“Professor born in the 1990s, just secured their fifth round of funding.”

Investors are fiercely competing.

Author: Liu Bo

The report is about "PEdaily" in the investment industry.

A funding round left a deep impression on the VC community.

The investment circle learned that Luchen Technology recently completed a several hundred million yuan A++ round financing with a stellar lineup of investors - including the Beijing Artificial Intelligence Industry Investment Fund, Shixi Capital, Capstone Capital, and Lingfeng Capital. It is worth noting that this is already the fifth round of financing for Luchen Technology since its establishment three years ago.

This AI dark horse has an impressive background. The man behind the wheel is a post-90s professor named You Yang. He graduated from Tsinghua University, went on to pursue a Ph.D. at UC Berkeley, and became the first Chinese recipient of the "President's Young Professor" award at the National University of Singapore's Computer Department. In 2021, he chose to return to China to start a business in the AI infrastructure field and has led Lucen Technology to an estimated 40-fold increase in valuation in three years.

This is undoubtedly the most vivid portrayal of China's AI era.

Post-90s, from Tsinghua to university professor.

Join the wave of AI entrepreneurship.

The story of Luchen Technology starts with a super study genius.

Founder and Chairman of Morning Light Technology, You Yang.

Born in the 1990s, You Yang, when studying in the Computer Science department at Tsinghua University, once won the highest Bell scholarship in the department and also received honors such as an Outstanding Graduate of Tsinghua University and an Outstanding Graduate of Beijing. In 2015, after earning a master's degree at Tsinghua University, You Yang decided to further his education abroad and went to UC Berkeley to pursue a Ph.D. degree, where he was under the guidance of Academician James Demmel, Dean of the College of Electrical Engineering and Computer Sciences.

During this period, You Yang received numerous awards for his research papers. He is the only person under 35 years old in the world who has been recognized as the best paper or outstanding paper at four top conferences (AAAI, ACL, IPDPS, ICPP) as the lead author or corresponding author. The method proposed in the paper "Imagenet training in minutes" has set a new world record for ImageNet training speed.

At the same time, he was the most highly cited doctoral graduate in the field of high-performance computing worldwide in 2020 on Google Scholar. He was also nominated by UC Berkeley for the ACM Doctoral Dissertation Award, which only selects 2 out of 81 doctoral graduates, illustrating its prestige.

After obtaining his PhD degree, You Yang received job offers as a young professor from several prestigious universities in the United States, including the Ivy League schools, as well as the National University of Singapore. After careful consideration, he decided to join the National University of Singapore in May 2020. With this decision, You Yang became the first Chinese recipient of the "President's Young Professorship" award in the Computer Science Department at the National University of Singapore.

Why choose entrepreneurship? You Yang frankly stated that the professor from the computer science department at Berkeley pursues the combination of research and production, and the wave of AI development has shown him the entrepreneurial opportunities. "In this era, one of the best ways for us researchers to improve ourselves is through entrepreneurship."

Moreover, the experience of being a professor also laid the foundation for You Yang's entrepreneurship. In his view, professors are different from ordinary scientists. On the one hand, professors need strong communication and expression skills; on the other hand, they also need the ability to manage a team. A professor leading a team of 20 to 40 people in a laboratory is very common, which is no different from running a small to medium-sized company.

In July 2021, You Yang decided to return to China to start a business and officially established Lu Chen Technology in Beijing, which is an AI Infrastructure company. What is AI Infrastructure? There is a popular saying within the industry that, "If developing AI applications is like building a house, then AI Infrastructure companies are like the construction teams providing cement and steel bars."

Yu Yang also fully agrees with this. He believes that finished houses can be used as shopping malls, hotels, office buildings, conference centers, factories, and so on, just as the application of AI models can be in finance, gaming, social networking, e-commerce, energy, pharmaceuticals, and so on. "The essential value of AI infrastructure companies is to help millions of industries quickly deploy large models and use them in practical businesses or applications."

In the early stages of its establishment, Lucheng Technology mainly provided distributed software systems and enterprise-level cloud computing solutions to B2B enterprises. In other words, it is like creating a plugin in the cloud, which can be applied on Alibaba Cloud, Tencent Cloud, and Huawei Cloud, and users pay to use the plugin.

However, You Yang and his team quickly realized that large cloud providers emphasize integrated support and do not prioritize small and medium-sized enterprises. As a result, Luchen Technology quickly adjusted its product form, transitioning to vertically connecting all elements on the cloud, and then matching Luchen Technology's software optimization capabilities to create an end-to-end solution to serve small and medium-sized clients.

