Slow business, fast iteration: Diptech's 7-year anti-consensus "breakthrough"

If a company spends seven years doing only one thing, everyone will think it is very focused; if it "misses" the Internet+, data middle platform and other opportunities in the past seven years, everyone will think it is a bit stupid; but if this company later becomes the number one player in the domestic enterprise-level large model, it is Diptech Technology.

The entrepreneurial story of Diptech Technology began in 2018, which was almost the most "expanded" year for China's software industry.

Under the acceleration of the capital market, a number of start-up companies with data middle platform as their selling point were born. The direct consequence of this vicious competition aimed at grabbing market share was the extreme distortion of PMF (Product-Market-Fit), which ultimately led to the failure of these companies in digitalization - they only had the appearance of scale growth, but the internal profit problems were not solved for a long time.

Similar dramas continued in the era of large models.

The generative AI revolution launched by OpenAI has excited countless entrepreneurs, but it has also revealed deep contradictions in the enterprise market: although general large models can generate text, pictures and even videos, they find it difficult to penetrate the "black box" of corporate data. Although Meta's open source Llama2 has lowered the technical threshold and DeepSeek has achieved top reasoning results with a low-cost architecture, the engineering capabilities of the industry know-how are still scarce.

Furthermore, the serious disconnect between AI benchmark settings and the real world has caused the industry to question whether large models have been deified. The most typical example is that intelligent systems in real-world scenarios are continuously interactive and have memories, but the existing evaluation settings ignore these key features, resulting in an increase in intelligence but no change in utility.

" Being good at To C does not mean you can do To B. There is a natural gap between the two ." Zhao Jiehui, who calls himself a newcomer in the software industry, has a very special background. Before starting his own business, he was a key member of Huawei's core router team. He experienced the golden age of high gross profit in the communications industry. After switching to enterprise services, he commented on the software industry: " Traditional enterprise IT is essentially a database plus an interface. The technical content is not high, but with the advent of large-model AI, the technical threshold of the industry will be raised ."

Zhao Jiehui has a very advanced understanding of advanced technologies. Almost as soon as ChatGPT became popular, he realized that this would be a complete subversion of the traditional software industry. When everyone around was focused on the industry's big model, he took a different approach and led the team to focus on the deep decoupling of industry know-how and the engineering capabilities of data assets.

In his opinion, unlike the past when the AI ​​community focused too much on how to train stronger models, the logic of enterprise-level large models lies in how to focus on real-world utility based on open source foundations. The challenge behind this is to define "what to do" and "how to measure progress", which means that the research focus should shift to problem setting and evaluation rather than the model itself.

From this perspective, how to make a company's data and knowledge truly "live" has been the only main thread that Diptech has used to break through in the past seven years.

So far, Diptech has completed eight rounds of financing and has been continuously bet on by many star institutions including Hillhouse Capital and IDG. According to Frost & Sullivan data, Diptech is the No. 1 professional provider of large-model artificial intelligence application solutions in China's enterprise-level large-model artificial intelligence application solution market in terms of revenue in 2024. Combined with the fact that it has just submitted its prospectus to the Hong Kong Stock Exchange, it is very likely to become the first domestic stock in this direction .

The starting point of Dipu Technology is not sexy. Most of the core members come from Huawei, Alibaba, and IBM. They are a group of "technical craftsmen" with more than ten years of experience. They advocate thinking from the first principles, which is reflected in product polishing. "What actual value can it bring to customers" is the most frequently used hot word in their internal discussions.

The first time I met the founder Zhao Jiehui offline was at a CIO closed-door meeting in 2019. At that time, he shared some business thoughts on the data middle platform on the stage. Although the speech exceeded the time limit, he was the sharing guest with the most "rewards" from the audience.

" This person is not a fancy person, and he has a deep understanding of technical architecture ." A senior CIO commented on him. In hindsight, the rapid decline of the middle platform craze is indeed as Zhao Jiehui judged. He believes that the ultimate goal of the middle platform is actually two points: to make the business agile and to share data online in real time. The simplest way to verify is whether the company can launch new business faster and whether the data can serve the business online. But it is obvious that the direction of the industry has gone off track at that time.

