--- title: "Baidu's Kunlunxin — the AI chip unit racing to a Hong Kong IPO" date: 2026-10-15 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-chips-kunlun language: en --- # Baidu's Kunlunxin — the AI chip unit racing to a Hong Kong IPO > Spun out of Baidu in 2021, Kunlunxin (昆仑芯) has gone from powering search to filing for a Hong Kong listing — and it shows how serious China's hyperscalers are about owning their compute. Baidu built its first AI chip to stop renting so much compute for search. Five years later, that chip business is knocking on the door of a Hong Kong IPO. The story of Kunlunxin (昆仑芯) is the clearest example of a Chinese internet giant turning a cost center into a standalone company — and it says a lot about where domestic AI silicon is heading. ## Where Kunlunxin came from Kunlunxin (昆仑芯) started life as Baidu's smart-chip and architecture team. It became an independent company in **April 2021**, with Ouyang Jian (欧阳剑), Baidu's chief chip architect, as CEO. Its first outside funding round that year valued it at about **¥13 billion (≈ US$1.83 billion / HK$14.3 billion)**. The technical root goes back further. The team released its self-developed **Kunlun XPU** architecture in 2017 and has iterated ever since: - **Kunlun 1** shipped in volume and was deployed across Baidu Search and Xiaodu, with reports of **over 20,000 units** in use. - **Kunlun 2** entered mass production in August 2021 and spread into internet, industrial, smart-city, and computing-center deployments. - **Kunlun 3 (P800)** reached mass production in 2024. The company describes it as built on a **7nm process** with **over 50 billion transistors**, **256 TOPS** of INT8 compute, and roughly **3x the performance** of the previous generation. Those spec lines are the company's own, and they should be read as claims, not audited benchmarks. ## The Baidu World 2025 roadmap At Baidu World 2025, executive Shen Dou (沈抖) laid out a strikingly concrete plan: ship a new Kunlunxin part **every year** for the next five years. - **Kunlun M100**, optimized for large-scale inference, is targeted for **2026**. - **Kunlun M300**, aimed at training and inference for very large multimodal models, is targeted for **2027**. - The long-term goal is to scale a **single cluster to the million-card (百万卡) level by 2030**. That last number is ambition, not a shipped product. But the cadence — a new chip annually — is the kind of commitment that turns a research project into an industrial pipeline. ## The IPO move In early 2026, Kunlunxin took a step few Chinese AI-chip firms have reached: it submitted a **confidential A1 filing to the Hong Kong Stock Exchange**, the first formal step toward a Main Board listing. For a business born inside Baidu, going public changes the calculus entirely. - It broadens the customer base beyond Baidu's own cloud and services. - It funds the multi-year R&D needed to keep pace with fast-moving accelerator technology. - It signals confidence that domestic AI silicon can stand on external revenue, not just parent-company demand. Investors already in the cap table read like a who's-who of Chinese strategic capital: CPE Source, IDG, Legend Capital, and Yuanhe Puhua in the early round, later joined by entities including China General Technology's investment arm, Societe Generale's China vehicle, CITIC Securities, and — notably — **BYD**, which invested in 2023 and is also a customer. ## Why this matters beyond Baidu Kunlunxin is not trying to be a general-purpose GPU. Its edge is the same as its hyperscaler peers: tight co-design with the parent's software. Kunlunxin silicon is built to run Baidu's ERNIE (文心) large model (大模型) and search workloads, which means it can be optimized for exactly those jobs rather than every job. According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the push to take chip units public is less about the IPO headline and more about giving domestic silicon a real external customer base beyond the founding parent company. ## Honest limitations - **Benchmarks are self-reported.** The 256 TOPS and 3x-generation claims come from the company. Independent, third-party comparisons against Nvidia or other domestic accelerators are limited. - **Foundry exposure is unstated.** Kunlunxin has not publicly named who manufactures its 7nm parts. Leading-edge production remains sensitive under export controls, and that supply line is a structural risk. - **The million-card target is far out.** A 2030 cluster goal is a vision statement. What matters is whether the yearly cadence actually ships competitive parts — and that is still unproven at the frontier. - **Concentration risk.** Much of the credible demand today is Baidu itself. A successful IPO depends on proving the chips sell to others, including rivals who may prefer their own silicon. ## What readers can do now - **If you buy AI compute in China**, watch Kunlunxin's M100 and M300 launches as real test events. A benchmark from an independent lab will matter more than the filing date. - **If you build on Baidu AI Cloud (百度智能云)**, ask which Kunlunxin generation backs your instances and whether roadmap parts are available for early testing. - **If you follow capital markets**, treat the Hong Kong filing as a bellwether: a successful listing would validate the "spin off the chip unit" playbook that Alibaba and others are also exploring. - **If you invest**, separate a charismatic roadmap from shipped, benchmarked silicon. Kunlunxin has real deployment history with Baidu — but the IPO thesis rests on customers beyond Baidu, and that is the number to watch. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-kunlun Free to quote with attribution and a link to the original.