ontology capabilities, which should be considered R&D investment.
If this portion is adjusted, the company's gross margin would be even higher.
Comments