Infinigence takes its AI chip-translator pitch to Hong Kong investors

With ties to Tsinghua University, the AI infrastructure specialist has won backing from prominent tech and venture investors as it reportedly prepares for an IPO

Key Takeaways:

• Infinigence has reportedly raised 2.3 billion yuan in a Series C funding round, building on earlier backing from Tencent, Baidu, Xiaomi and HongShan
• The company makes software that lets AI models work across different chips, and is expanding into broader enterprise services

By Hu Minghe

With all the different AI models and chips out there, the landscape can often look like a modern Babel, filled with different software and hardware speaking languages that may sound foreign to one another. That’s providing fertile ground for companies like Infinigence, whose software lets AI models run across many different types of chips.

Now, Infinigence is also finding fertile ground in the investment community, where it has reportedly raised 2.3 billion yuan ($344 million) in a Series C funding round, according to Chinese media reports. The financing, reported last week, lifts the company’s cumulative fundraising to 4.3 billion yuan, with investors over successive rounds including internet groups Tencent and Baidu, smartphone maker Xiaomi and Lenovo Capital. Venture capital firms HongShan, formerly known as Sequoia China, and Qiming Venture Partners have also invested.

Stock market investors may become the next to join that roster, following a Bloomberg report this month that Infinigence had confidentially filed for a Hong Kong IPO.

Tsinghua roots

Infinigence is closely tied to Wang Yu, an electronic engineering professor at Tsinghua University, China’s leading sciences school. Company co-founder Xia Lixue, Wang’s first doctoral student, chose industry over academia and was working at Alibaba’s cloud division, where his work included language-model acceleration and AI-chip projects. When ChatGPT-style models emerged in late 2022, Xia saw a commercial opening and, with Wang’s encouragement, launched Infinigence in May 2023. Xia is now the company’s CEO.

The opportunity lay in a market with both a glut and a shortage of computing power. In an April 2024 interview, Xia described an example of operators sometimes being forced to leave 100 machines idle unless one buyer rented them all, because dividing them among 10 customers was too complicated. Infinigence set out to make better use of that underused capacity by making it accessible to more businesses.

Translating between chips

Running AI software across different hardware is difficult because each chip family uses its own software tools and instructions. U.S. export restrictions on advanced semiconductors have pushed Chinese companies to test a wider range of domestic alternatives, multiplying the hardware systems that businesses need to support. That variety has created demand for a software layer that can translate between AI models and the diverse chips beneath them.

In his 2024 interview, Xia described an approach that works like a common language. Normally, every AI model has to be separately adjusted to work on every type of chip, so 10 models and 10 chips would mean 100 separate jobs. Infinigence instead converts each model and each chip into a shared in-house format, so everything can work together. A new chip then needs to be adapted only once, not once for every model.

Xia said the Tsinghua lab where he studied had been developing ways to make software work across different types of chips since 2008. He likened choosing model-and-chip combinations to choosing clothing sizes and colors on Taobao, Alibaba’s online marketplace.

Three-dimensional content startup Tripo illustrates how customers use the service. Tripo’s software generates 3D digital models from text prompts or images. According to both companies, Infinigence provides the computing capacity, storage and engineering support required to train and operate that service. The companies work together to combine Tripo’s underlying systems so its developers can focus on their products rather than managing infrastructure.

In an April 2025 interview, Xia described his company’s core business as renting out computing capacity, and charging customers either by time or by actual usage. Infinigence bundles capacity rented from data-center operators with its own software. It also provides software for customers to run AI models on their own systems.

By May 2026, Xia was emphasizing a broader service role. Customers increasingly needed help turning raw computing power and models into reliable working applications, he said. He described expanding beyond hardware efficiency to build more complete operational services around the computing platform. That would bring Infinigence closer to customers’ day-to-day operations, where stability matters alongside computing prices.

A rival heading to Hong Kong

Qiming Venture Partners, one of Infinigence’s backers, said in December 2025 that the company’s customers included Moonshot AI, the developer of the Kimi chatbot, electric-vehicle maker Li Auto and telecom carrier China Mobile. That shows how the company’s lineup stretches beyond AI developers to established businesses bringing AI into their operations. Infinigence also says daily usage of its AI model platform increased 40-fold between December 2025 and July 2026.

Domestic rival SiliconFlow is also racing towards a Hong Kong IPO, posting its listing application on the stock exchange’s website on June 30. SiliconFlow’s filing reported more than 10 million registered users by April 30 and more than 13,000 enterprise customers served by June 21. By comparison, Infinigence’s website says it serves more than 100 corporate customers, though the two companies may count customers differently.

SiliconFlow prospectus

Infinigence develops compact models such as Megrez, first released in December 2024 and designed to run on smartphones and other devices. SiliconFlow says it does not develop its own models or end-user applications. Both companies share a backer in AI model developer Z.AI, also known as Zhipu.

Their commercial services overlap as well. While SiliconFlow runs an online platform that lets developers tap different AI models on demand, it also sets up dedicated systems inside enterprise customers’ own facilities. Those on-site deployments generated 47.1% of SiliconFlow’s revenue last year, underscoring that Infinigence competes for enterprise contracts against a rival offering both general online access and tailored installations.

Both startups also compete with Alibaba Cloud and Volcano Engine, the cloud unit of ByteDance. Xia argued in 2024 that supporting many model-and-chip combinations would be especially costly for cloud giants already managing complex businesses and customer demands.

Keeping its services useful requires constant adaptation. In an August speech, Infinigence chief scientist Dai Guohao pointed to three ongoing industry challenges: fragmented computing resources, high model-operating costs and rapidly evolving models. Each new generation of AI models creates fresh demand for optimization specialists, but also requires them to continually update their software.

To widen its market, Infinigence is teaming up with China Telecom, one of China’s three big state-run telecom carriers. Under a July agreement, the partners plan to build an international computing-services platform that uses the telecom operator’s overseas network to serve Chinese enterprises operating abroad as well as global developers. Planned locations include Southeast Asia, giving Infinigence a path to expand beyond domestic clients.

To subscribe to Bamboo Works weekly free newsletter, click here