Baidu’s Chip Unit Asked IPO Investors to Buy Its Semiconductors
The Takeaway
- Kunlunxin, Baidu’s AI chip unit, targets a $50 billion Hong Kong IPO.
- IPO investors must commit to buying chips worth 3-7x their subscription.
- Company seeks SMIC production after missing China’s secure chip list.
In China, chip companies may have found a new clientele for their semiconductors: IPO investors.
Kunlunxin Technology, an AI chip firm majority owned by search engine company Baidu, is planning to go public in Hong Kong at a target valuation of $50 billion, according to a person who participated in a recent investor road show by the company and another person with direct knowledge of Kunlunxin’s plans. And as it lines up a group of investors to take a chunk of the offering, Kunlunxin is prioritizing those who commit to buy its chips, according to two people who participated in a road show.
Such investors include investment funds backed by local governments that have a mandate to invest in AI and semiconductors. They have been asked to buy chips with a value three to seven times the worth of their planned subscription in Kunluxin’s initial public offering shares, according to the people. The timing of the IPO is unclear, as the listing is still going through regulatory approval.
It’s not uncommon in the U.S. for tech firms going public to ask investment banks handling the offering to become customers. SpaceX, for instance, required the banks arranging its recent IPO to buy into its Grok AI model. However, these arrangements are not common in China’s tech IPOs. Moreover, asking investors to buy chips, a huge commitment that would only make sense for companies that operate their own data centers, reflects the fact that it is getting tougher to compete in China’s AI chip market.
China has pushed the local tech industry to become more technologically self-reliant, responding to U.S. chip export restrictions imposed against Beijing in 2022. That has spawned a group of startups like Kunlunxin, all tackling the same part of the AI chip market—inference, the process where AI models generate responses or perform tasks.
Kunlunxin’s target valuation of $50 billion is almost 40% higher than the $36 billion market value commanded by Baidu, which currently owns 58% of Kunlunxin and is traded both on the Nasdaq Stock Exchange and in Hong Kong.
What gives Kunlunxin confidence to shoot for this valuation is that it has a built-in customer in Baidu, which gives it an advantage over most domestic chip firms that are still struggling to establish a significant client base.
Kunlunxin would be the sixth Chinese AI chip designer to go public since December 2025. It couldn’t be learned how much it is looking to raise in the IPO. The company earlier sought to raise as much as $2 billion in the offering, Bloomberg reported in January.
Kunlunxin, named after a 1,864-mile-long mountain range in China revered in Taoist mythology as the “ancestor of mountains,” was founded in 2011. Its long track record, plus the compatibility of its product with Nvidia’s Cuda software system, make it easier for developers to move some tasks away from Nvidia hardware.
Baidu has been training new versions of its Ernie model on Kunlunxin chips, partly replacing Nvidia graphics processing units, The Information reported. In May, Baidu said a key version of Ernie 5.1 had been trained on the chips.
Crowded Market
One reason most Chinese AI chip designers have focused on inference is because they are less difficult to make than chips used in training models.
Kunlunxin’s chips sit mainly in this segment. Its P800 series is used largely for inference, or running already-trained AI models, while also supporting the post-training steps of fine-tuning and parameter tuning, or using small datasets to improve models. Baidu said Kunlunxin’s M100 chip series is ready for large-scale inference this year, and its M300 series will be ready for training and inference in 2027.
Kunlunxin is chasing three kinds of customers: central government–owned enterprises, large internet companies and large AI-model developers, according to two employees. Tencent has become a major external customer, they added.
Kunlunxin’s pitch to state-linked buyers took a hit after China left it off the first “secure and reliable certification list” for domestic AI training and inference chips. The list, released in May, includes hardware from many of the company’s rivals, such as Huawei, MetaX, Moore Threads and Alibaba’s T-Head unit, but not Kunlunxin. This certification is an important procurement guide for government agencies and state-owned companies, which are under pressure to use trusted domestic technology.
Part of the problem is where Kunlunxin makes its chips. The company in the past has relied on Samsung Electronics to manufacture them, undercutting its pitch as a domestic supplier. It is now in talks with China’s state-owned foundry, Semiconductor Manufacturing International Corp., to shift some of the production home, according to two separate Kunlunxin employees.
Alibaba is also planning an IPO for its T-Head unit, Bloomberg reported in January. In June, the unit tripled its registered capital to $148 million, its first increase in more than three years, and underwent a corporate restructuring.