Silicon Flow, the first to catch the traffic from DeepSeek, has newly received hundreds of millions of yuan in financing led by Alibaba | Exclusive

Anchong Waves has learned that AI Infra company Silicon Flow recently completed a financing round worth hundreds of millions of RMB, led by Alibaba Cloud. Existing shareholders such as Sinovation Ventures and other institutions over - subscribed, and China Renaissance served as the exclusive financial advisor. Earlier investors also include Meituan (strategic investment), Sinovation Ventures, Glory Ventures, Miracle Plus, Huachuang Capital, Puhua Capital and other institutions.

Yuan Jinhui, the founder of Silicon Flow, compared this financing round to "a two - way journey". Breaking it down, Alibaba has always made strategic investments in AI infrastructure. At the beginning of the year, Alibaba CEO Wu Yongming announced a huge investment of 380 billion RMB in the fields of cloud and AI hardware infrastructure. This is also the largest investment record in this field by a Chinese private enterprise ever. From Silicon Flow's perspective, in addition to obtaining financing, in the future, "it can also have better ecological cooperation with Alibaba's Tongyi Qianwen and extensive cooperation in aspects such as computing power and domestic and international market expansion," Yuan Jinhui said.

This is also a financing round for Silicon Flow after its explosive growth. The last financing was at the end of 2024 before the explosion. At that time, Silicon Flow completed a Pre - A round of financing worth hundreds of millions of RMB, led by Huachuang Capital, followed by Puhua Capital, and existing shareholder Glory Ventures over - subscribed. It is understood that this financing will be mainly used for talent recruitment, product R & D, and domestic and international market expansion.

Two Bets: DeepSeek and Domestic Chips

Silicon Flow is an important part of the magical year brought by DeepSeek.

As the fastest To D (developers) and To B AI cloud service product to handle DeepSeek traffic, its traffic once surged to exceed that of many To C applications.

Yuan Jinhui said that Silicon Flow's familiarity with adapting to DeepSeek - R1 & V3 stems from "for a long time after the open - sourcing of DeepSeek - V2, we were the only third - party MaaS platform that supported DeepSeek".

Looking back further, it is because they bet on open - source earlier. And DeepSeek happened to be a popular open - source model at that time, highly sought after by many due to its outstanding programming ability.

In addition, another important bet of Silicon Flow is domestic chips. In early February, Silicon Flow cooperated with Huawei Ascend to achieve efficient deployment of the DeepSeek model on domestic chips.

As Yuan Jinhui introduced, the reason why they, as a third - party platform, launched the DeepSeek service based on domestic chips at the first time is that as early as last year, Silicon Flow had done a lot of R & D work on domestic chips, solving many problems in aspects such as operators, communication, and system optimization. This enabled them to be earlier than those suppliers deploying DeepSeek based on NVIDIA GPUs. Of course, the cost also included "colleagues basically didn't rest during the entire Spring Festival, but we paid triple salaries to our overtime colleagues".

Even today, Silicon Flow remains the only supplier in the market that provides large - scale DeepSeek API services using domestic chips.

Regarding the recent progress of domestic chips, Yuan Jinhui said that on the one hand, domestic chips are very easy to use and have no problems in terms of reliability. "In the inference scenario, since the model architecture is relatively convergent, as long as the inference engine is adapted and deployed on the MaaS platform, developers can't even perceive what kind of chips are used at the bottom layer." On the other hand, from a practical perspective, "domestic chips have high cost - effectiveness in large - scale inference services".

After the Explosion: What Can and Cannot be Replicated

The successful bets on open - source large models including DeepSeek and domestic chips have enabled Silicon Flow to achieve a growth rate far exceeding expectations. As of now, the total number of Silicon Flow users has exceeded 6 million, with thousands of enterprise users and over a hundred billion daily Token generations.

Of course, DeepSeek's open - source strategy, while expanding the industry scale, has also intensified the competition among downstream MaaS service providers.

There have always been various debates about whether MaaS can make money. Yuan Jinhui believes that MaaS is a product form that is user - friendly but challenging for suppliers.

In his view, to complete the business cycle, a series of problems in technology, operation, supply - chain optimization, customer acquisition and other aspects need to be solved. If these problems are solved well, MaaS can achieve a win - win situation for users and suppliers. Currently, in addition to exploring development paths in the domestic market, they are also actively expanding overseas markets with better payment capabilities and industrial ecosystems.

In addition, he believes that Silicon Flow's unique advantages also include: neutrality in computing power, models, and scenarios.

Yuan Jinhui is a serial entrepreneur. Before starting his business, he invented the world's fastest topic model training algorithm and system, LightLDA, at Microsoft Research Asia.

Before Silicon Flow, his entrepreneurial resume also includes two experiences: FirstRate Technologies, which developed the open - source deep - learning framework OneFlow, and the once - sensational Yuanchuang Zhizhi.

Last year, when being interviewed by Anchong Waves, he said that the first experience was more like a performance art: obsessed with technological pursuit but finally trapped in commercialization. And Yuanchuang Zhizhi was more like an accident for everyone.

In this ongoing business venture, he believes that he has begun to get a feel for business, and the entire team "has gradually changed from the original laboratory form to a more complete and mature business organization".

He attributes Silicon Flow's achievements to a series of correct choices made in the past and the team's execution ability: being optimistic about open - source large models, focusing on AI inference services, and fully committing to adapting to domestic chips. "Although explosive growth requires timing, making a series of correct choices is replicable and sustainable," he said.