Looking back on the three-year entrepreneurial journey, Yu Yang is most deeply touched by the rapid changes in the market and technology. In order to adapt to the changing demands for infrastructure in the AI era, it is necessary to keep pace with the times and make quick adjustments. He said, "Our team has also grown very quickly."

How is an AI dark horse trained?

Entering the era of large models, Luchen Technology has launched the Luchen Cloud Platform, which includes the large model training and reasoning system, Colossal-AI. It is dedicated to helping enterprises reduce the cost of implementing large models and improve training and reasoning efficiency.

In You Yang's vision, Luchenyun is more like a computing version of "DiDi Chuxing," creating an automated platform while leveraging the computing power optimization capabilities of Luchen Technology, allowing users to train or deploy AI models 2 to 5 times faster at the same price, and creating models with parameters ten times higher at the same cost. "We want to build a true computing power ecosystem platform, allowing companies and developers in need of computing power to quickly find the most cost-effective computing power."

If you are looking for benchmark companies overseas, Together AI, CoreWeave, and Lambda Labs could be considered. However, in reality, the latter can be seen as a "compute power real estate agent," more like an intermediary providing compute power sales, not a standardized product or platform.

Lu Chen Technology aggregates the existing legal high-performance computing power from Chinese individuals and institutions. For example, if various local computing centers do not possess optimization capabilities, Lu Chen Technology can acquire computing power at a low price or directly deploy it on the Lu Chen Cloud platform. After optimization, it provides services to various users.

In You Yang's view, for Luchen Technology to become the "Didi Chuxing" of computing power, the most crucial point lies in the user ecosystem. It is not about who has the most computing power, but rather who can reach the most end users. Ultimately, downstream users determine the fate of upstream providers. In other words, computing power will eventually become redundant, and only by establishing a user ecosystem that utilizes computing power can a long-term competitive advantage be created.

In this way, the advantage of Morning Tech becomes prominent - records indicate that Morning Tech's Colossal-AI system ranks first in the world in terms of the GitHub ecosystem index in the field of MLSys (Machine Learning and Systems), with an open-source community of approximately 40,000 to 100,000 developers deeply utilizing Morning Tech's products.

Meanwhile, Luchen Technology's technical capabilities have been recognized by a number of tech giants. As early as 2018, the team had helped Google accelerate the world's best large model from 3 days to 76 minutes on the world's first milliwatt AI-specific cluster. In addition, key technologies from Luchen Technology have been used by Microsoft, NVIDIA, ByteDance, and others.

The large-model video Open-Sora is another flagship product of Lu Chen Technologies. Looking around, it is rare to see domestic AI infrastructure companies building large models themselves. So why did Lu Chen Technologies take this step? You Yang said that first of all, Lu Chen Technologies has a good accumulation of video large-model technology. Previously, it has received outstanding papers from AAAI and ACL, and has published 50 papers on top AI video large-model conferences such as CVPR, ICLR, NeurIPS, ICML, and others.

He pointed out that large-scale video models have just emerged, and there is no leading player in China yet. Currently, some large companies' apps rely on a large amount of short video data not because they have a technological advantage, but because the success of video large-scale models ultimately depends on the quality of data. Therefore, Luchen Technology can quickly catch up.

Next, Luchen Technology is optimistic about this direction for the long term, believing that video large models are more likely to follow long-term scaling laws than LLM (large language models). This is because the training data for video large models is a true reflection of the objective world, and the ultimate creator of the data is the Creator, so large models can certainly discover the inherent patterns within it. The training data for LLM comes from the Internet and books, and the levels of the data creators vary, leading to a lot of ambiguity and junk information.

"The most important thing is that the field with the highest demand for computational power optimization in the future is likely to be large-scale video models. The super apps for large-scale video models will ultimately need to compete in terms of computational efficiency, which is our advantage."

Behind the Chinese AI financing competition

This means "commercialized competition" in English.

Along the way, Lu Chen Technology has left a deep impression on investors.

In 2021, at that time, You Yang was selected for a list by Forbes Asia. Kai-Fu Lee, the founder of Innovation Works, was one of the judges. After seeing You Yang's resume, he recommended him to the investment manager of Innovation Works, and the two sides established contact. You Yang recalled that later on at Innovation Works, he had a face-to-face meeting with Teacher Kai-Fu Lee, and they quickly reached a consensus. From the initial meeting to the funding, it only took a week. At the same time, ZhenFund also got in touch with You Yang, and together they led the seed round of investment.

"I was alone at that time, with only a few papers and code. What surprised me even more was that there were no hedging clauses or even buyback clauses in the investment trading documents at that time, which made me feel the support and trust they had for entrepreneurs." Yu Yang still vividly remembers this."

