Two years after leaving WANG Huiwen, this entrepreneur caught the first bus with LIANG Wenfeng.

The AI startup scene is booming. In the past two years, entrepreneur Yuan Jinhui has experienced two fantastic journeys.

The first one was in March 2023, when Wang Huiwen, the co-founder of Meituan, acquired Yuan Jinhui's first startup project, OneFlow. OneFlow is a deep learning framework that can accelerate the training efficiency of large models. Wang Huiwen took a fancy to Yuan Jinhui's technical ability.

As a result, Yuan Jinhui and his team joined Beyond Light - Years founded by Wang Huiwen. In June 2023, Wang Huiwen announced his "retirement due to illness", and his old friend Wang Xing stepped in to take over. Beyond Light - Years was acquired by Meituan, and Yuan Jinhui entered the Meituan system accordingly.

The second one was in August 2023. Yuan Jinhui decided to leave Meituan and start a new business again, founding Silicon - based Flow. He targeted the AI Infra startup field, which is an intermediate layer above the chips and below the large models and applications. It can achieve stable and efficient model invocation through technical means.

On January 20th this year, DeepSeek released and open - sourced the R1 inference model, which quickly became a hit. In February, Yuan Jinhui led his team to work overtime without a break during the Spring Festival. Silicon - based Flow jointly deployed the R1 model with Huawei Cloud. This caused the traffic on Silicon - based Flow's official website to soar 40 times, even exceeding that of Tencent Cloud and Alibaba Cloud for a while.

Yuan Jinhui also has another connection with Liang Wenfeng, the founder of DeepSeek.

After leaving Meituan, Yuan Jinhui was looking for financing everywhere. At this time, Liang Wenfeng showed the "willingness to invest", but the two had a disagreement on the issue of open - sourcing. Liang Wenfeng insisted that the model inference deployment engine should be open - sourced, believing that there was no way out without open - sourcing. However, Yuan Jinhui hadn't got over his previous startup experience with OneFlow - due to insufficient commercialization, the project had been in financial difficulties several times.

"We previously developed the open - source deep learning framework OneFlow. Open - sourcing doesn't necessarily contribute to success. One has to figure out the business logic first," Yuan Jinhui told China Entrepreneur.

So, the investment didn't materialize. However, the two companies have always had business intersections. Whenever DeepSeek updated its large models, Yuan Jinhui's team would be the first to deploy them on Silicon - based Flow's servers, including the V2 model released in May 2024 and the V2.5 released in September. For some time, Silicon - based Flow was the only third - party service provider to deploy DeepSeek - V2 and V2.5.

From starting his first business in 2016 to his second in 2023, Yuan Jinhui summarized entrepreneurship: "Now, starting a business is like crossing a desert. You carry some water supplies on your back. If you don't find an oasis on the way, you may run out of water and fail to cross. But now, we've found an oasis. If it's big enough, I'll stay here. Or maybe after replenishing supplies here, I'll continue moving forward."

The supplies refer to getting money from investors, and the oasis represents the commercial opportunities brought by the popularity of DeepSeek, which gives Silicon - based Flow cash flow and growing business.

Silicon - based Flow completed its latest round of financing in June this year, led by Alibaba Cloud with an investment of hundreds of millions of RMB. Old shareholders such as Sinovation Ventures over - subscribed. Earlier investment institutions also included Meituan (strategic investment), YaoTu Capital, and MiraclePlus.

This is the rhythm of a serial entrepreneur in the AI era. With the rapid iteration of technology, there is no eternal barrier. One can only evolve with technology until finding that large enough oasis in the desert.

The Battle of Large - scale Expert Parallelism

Before DeepSeek released the V3 model in December 2024, Liang Wenfeng asked Yuan Jinhui's team if they wanted to deploy it first. After calculating the costs, Yuan Jinhui found that they might not have enough money (to rent servers). At that time, everyone joked that they should quickly raise funds and ask investors to transfer money.

There are two difficulties in deploying the V3 model. First, the optimal deployment structure requires large - scale expert parallelism. Without expert parallelism, the production efficiency is very low, and it's hard to make the numbers work. To reproduce large - scale expert parallelism, a large amount of R & D work is needed.

Second, to deploy large - scale expert parallelism, dozens or even hundreds of graphics cards are required. The monthly rental cost for these cards is about 70,000 - 80,000 RMB. If dozens of servers need to be rented, it's a significant cost for a startup company.

By January 2025, the R1 model based on the V3 model architecture was launched. The traffic storm it triggered caused the DeepSeek APP and website to crash directly. It could only answer one question per user per day. In the first week, the number of downloads of the DeepSeek APP was 6 million, and by the second week, it had soared to 20 million. Its popularity combined with the population flow during the Spring Festival made DeepSeek a hot topic on the streets.

Source: AI - generated

It happened to be the Spring Festival at that time. Yuan Jinhui had to mobilize the entire team. The team didn't stop working during the Spring Festival, and the technical team also worked overtime on weekends. This situation lasted until April. Except for the DeepSeek official, no one was able to break through the large - scale expert parallelism technology in a short time. Everyone could only "replace the engine while flying the plane". First, they accessed the traffic by piling up resources, then updated the inference engine architecture, and gradually replaced the old architecture. This process lasted for about two months.

