The article details Baidu's strategic ownership of four out of five layers in the AI race (chips, cloud, models, apps) showcased at the Baidu Create event, highlighting its resilience against export bans and its ecosystem built over a decade.
๐ง๐ต๐ฒ ๐ฒ๐ ๐ฝ๐ผ๐ฟ๐ ๐ฏ๐ฎ๐ป ๐๐ฎ๐ ๐๐๐ฝ๐ฝ๐ผ๐๐ฒ๐ฑ ๐๐ผ ๐๐น๐ผ๐ ๐๐ต๐ถ๐ป๐ฎ'๐ ๐๐ ๐ฑ๐ผ๐๐ป. Today I attended Baidu Create live in Beijing and talked to people there. Every demo I saw on stage was running in Baidu's cloud, on a Baidu model, inside a Baidu app. Some of those models were trained on Baidu chips. There are five layers to the AI race. Energy, chips, cloud, models, apps. Whoever owns more of them wins the decade. Baidu owns four of those five. They have been building them for years. Baidu made AI a strategic priority in 2010. The Kunlun chip project started in 2012. PaddlePaddle was open-sourced in 2016. Baidu embarked on autonomous driving in 2013. ERNIE shipped its first version in 2019. Kunlunxin spun out as a standalone chip company in 2021. Years of quiet work, while the rest of the industry argued about model benchmarksโฆ Baidu build infrastructure in all layers >>>> ๐๐ต๐ถ๐ฝ๐ Kunlun is Baidu's own chip, designed by Kunlunxin. The P800 cluster has crossed ten thousand cards in production. Kunlunxin is filing for an IPO in both Shanghai and Hong Kong. The chip arm is being valued as its own company. >>>> ๐๐น๐ผ๐๐ฑ Baidu AI Cloud uses Kunlun alongside Nvidia today. The roadmap aims to reach a one-million-card Kunlun chip cluster by 2030. >>>> ๐ ๐ผ๐ฑ๐ฒ๐น๐ ERNIE 5.1 is live. Pre-training cost is 94% lower than comparable models. It runs on PaddlePaddle, the Chinese alternative to PyTorch, with 21.85 million developers and 670,000 enterprises using it. >>>> ๐๐ฝ๐ฝ๐ DuMate, Baidu's desktop agent, took a messy Chinese-English spreadsheet, wrote its own scripts, found a hidden sales drop, and produced a clean dashboard in minutes. The same DuMate built a flash-sale page, wrote daily reports, scheduled tasks, and ran the campaign. Baidu kept calling the human operator a "super-individual." One person plus one agent does the work of a team. MiaoDa, the coding agent, launched on stage. The audience pulled out their phones and built apps from their seats in real time. Baidu Yijing, the digital-human platform, shipped voice, facial expression, and emotion-sensing for livestream commerce, the biggest e-commerce channel in China. ๐๐ด๐ฒ๐ป๐๐ ๐ฎ๐ ๐ฆ๐ฐ๐ฎ๐น๐ฒ. That was the theme of the event. It is also what owning four layers actually buys you. If Baidu pulls it off, the next decade of AI does not look like better chatbots. It looks like agents doing real labor for 700 million users, across every product Baidu ships: search, commerce, code, customer service, robotaxis.
Baidu has quietly built a trillion-dollar infrastructure including chips, cloud, framework, models, and agents, serving 700 million users, unveiled at Baidu Create.
Baidu invited a visitor to their Beijing office to demo a robot hand that mimics human hand movements with minimal delay, highlighting the rapid growth of China's humanoid robotics ecosystem with over 150 companies, as noted by Elon Musk.
An analysis of how Meituan's LongCat Lab trained the LongCat 2.0 model on Huawei's 910C AI chips and CloudMatrix superpods, showcasing China's AI ecosystem overcoming US export controls. The article highlights Meituan's strategic pivot to an AI-driven lifestyle super app and Huawei's innovations in AI factory infrastructure.
An analysis of DeepSeek's long-term strategy, arguing that their innovations in MoE, GRPO, and KV cache reduction are aimed at building a 10T USD Chinese AI hardware ecosystem rather than selling immediate applications, potentially achieving a 1T USD valuation.