Author｜Sang Mingqiang

The vision of the next-generation data intelligence infrastructure, including upgrades to existing IT systems and intelligent decision support driven by data and computing power, has been a hot topic in the industry.

In the newly released "China Data Analysis and AI Technology Maturity Curve", Gartner believes that the "Data Platform" is on the verge of disappearance, to be replaced by "Digital Intelligence Infrastructure". This shift stems precisely from the fundamental differences between digital intelligence infrastructure and the data platform.

Unlike the data platform, which emphasizes "centralized management", digital intelligence infrastructure is more convinced of the power of ecosystems.

It is important to note that businesses today are more pragmatic than ever, especially for medium to large-sized enterprises. Rather than blindly chasing technological trends, they are eager to build a flexible and sustainable data analysis and AI platform to optimize business processes and aid decision-making, rather than simply moving data online for a dashboard display.

Reflecting this trend, OpenAI recently acquired the data warehouse startup Rockset for $3.6 billion, and there have been intense moves by Databricks and Snowflake. The former acquired MosaicML, which focuses on improving AI model training, for $1.3 billion last year, while the latter was rumored to be in talks to acquire Reka AI, an AI startup, for over $1 billion. These signs indicate that the advancement of both the data layer and model layer has become an industry consensus, with the To B market poised to become the preferred landing spot for large models due to its ability to quickly form a value loop.

Similar upheavals are also occurring domestically. Specifically, unicorns in the Data+AI field, led by Dipu Technology, are undergoing iterative upgrades to their existing product and technology architectures based on AI and large model capabilities, evolving from early data platforms to unified lakehouse solutions, and finally to digital intelligence infrastructure for the To B sector in conjunction with large model technology.

However, due to a mismatch between investments and business value, as well as fluctuations in GPU market prices, the large model industry is currently experiencing a steep decline.

But this is not necessarily a bad thing. "Large models are gradually changing the rules of the game in the To B market," said Zhao Jiehui, Chairman and CEO of Dipu Technology. In his view, "by integrating large models with cloud ecosystems, enterprises and organizations can better integrate structured and unstructured data, leading to a qualitative leap in data platform capabilities."

According to Zhao Jiehui's logic, there are two main disconnects in the current value scenarios for large models in enterprises. On the one hand, while many data platforms collect vast amounts of real-time data, they lack flexible business context logic. On the other hand, while large models are well-known for their understanding capabilities, they lack real-time business details reflected in real-time business data.

To some extent, this also explains why the large model sector has been a mixed bag over the past two years. Despite significant investments and the training of more parameters with greater computing power, the ROI in value scenario implementations has been alarmingly low. What the market lacks are players who can truly integrate the data layer, application layer, and model layer into one, providing comprehensive data solutions.

In other words, setting aside the C-end market, the key to rapid adoption of large model technology in the B-end lies in forming an overall "Data+AI" infrastructure and innovative applications that enable large models to form deeper real-time business reasoning capabilities based on enterprises' real-time and unstructured data-driven business logic, thereby building intelligent applications across various sectors.

This is also the reason why Dipu Technology stands out among domestic data intelligence companies. Those familiar with the company know that as a newly designated national-level "little giant" enterprise in 2024, Dipu Technology is known for its low profile and non-conformity. The core members of the team include former Huawei veterans, and the company has inherited Huawei's down-to-earth approach to product development and technical research.

Similar to Databricks' growth trajectory, Dipu Technology has consistently focused on the "Data+AI" strategy since its inception. From initially positioning itself as a bottom-level data platform, to developing the real-time intelligent lakehouse platform FastData, and finally to exploring DeepexiOS, the only enterprise AI platform currently capable of practical implementation in China, Dipu Technology has precisely stepped onto the wave of digital intelligence infrastructure development at every step.

Take DeepexiOS, Dipu Technology's latest flagship product launched at this year's China International Fair for Trade in Services, which comprises three core components: the FastAGI enterprise large model service platform, the FastData enterprise converged data platform, and the Fast5000E enterprise computing platform. Based on these, DeepexiGenAI, Dipu Technology's generative AI application for enterprises, has already been implemented in core areas such as supply chain AI rapid response, process AI formulation, AI-assisted engineering design, and data AI analysis, and has passed the Ministry of Industry and Information Technology's China Academy of Information and Communications Technology (CAICT) model capability standard compliance level 4+ certification.

"Frankly speaking, customers are not concerned about how many parameters a large model has been trained on or how high its accuracy is. Instead, they are more interested in the practical business value that this technology can bring," said Zhao Jiehui.

Taking Belle as an example, as an established giant in China's footwear and apparel retail industry, it is most concerned with achieving supply chain AI rapid response in its digital construction. With the DeepexiOS enterprise AI platform, it can analyze real-time business data and make scientific replenishment and production decisions based on historical data and current market conditions.

In fact, the combination of data governance and model capabilities brings about a level of disruption far more profound than digitization itself in previous years.

If the main task of the previous era was to build data infrastructure and assetize data elements, the primary mission of "Data+AI" is to decouple and reconstruct the complex scenarios of various enterprises, combined with the inherent ease of use of large model technology, which is likely to spark a new wave of digital intelligence infrastructure in the market.

This is not a baseless claim.

According to the "IDC MarketScape: China Real-Time Lakehouse Market Vendor Assessment, 2024" report, the proportion of enterprises choosing external partnerships to build data management services will rapidly increase from 58% to 85% over the next 12 months. In other words, as data volumes grow rapidly and demands for data management escalate, along with increasing technical architecture complexity and standalone development costs, superior data governance capabilities will be the key driver for the implementation of the large model industry.

Let's take the manufacturing industry as an example. As is well known, China is the only country in the world with all industrial categories in the United Nations Industrial Classification, encompassing 39 major industries, 191 medium-sized industries, and 525 minor industries. With the most comprehensive industrial system and supporting capabilities globally, its high scene complexity means that its digital transformation journey will not be easy.

However, just as a coin has two sides, "focusing on digital intelligence infrastructure in manufacturing is like opening pearls in the deep sea; only by breaking open the hard shells can treasures be found," said Zhao Jiehui, citing Dipu Technology's practical experience in a certain equipment manufacturing field. The entire process is divided into two steps: first, establish the FastData enterprise converged data platform, and then delve into the various data generated during the production process based on the FastAGI enterprise large model service platform to refine and adjust the existing process flow, generating more efficient and precise "process cards."

The benefits of this approach are evident. On the one hand, it ensures quality control during the product design stage, avoiding later inspections and rework. On the other hand, through standardized and modular design, it shortens preparation time for each production batch, significantly improving production efficiency and ultimately ensuring the consistency and stability of product quality.

It is worth mentioning that as more value scenarios are implemented, the scale effect of digital intelligence infrastructure will follow.

"China is not short of data intelligence and large model vendors. Many people complain about the industry's downturn, but the crux of the problem is the desire to make quick money. For a new technology, in addition to continuous learning and iteration, we must understand what value it can bring to customers," Zhao Jiehui admitted candidly.

To some extent, while there is still considerable uncertainty about the potential of generative AI, one thing is undeniable: large models are not traditional AI bubbles, and the moon is not necessarily rounder abroad. The growth in their penetration rate is just a matter of time, much like the central platform boom in the To B sector years ago. While some aspects of it may have been discarded by the market, the good aspects will still be retained, even delivered to customers in a new way. Ultimately, the success of digital intelligence infrastructure must be judged by its value itself.