Since 2021, I have made an appointment with Zhao Jiehui and the core executives of Dipu Technology to have an in-depth communication every year. The discussion is completely open but the information density is very high, from products, technology, to organization, business, from domestic to foreign, and how to balance daily work and family life. Compared with other entrepreneurial companies, Dipu Technology gives people the overall feeling of simplicity and altruism - making valuable products and letting all employees of the company feel identified .

Unlike other industries, the digital economy is a very broad concept. In this system, each player has his or her own responsibilities, so it is important to anchor your own strategic positioning.

As an "old Huawei man", Zhao Jiehui also believes in the law of value. "Deepen the beachhead and build low dams" was first proposed by Ren Zhengfei and is now also the entrepreneurial secret of Diptech Technology. Deepen the beachhead means to practice internal skills to lay a solid foundation and then create value. As for "low build dams", it means to restrain one's greed and not think about taking in all customer needs. The core of the strategy is not only the war (what to do), but more importantly the strategy (what not to do).

So when a number of giants invested resources in general large models, Diptech chose a more covert path - going deep into the "deep waters" of the industry and building a moat outside the technological range of the giants.

From accurately positioning itself as a bottom-level data platform at first, to creating the real-time intelligent lake warehouse platform FastData, to the earliest domestically launched FastAGI enterprise-level artificial intelligence solution, and then to becoming one of the largest domestic enterprise-level large-model AI application players, unlike the "stitched and pieced together" industry large-model products on the market, Diptech Technology's core thinking in making products is how to get rid of the "death spiral of functional superposition" - while continuously enriching the product's technical depth, ensuring its unity.

"In China, the focus of big model products for the B-side is data governance capabilities. The success of enterprise-level big models does not require higher accuracy, but truly achieving zero illusion. From this perspective, we are on the same starting line as those giants, and we all need to spend time to solve the problems of data set processing and permission control."

Regarding the big model startup, Zhao Jiehui made a very interesting analogy: "The C-end market is like hunting antelopes on the grassland. If you are not careful, you will be surrounded by big companies. On the other hand, the enterprise-level big model market is like prying shells in the deep sea. There are pearls in each shell, but if you pry them quickly, you can eat your fill." 

"I've been thinking about one thing recently. Many entrepreneurs are obsessed with going from 0 to 1, but tend to ignore what '0' actually represents."

Zhao Jiehui explained that "0" is the original intention of a company's entrepreneurship, that is, what value can you create for the other party. Just like helping retail companies clean up the data sets that have been accumulated for several years, it is often the long-term accumulation of such "hard work" that others cannot do and look down upon that has created the key ability of Dipu Technology to atomize and encapsulate industry know-how into replicable data assets, which eventually broke out in the wave of large-scale model industrialization.

Taking the enterprise-level large-model AI application as an example, the essence behind this matter is to transform the "experience" of the master craftsman into replicable data logic. When the master craftsman leaves or retires, his experience will not disappear, but will be deposited as the digital assets of the enterprise.

Belle Fashion Group (hereinafter referred to as Belle Fashion) is one of the most in-depth partners of Diptech since its establishment. The reason behind this is not only the long-term cooperation between the two parties, but also the fact that Belle Fashion has in-depth thinking and accumulation in digital exploration.

As a well-deserved leading footwear and apparel company in China, Belle Fashion has more than 8,000 directly-operated stores in more than 300 cities across the country, and its business scenarios are extremely complex.

According to Zhao Jiehui's recollection, "Before deciding to fully invest in AI in a systematic manner, the other party was most worried that the large model illusion problem would be infinitely magnified in the scenario." For example, when managers asked "How is a certain SKU priced today?", the answer given by AI is often unsatisfactory.

The lack of rule anchors for the company's own scenarios is the most difficult problem in the eye droop model industry. You should know that this ability does not rely on parameter scale, but on the deep deconstruction of the relationship between "people, goods and places".

The solution provided by Diptech is divided into two steps: the first step is to use the Deepexi enterprise model as the basis, inject Belle Fashion's industry data, and train a Deepexi-RM model specifically for the commercial circulation field (equivalent to equipping general AI with an enterprise knowledge base); the second step is to use the FastAGI solution to transform the trained enterprise model into an "AI salesperson". These AI specialists can accurately handle specific business such as product scheduling, inventory forecasting, and supply chain optimization .

