The startup allows other companies’ models to run across many types of chip, a neutrality that is useful in a fragmented Chinese market

image credit: Bamboo Works

Key Takeaways:

●   SiliconFlow has raised a fresh $300 million in funding, betting that the layer between AI models and computing chips can become a durable business opportunity

●   The startup hosts and optimizes models from companies including DeepSeek and Alibaba rather than developing a foundation model of its own

Less than three years after its founding, AI startup SiliconFlow has raised more than 2 billion yuan ($294 million) in a new funding round without building a foundation model of its own, becoming one of the latest beneficiaries of China’s AI investment boom. The company is betting that the layer between AI models and the computing chips and other hardware that power them can become a durable business opportunity.

That business model looks lucrative in a currently fragmented environment populated by many types of chips and other hardware, both locally and globally developed. But it carries risk if large language model (LLM) developers, cloud providers and chipmakers eventually decide to integrate their products directly, reducing or even eliminating any role for such neutral middlemen.

The latest funding is the fifth for SiliconFlow in its short lifetime, and includes backers like online travel giant Trip.com, AI specialist SenseTime, telecoms heavyweight China Unicom and Nio Capital. They join a group of earlier big-hitter investors like Alibaba Cloud and Zhipu AI.

SiliconFlow’s work is less visible than a chatbot like ChatGPT, but matters to anyone trying to run one reliably. Large language models generate answers one "token," or fragment of a word, at a time. SiliconFlow hosts models made by DeepSeek, Alibaba’s Qwen, Zhipu and Moonshot, then lets developers query them through a common interface. Its software optimizes model inference for different processors, manages memory and batching, and shifts workloads among chips and data centers.

This kind of model-as-a-service (MaaS) lets customers use a hosted model through an application programming interface (API) rather than buying servers and operating the model themselves. OpenAI’s API is a form of MaaS, with its ChatGPT as a finished consumer product. SiliconFlow’s twist is that it offers models from multiple developers and runs them across multiple types of hardware. SiliconFlow does not make the AI models or chips. It provides the software and computing services that make those models run efficiently on different hardware.

"Token factory"

The company describes itself as being a "token factory." Instead of renting computing power by the hour, SiliconFlow gives customers access to models such as DeepSeek, Qwen and Kimi, at a predictable cost and speed. Its Siliconcloud platform provides public APIs, while larger customers can reserve capacity, deploy SiliconFlow’s software in their own cloud or data center, or use it to manage computing clusters they already own. The company says it serves more than 10 million users and 10,000 enterprise customers.

The strategy reflects founder Yuan Jinhui’s long focus on the infrastructure behind AI models. Yuan studied computer science at Xidian University and earned a doctorate at Tsinghua University, China’s leading science school, under AI scientist Zhang Bo. Both institutions were among China’s first national training bases for integrated-circuit talent and have produced generations of electronics and semiconductor engineers.

After a stint at Microsoft Research Asia, Yuan founded Oneflow in 2017 to improve the efficiency of training large neural networks. Oneflow was later purchased by internet giant Meituan, after which Yuan and former colleagues founded SiliconFlow, shifting their focus from training models to running them.

Yuan’s background became useful during the DeepSeek boom. In February 2025, as demand for DeepSeek’s R1 and V3 models overwhelmed some services, SiliconFlow and Huawei Cloud launched versions running on Huawei’s Ascend processors. That brought visibility to both SiliconFlow and Huawei, but the more important signal was technical. Researchers from the two companies later detailed how they served DeepSeek-R1 on Huawei’s CloudMatrix384 supernode, showing SiliconFlow was doing more than simply reselling access to the model.

China’s fragmented computing market makes the opportunity for companies like SiliconFlow more pronounced at home. Customers may run Nvidia processors alongside rival products from Huawei, Biren, Moore Threads, Metax and other domestic suppliers, each with different software stacks and performance characteristics. A neutral layer that can make several models run across several chip families could save customers significant integration work.

Crowded field

Despite the big market opportunity, SiliconFlow still faces some formidable domestic rivals. Volcano Engine, Alibaba’s (NYSE:BABA)(9988.HK) Alibaba Cloud and Baidu’s (NASDAQ:BIDU) (9888.HK) Baidu AI Cloud ranked ahead of it by token volume in China’s public-cloud MaaS market in 2025. Those rivals have an advantage in their ability to bundle model access with storage, databases, security and existing enterprise contracts. Ppio, Infinigence AI and Luchen Technology compete more directly in inference optimization, multi-chip deployment and their ability to work with private AI infrastructure.

Demand for the types of services that SiliconFlow offers is rising quickly. IDC said Chinese enterprise MaaS consumption jumped from 114 trillion tokens in 2024 to 1.94 quadrillion in 2025, while public-cloud MaaS revenue reached 3.07 billion yuan ($453 million). China’s Ministry of Industry and Information Technology (MIIT) said the country’s core AI industry exceeded 1.2 trillion yuan in 2025 and included more than 6,200 companies. SiliconFlow ranked fourth by token volume but was not among IDC’s five largest providers by revenue, suggesting the company has been quick to implement its technology but slower to monetize it.

The funding boom made abundant capital available across the AI value chain. Twelve recently listed Hong Kong businesses, including Zhipu AI, Minimax and Biren, raised a combined $4.9 billion through their IPOs, according to the stock exchange. Others have raised money privately, such as DeepSeek, which reportedly recently raised more than 50 billion yuan in its first external round – making SiliconFlow’s 2 billion yuan look relatively modest.

So much investment could ultimately undermine middlemen like SiliconFlow, since cash-rich companies like DeepSeek may use their extra funds to independently develop the kinds of services where such middlemen are finding business.

That pressure helps explain why SiliconFlow is moving beyond public APIs. It is expanding its offerings for private enterprises, where customers may pay for dedicated capacity, security, governance and the management of their own hardware. The company also plans to use part of its new funds for overseas expansion. But that terrain is also crowded with other rivals, including U.S. specialists Together AI, Fireworks AI and Baseten, which have also moved beyond model catalogs into customization, routing and dedicated infrastructure.

SiliconFlow has a clearer advantage at home, where China’s fragmented patchwork of domestic chips creates demand for a neutral layer connecting AI models and hardware. That edge may be less pronounced overseas, where the Nvidia ecosystem used by many is more standardized. SiliconFlow’s best chance of protecting margins may therefore lie in embedding its software inside the computing infrastructure of its customers, rather than selling interchangeable API calls. SiliconFlow has shown it can empower massive token consumption. Now, the remaining question is how much of that value it can keep for itself.

To subscribe to Bamboo Works free weekly newsletter, click here

Benzinga Disclaimer: This article is from an unpaid external contributor. It does not represent Benzinga’s reporting and has not been edited for content or accuracy.