Catching the traffic dividend made Silicon - based Flow well - known in February. It suddenly became a C - end application overnight. C - end novice users mainly tried it out for novelty, but the retention rate was not good. When large companies' C - end products later deployed DeepSeek, they would switch to the products of large companies. For example, the monthly active user number of Tencent Yuanbao soared from millions to 40 million in the first quarter.

This traffic migration battle established the "brand awareness" of Silicon - based Flow. In the API call windows of some AI manufacturers, in addition to Tencent Cloud, Volcengine, and Alibaba Cloud, "Silicon - based Flow" is also an option.

Looking back on this suddenly erupted battle, it can be said that Yuan Jinhui had some gains: First, he made large - scale expert parallelism "usable" and then "easy to use"; second, he fought a big battle, opened up the market, trained the team, and increased brand awareness.

Commercialization in the Era of Large Models

MaaS (Model as a Service) is a business model born in the era of large models. For users, they pay according to the number of Tokens they use. However, some entrepreneurs hold the opposite view on this model.

Yuan Jinhui told China Entrepreneur that whether MaaS services can make money depends on the company's ability. For example, in terms of technology, using the same computing power to achieve better efficiency.

After achieving large - scale expert parallelism in April, its Token output efficiency was dozens of times that of non - expert parallelism, indicating there is room for making money.

To solve the deployment problems of external companies, DeepSeek specially held an "Open - source Week" and released relevant papers. DeepSeek mentioned that "the optimization goal of the V3/R1 inference system is: greater throughput and lower latency. To achieve these two goals, our solution is to use large - scale cross - node expert parallelism."

Yuan Jinhui gave an example. Suppose he has 40 machines, and the monthly rental cost is about 3 - 4 million RMB. If they are fully utilized, the monthly income from the Tokens produced can reach over ten million. Yuan Jinhui said, "The key is to fully utilize them." This involves operational efficiency issues such as customer acquisition and staggered operation of machines.

Yuan Jinhui said: "As a supplier, whether one can do well in MaaS depends on whether one can handle this systematic project. For companies doing Infra, their main task is to solve users' pain points or industry difficulties."

So, when some startups gave up deploying DeepSeek to provide MaaS services, Yuan Jinhui seized this opportunity.

To attract users, Silicon - based Flow launched an operation strategy of "get 14 RMB when registering and another 14 RMB when inviting successfully" soon after the launch of the MaaS platform. Yuan Jinhui joked that this popularity almost made his cash flow unsustainable.

But Yuan Jinhui doesn't regret it. He said that there are risks in making any decision. "If you don't take a gamble, you can only wait for the next opportunity. At least in this opportunity, you're not qualified to participate. You can only do projects like software sales or all - in - one machines. In fact, no one doesn't want to do the best business, and the best business is AI cloud services."

Yuan Jinhui said: "During the startup period, what you should worry about is not that the problems are too challenging, but that there are no challenges. Always encountering easy problems is meaningless."

In 2025, Yuan Jinhui believes that the key issue that the entire AI industry needs to consider is commercialization.

He said: "Previously, everyone only cared about technical problems. But now, the whole industry is thinking about what products to make and how to complete the commercial closed - loop of products."

Entrepreneurship in the AI Era - "Anxiety"

For Silicon - based Flow, the keyword for 2025 is high growth.

Yuan Jinhui said: "How can we increase our market share? Whether it's small and medium - sized B - end customers or large B - end customers, whether it's public cloud or private deployment, and even we can explore the overseas market and turn it into an oasis."

To cross the "desert", in addition to the supplies carried with you, you must constantly look for oases.

Yuan Jinhui said that the startup process may not be a one - step - to - success process. ByteDance didn't start with Douyin in the first step. It had several leap - forward curves in the middle. "Seize the existing business opportunities, standardize products, and increase market share as much as possible to become a safe haven against future risks. Only then can you seize better opportunities in the future and make better products."

Doing AI Infra entrepreneurship essentially means looking for the gaps between large companies and finding trenches between the dense firepower. However, the product's shelf - life is longer than that of pure C - end products, and the barriers are thicker.

One of the advantages of Silicon - based Flow is "neutrality". Compared with model manufacturers, the latter promote large - model cloud services and give priority to selling their own models.

Yuan Jinhui said that the AI industry is still in a wild - growth stage. The cannons of large companies can't cover every corner. At this time, the most important thing is not to think too much and just run forward desperately.

Just as there is a ceiling for model capabilities, there is also a ceiling for AI Infra technology. It's doomed not to stay far ahead of competitors for a long time, but the key is to do a good job in detail optimization and engineering focus.

"Anxiety is a characteristic of AI companies at this stage. Each company shines for a while, and you can't stop." Yuan Jinhui summarized.

Starting over, Yuan Jinhui wants to challenge himself again. "In terms of excitement, not only do you have to do well in technology, but also in products and business. It's like playing a game. You have to pass one level after another. If you always repeat the same level, you'll quickly get bored and lose interest."

This article is from the WeChat official account "China Entrepreneur Magazine" (ID: iceo - com - cn). Author: Yan Junwen, Editor: Zhang Xiaodi. Republished by 36Kr with permission.