It is understood that the "AI single product operation brain" jointly created by Diptech and Belle Fashion has compressed the decision-making process such as directly calculating the inventory status of goods and proposing replenishment recommendations to minutes. In addition, the system can automatically generate business analysis reports that meet the needs of business decisions, and can comprehensively use AI to decompose more than 500 indicators of a single store and form insights. 

In fact, Diptech has a high penetration rate not only in shoes and clothing, but also in major industries such as medical care, manufacturing, and transportation. The reason for this is that its secret lies in the layer-by-layer control over the closed loop of "data-model-strategy".

China Haicheng, a leading state-owned listed engineering design company in the domestic light industry, has cooperated with Dipexi Technology to build a customized enterprise-specific big model deployed locally by China Haicheng based on the desensitized training data processed by FastData solution, relying on the industry-specific capabilities of Deepexi enterprise-level big model platform and the public knowledge base of the manufacturing industry.

The model collects nearly a thousand professional technical documents to build a complete knowledge base, generating a huge corpus containing corpus slices, question and answer groups and standard graphics. It can process text, images, tables and formulas, and support applications such as file classification, layout analysis, and image and formula recognition.

Actual tests show that the system achieves 90% accuracy in drawing compliance review. When AI can respond to professional consultations within 5 seconds, it means that cognitive intelligence has broken through the "hard barriers" of the manufacturing industry - decades of engineering wisdom has been transformed into digital assets that can be calculated, traced, and evolved.

Low gross profit, excessive competition and overcapacity were the three mountains that weighed on the Chinese software industry in the past. The fundamental reason is that there were no revolutionary breakthroughs in the software industry in the past. At the industry level, the current "language pre-training + reasoning + RL" formula can already solve most tasks, and many new technologies only bring small improvements, and even deviate from the normal ROE track.

What is important is how to fine-tune the model based on specific enterprise data sets and business logic . Recalling the experience of prying Shell in the deep sea of ​​digitalization over the years, Zhao Jiehui believes that the key to Dipu Technology's breakthrough in the fierce competition is to dispel the mystery of the trend and look at and think about the problem from a higher dimension. "This matter is insignificant, but it is crucial."

Taking the concept of industry big model, which is the most talked about nowadays, as Zhao Jiehui understands, it is completely different from enterprise-level big model. The former is an improvement under the original process framework, while the latter is a new thinking framework based on data + model. "Whether it is retail, manufacturing or other industries, starting from data and documents, the essence of the model lies in the manipulation of data and fine-tuning of all parameters, and being able to understand the data set on this basis."

It is this "anti-consensus" perspective that has allowed Dipu Technology to avoid the industry's internal competition of "piling up functions and competing on price" and instead bet on the deep integration of data governance and AI capabilities. In fact, before the big model broke out, Zhao Jiehui had made a judgment based on China's data intelligence industry. At that time, he mentioned, "The most difficult part of China's digitalization lies in the scenarios, but if we can chew these hard bones, we will be more than half successful."

Taking the manufacturing industry as an example, the complexity of China's manufacturing industry is the highest in the world: 39 major industrial categories and 525 minor categories. Each scenario is a unique "data maze" and also a "gold mine of value."

In hindsight, Diptech made the right bet again. According to its prospectus, by the end of 2024, the total number of customers will exceed 245; at the same time, the FastAGI enterprise-level artificial intelligence solution business will usher in rapid growth (the proportion will jump from 5.1% to 37.2%), and the scale effect will begin to emerge (the gross profit margin will rise to 51.9% in three years) .

But this is just the beginning. According to Frost & Sullivan data, China's enterprise-level large-model AI application solutions will show explosive growth in the next five years, and are expected to grow from 5.8 billion yuan in 2024 to 52.7 billion yuan in 2029, with a compound annual growth rate of 55.5%.

From data platforms to AI-Ready, from tools to operating systems, from edge innovation to ecological reconstruction, these achievements are supported by a survival philosophy of "slow business, fast iteration". While Silicon Valley is obsessed with the grand narrative of AGI, Diptech has revealed a truth through seven years of practice: the key to China's AI success lies not in the parameter competition, but in the penetration of the industry's capillaries.