What happened
At its annual Baidu World 2026 conference in Beijing, Baidu founder and CEO Li Yanhong said demand is now as important as training demand and unveiled a multi-year in-house AI chip roadmap, Chosunbiz reports. The centerpiece is the Tianchi 1024 supernode, slated for 2027, which links 1,024 AI chips into a single computing system. Baidu claims three times higher chip-to-chip bandwidth than existing systems, 2.5 times better training performance, and three times higher inference performance. The roadmap continues with the high-performance M300 chip next year, following this year's cost-focused M100, and several next-generation chips built on a new 'N architecture' planned for 2028. Baidu also officially launched Ciyuansuo, an AI model marketplace that aggregates model providers, and introduced two application models: a deep reasoning model for fast-response services like Baidu Search and a micro-expression model for more natural AI-generated video.
Chosunbiz reports that Baidu held Baidu World 2026 on November 12 at the China National Convention Center in Beijing, where founder and CEO Li Yanhong said that while most computing demand previously went into developing and training AI models, demand has become equally important, and that inference will be the main long-term growth engine for the AI semiconductor market as AI agents take on more complex tasks directly.
The company's most prominent reveal was the Tianchi 1024 supernode, planned for 2027, which connects 1,024 AI chips into a single computing system. According to Baidu's stated figures, chip-to-chip data transfer bandwidth is three times higher than existing systems, AI model training performance is 2.5 times better, and performance is three times higher. A supernode links multiple AI accelerators over a high-speed network so they operate as one large system, addressing the problem that slow communication between chips limits overall efficiency.
Baidu also laid out a chip roadmap aligned with frontier model advances: after this year's cost-focused M100, the high-performance M300 arrives next year with stronger compute, memory, and chip-to-chip connectivity, followed in 2028 by several next-generation chips based on a new design structure Baidu calls the 'N architecture.' The company plans to combine its own chips with optical interconnects and high-speed linking technologies.
On the software side, Baidu officially launched Ciyuansuo, a marketplace aggregating AI model providers so users can purchase model services. Baidu said average daily model calls surpassed 50 million one month after launch, with more than 700 billion tokens served daily. Li said some existing intermediary services do not guarantee model authenticity, performance, response speed, or stability, and Baidu plans to verify models, monitor service quality, and assign credit ratings to providers.
Alongside its recently released ERNIE 5.5 large language model, Baidu introduced a deep reasoning model that improves ability and processing speed for quick-response services such as Baidu Search, and a micro-expression model that makes facial expressions, emotions, and voice more natural in AI-generated video. The company said the new application models cut inference service delivery costs by more than 30 percent and video production costs by 50 percent. Li also argued AI is an 'organizational revolution,' predicting companies will reorganize around tasks rather than job titles, and Baidu said its internal active AI agents grew from about 5,000 a year ago to 37,000, with weekly agent-completed tasks rising from 84,000 to more than one million, and its Baidu Dazi agent's first-pass acceptance rate on tasks of at least 15 steps rising from 54.6 percent to 88.2 percent.
Source details: biz.chosun.com
Why it matters
The announcement signals a faster Chinese push for semiconductor self-reliance under U.S. tech controls, with one of China's largest tech companies committing to a multi-year in-house accelerator roadmap rather than relying on foreign suppliers. If the claimed performance gains hold, a 1,024-chip supernode could materially lower the cost and raise the scale of training and serving large models for Baidu's own products and customers. The shift Li described from training-bound to -bound demand tracks where much of the industry's compute spending is heading as AI agents take on complex, multi-step tasks. The Ciyuansuo marketplace, with Baidu-reported figures of 50 million daily model calls and 700 billion daily tokens one month after launch, plus verification and credit ratings for providers, is an attempt to solve trust and quality problems in a crowded Chinese model market. Baidu also reported that internal active AI agents grew from about 5,000 to 37,000 in a year, with weekly completed tasks rising from 84,000 to more than one million.
The chip roadmap is a concrete self-reliance move under U.S. technology controls on China, extending Baidu's in-house accelerator program across at least three generations through 2028 rather than a single product announcement.
The claimed supernode gains, if independently verified, would affect the economics of training and serving large models at a time when Baidu itself says is becoming the dominant compute demand.
Ciyuansuo addresses authenticity and quality problems in a crowded model market, and Baidu's reported call and token volumes suggest real early usage, though these figures are company-reported and not independently confirmed.
Baidu's internal agent statistics, including the growth in active agents and the Dazi acceptance-rate improvement, offer a company-provided data point on how quickly agentic workloads are scaling inside a major Chinese tech firm, which is the stated driver behind the -focused chip strategy.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Whether Baidu delivers the M300 on schedule next year and the Tianchi 1024 in 2027, and whether independent benchmarks support the claimed bandwidth, training, and gains. How Ciyuansuo's verification and credit-rating system affects model providers and adoption. Whether the reported inference cost reductions of more than 30 percent and video production cost cuts of 50 percent translate into observable pricing for customers. Details on the N architecture as 2028 approaches.
Delivery of the M300 next year and the Tianchi 1024 supernode in 2027, and any independent benchmarking of the claimed 3x bandwidth, 2.5x training, and 3x improvements.
How Ciyuansuo's verification, quality monitoring, and provider credit ratings are implemented, and whether the marketplace's reported usage grows.
Whether the claimed cost reductions of more than 30 percent for services and 50 percent for video production show up in customer-facing pricing.
Details and specifications of the N architecture as the 2028 timeframe approaches.