Quickly, in September 2022, Luchen Technology completed its angel round financing, led by Lan Chi Venture Capital. In 2023, the "Battle of the Models" began, with major technology giants issuing recruitment notices. Some large model manufacturers also expressed acquisition intentions toward Luchen Technology.

There is no doubt that this is a crucial juncture for You Yang and Lu Chen Technology. However, after careful consideration, You Yang believes that the AI large model has not reached its final stage, and being acquired so early has no prospects and bears significant risks. At the same time, he also believes that Lu Chen Technology has long-term value, can survive independently and go public, and has the potential to become the new DataBricks of the AI era.

As a result, Luchen Technology ultimately rejected the acquisition offer and quickly secured hundreds of millions of dollars in Series A funding from two renowned USD venture capital firms. Subsequently, the company completed a Series A+ funding round at the end of 2023, with a leading technology giant from the Fortune 500 as the main investor, and the Greater Bay Area Fund and Singapore Telecommunications Investment Company (SingTel Innov8) participating in the investment as well.

During this period, the valuation has also skyrocketed. The company has grown by about 40 times since its establishment. It is reported that Lu Chen Technology still holds several investment TS at present, but the company's cash flow is relatively sufficient. They will carefully consider the pace of future financing. "I think we should first develop our business and elevate the company to a new height. If we raise a lot of money and push up the company's valuation but cannot create matching business to support the valuation, it will be easy to fall from a high point at that time."

Right now, the AI industry is booming. The latest groundbreaking financing news is that OpenAI announced on its official website that it has raised $6.6 billion (about 46 billion RMB) in its latest round of financing, with a post-financing valuation of $157 billion (over 1.1 trillion RMB).

And turning attention back to domestic markets, large-scale model financing is equally crazy. At the end of July, Baichuan Intelligence announced the completion of Series A financing, with a total financing amount of up to 5 billion RMB. In addition, the other four companies in the "Big Model Five Tigers" - Dark Moon, Zhipu AI, Minimax, and Zero One Wan Wu - have also intensively announced financing from last year to this year, all joining the unicorn ranks.

But controversies also come along with the financing boom. Previously, an investor from Sequoia Capital openly stated that the companies in the current AI race had spent over $50 billion purchasing NVIDIA GPUs, yet the revenue generated during the same period was only around $3 billion. Implicitly, this suggests that the bubble is too large.

It is obvious that the key for AI startup companies is how to achieve commercialization. Relying solely on financing and burning through money without generating revenue will not sustain the company in the long run. Therefore, for AI infrastructure companies like Luchen Technology, choosing a viable and effective business model is crucial.

Looking around, most AI infrastructure companies adopt the MaaS model, but it is not suitable. It is well known that such AI infrastructure companies provide various open-source model API calling services, with shortcomings in independently developed cutting-edge models, essentially competing with ChatGPT, Kimi, Zhipu, and others. In the price war, API prices can drop as low as 1 million tokens for 1 cent, with daily token usage for AI infrastructure companies on MaaS ranging from 100 million to 1 billion, resulting in daily revenue of only a few hundred yuan. Therefore, it is difficult for the MaaS model to increase revenue in the short term.

The open-source model can achieve greater value after fine-tuning with private data, and Luochen Cloud's high-performance GPU and extensive image library allow users to efficiently fine-tune models to meet business needs. From this perspective, Luochen Cloud's business prospects seem more promising.

When talking about the heterogeneous cluster mode of domestic chips, You Yang believes that using multiple different chips in one cluster for model training is a waste of resources. "If domestic AI training chips can eventually develop, they must comply with market rules, perhaps like the US market, where 1 or 2 manufacturers' chips occupy over 95% of the market."

Of course, Youyang also agrees that heterogeneous chip clusters are helpful for inference, as serving different batches of user computations can be completely independent without technical difficulty. However, precisely because the technical barriers are low, the profits may not be considerable. "I found that solutions claiming to accelerate inference by more than 3 times compared to TensorRT/vLLM/SGLang mostly utilize quantization/pruning/distillation, essentially shrinking a large model into a smaller one, without much technical depth."

As of today, Luchen Technology has also delivered a report card—The pure software gross profit margin has reached around 90%, including customers from four Fortune 500 companies contributing tens of millions of yuan in revenue; The number of Luchen Cloud customers is also expected to reach 10,000 by the end of the year. It is also projected that by 2025, revenue (confirmed payment) will reach 150 million yuan.

Looking into the future, he has already set his sights on a goal, "We will achieve revenue of 1 to 3 billion RMB, and then go public with a valuation of 30 to 60 billion RMB." In his words, China's AI entrepreneurial wave is surging, with many players emerging.