Over the past few years, the AI industry has experienced one technological leap after another.

From large models to multimodality, from training to inference, from Copilot to Agent, technological innovation has continuously reshaped people's imagination of artificial intelligence. At the same time, computing power, models, and applications have also become the most concentrated keywords in industry discussions.

But as AI truly moves from technological enthusiasm into the deep water zone of industry, an increasingly realistic question begins to surface:

What ultimately determines whether AI can truly run in enterprises?

The answer is becoming clear.

Models determine what AI "can do," computing power determines how fast AI "can run," and data infrastructure determines whether AI "can truly enter business."

This is also why, as AI enters the stage of large-scale implementation, data is gradually moving from being a "raw material" in the AI industry chain to becoming the key foundation that determines the effectiveness and scalability of AI applications.

And this is precisely the direction that KeenData has persistently adhered to over the past years.

Rather than chasing every concept, it has focused on the most fundamental and long-term issue of AI industrialization: continuously building enterprises' AI data infrastructure.

Looking back today, the persistence of the past is gradually converging with the new demands of industrial development.

In September 2026, the Bund Conference released a set of widely cited data: IDC predicts that by 2035, China's AI data infrastructure market will reach US$154.83 billion, more than a tenfold increase from 2025, with a compound annual growth rate of 37.7% from 2025 to 2030.

The market is new, but the problem is not.

A survey by Dun & Bradstreet shows that 97% of enterprises are advancing AI projects, but only 5% believe their data is ready to support large-scale AI implementation. Gartner has also pointed out that as many as 50% of generative AI projects are abandoned at the proof-of-concept stage, with insufficient data quality being the primary reason.

These numbers reveal a simple truth: the upper limit of AI is not determined by models, but by data.

This is not a judgment chasing hot topics. As early as 2019, when the company was founded, when industry attention was still focused on algorithm competitions and computing power arms races, KeenData chose a less "sexy" direction—building AI data infrastructure. Over seven years, the AI-in-Lakehouse intelligent driving architecture has continued to iterate and evolve, with a core code self-development rate of 97% and more than 200 technology invention patents accumulated.

There are no shortcuts on this road, but when an industry moves from frenzy to rationality, true value begins to emerge.

When AI moves from demonstration to production, from "being able to converse" to "being able to make decisions," what enterprises face is no longer a question of technology selection, but a question of engineering implementation.

KeenData summarizes the challenges facing large-scale AI implementation into four structural constraints: data silos, data quality, data security, and the engineering gap. None of them can be solved by a single model or algorithm; they require long-term accumulation of data governance, engineering capabilities, and systematic platforms.

This is precisely the path KeenData has chosen.

"Deep integration of data governance and data engineering" and "centralized management, decentralized empowerment"—this hybrid data intelligence implementation system has been gradually formed by KeenData in the process of serving industries such as energy, finance, manufacturing, and government affairs. It does not rely on a breakthrough in any single technology, but comes from continuous understanding and repeated verification of data scenarios across different industries.

KeenData continues to build an intelligent foundation for the Agentic AI era—the KeenData Agentic Lakehouse Platform. Its core consists of the KeenData Lakehouse multimodal lakehouse management platform and the Keen Agentic OS intelligent agent development operating system.

Among them, KeenData Lakehouse is responsible for solving the problem of "where data comes from, how it is governed, and how it becomes data usable by AI"; Keen Agentic OS further solves the problem of "how AI understands data, invokes data, and completes tasks based on data."

Together, they make data and Agents no longer two separate systems, but form a complete closed loop from data entry, governance and processing, to AI consumption, intelligent action, and then business feedback.

Things that were once regarded as underlying foundational capabilities are now becoming key capabilities for large-scale AI implementation.

The best way to judge whether a direction is right is not to look at what it says, but at what it has accomplished.

To date, KeenData has served more than 300 large organizations, covering core industries such as government affairs, energy, finance, and manufacturing.

In the energy sector, a large energy state-owned enterprise, through the AI data foundation built by KeenData, integrated 61 core systems and 1.2PB of data, reducing the efficiency of viewing business analysis reports from 1 week to 4 hours, and provided compliant, auditable data support for emerging businesses such as green electricity trading and carbon footprint tracking.

In the financial sector, the real-time data foundation built by KeenData for a large central state-owned bank supports millisecond-level anti-fraud and intelligent risk control scenarios.

In the government affairs sector, KeenData has deeply participated in the construction of trusted data spaces and pilot demonstration zones in multiple key cities including Beijing, Hangzhou, Suzhou, and Changsha, supporting the access of more than 1,000 data entities and the release of more than 2,000 data products.

In the manufacturing sector, KeenData has connected the "vehicle-cloud" data link, achieving full lifecycle connectivity of vehicle identification numbers and supporting AI applications such as intelligent customer service and vehicle profiling.

The common point of these cases is that they are not proof-of-concept, but systems truly running in production environments. What they solve is not the question of "whether AI can do it," but the question of "whether AI can do it stably, reliably, and at scale."

In the stage of faith, people believe that models can solve everything; in the stage of engineering, people begin to realize that what truly determines success or failure is often the foundational work that is not in the spotlight—how data is governed, how quality is ensured, how systems collaborate, and how engineering is scaled.

KeenData has chosen a less bustling position, but it happens to be the position that most needs to be filled as the AI industry moves from the first half to the second half.

Yu Yang said: "Our goal is to advance the construction of AI data infrastructure and empower thousands of industries to achieve digital and intelligent transformation."

This sentence is not a slogan, but something a group of people firmly believed in seven years ago.

When the spotlight finally turns to data, those who have been standing there all along will naturally be seen.