Kunlunxin

What Kunlunxin is

Kunlunxin is a Chinese company that designs processors for artificial intelligence workloads in data centres. It was carved out of Baidu: in April 2021 it became an independent company after completing a first funding round that valued it at about 13 billion yuan, with Baidu's chief chip architect Ouyang Jian as chief executive^[1].

AI chips of this kind are built for the heavy numerical work behind training and running large models — the matrix operations that general-purpose processors handle less efficiently. A company that designs its own chip can shape the hardware around the software it already runs.

Where it came from

Baidu launched its first Kunlun AI chip in July 2018. The chip was built on Baidu's self-developed XPU architecture and manufactured on Samsung's 14 nm process^[1]. The architecture describes how the processor's computing units are arranged and which instructions they execute; keeping that design in-house means the chip can be tuned to the workloads Baidu itself uses.

Baidu founder Robin Li has said the company began building its own chips because buying others' chips was too expensive^[1]. That motive — cost and supply of purchased hardware — sits behind the spin-out as well: the chip line began as an internal project, and became a separate company once outside investors came in.

Products

Kunlunxin's first-generation products were deployed in Baidu's own data centres in 2020, its second generation entered mass production in 2021 and reached the open market in 2022^[1]. That sequence — inside first, outside later — is a common path for chips developed by a large internet company: the parent's data centres act as the first customer and the first test of whether the hardware runs reliably at scale.

The third-generation P800 series launched in 2024 on the self-developed XPU-P architecture^[1][2]. In 2025 the company lit up China's first self-developed 32,000-card cluster and kept it running stably^[1]. A cluster on that scale means thousands of chips wired together to work as one system; keeping it stable is a test of networking and scheduling software as much as of the chips themselves.

The roadmap

At Baidu World in November 2025, Kunlunxin published a five-year roadmap: the M100 for 2026, the M300 for 2027, an N series for 2029, and a million-card single cluster for 2030^[1]. The plan moves in two directions at once — new chip generations on a yearly rhythm, and a jump in the size of the clusters those chips are meant to fill.

Position in the market

IDC data put Kunlunxin second among domestic AI chip shippers in 2024 with 69,000 units, behind Huawei Ascend^[1]. That places it among the larger of China's home-grown suppliers, in a market where domestic chips are measured against each other as well as against imported hardware.

Investors in Kunlunxin include CPE, IDG Capital, Legend Capital, Yuanhe Puhua, BYD and funds linked to China Mobile^[1]. The list mixes financial investors with strategic ones — an automaker and a telecom operator — which fits a company whose customers are the operators of data centres and large computing clusters.

Kunlunxin started IPO guidance for the STAR Market in May 2026 while continuing to pursue a Hong Kong listing^[1]. IPO guidance is the preparatory stage before a formal listing application, in which a company works with advisers on the disclosure and structuring it will need. Pursuing two venues at once is unusual but not unheard of for Chinese hardware companies.

Why it matters

Kunlunxin's history traces a wider shift: a search and AI company building chips for its own data centres, then turning that line into an independent supplier with outside capital, an open-market product and a public listing in view. Its roadmap and its cluster targets show where the ambition now points — not a single chip, but the systems built out of many of them.