From Big Data to 30,000‑GPU Clusters: The Evolution of China’s AI Infrastructure

In the second half of the AI era, algorithmic innovation remains important, but compute power is becoming the ultimate variable, turning the AI arms race from code and models to chips, electricity, and data centers.

Recently, Baidu Intelligent Cloud and the podcast platform "Crossroads" launched a joint podcast where Baidu AI Computing chief scientist Wang Yanpeng and host Koji discuss the evolution of Chinese internet infrastructure since the big‑data era.

"AI Infra is destiny" – Wang Yanpeng recounts the three major stages of China's compute evolution.

The first stage, the big‑data era, was led by Google with seminal papers such as MapReduce, BigTable, and GFS, which laid the theoretical foundation for large‑scale internet infrastructure. Commercial hardware like IBM mainframes and HP/DELL servers were insufficient, prompting Google to use cheap, PC‑class hardware to build massive systems, a philosophy championed by Jeff Dean.

This shift disrupted traditional high‑end server vendors and spurred the rise of distributed systems and open‑source ecosystems like Hadoop, enabling companies to build their own large‑scale data centers.

Koji asked whether Baidu faced similar challenges; Wang confirmed that Baidu's largest monolithic application, search, required processing volumes far beyond e‑commerce and social media, making Baidu one of the earliest Chinese companies to develop its own hardware and software infrastructure.

The second stage, the cloud era, was exemplified by Amazon, which rented out idle server capacity as a cloud service, introducing elastic computing that abstracts away the physical layer.

This era saw innovations such as intelligent NICs that virtualize resources, allowing massive scaling across thousands of servers.

The third stage is the current AI era, a fundamental shift in the computing paradigm. While CPUs remain the backbone of general‑purpose computing, GPUs allocate nearly 100% of transistors to computation, delivering orders‑of‑magnitude speedups for deep‑learning workloads.

Wang explained that NVIDIA’s CUDA, originally not aimed at deep learning, became a crucial abstraction for custom programming on GPUs, and early adopters in academia helped drive the deep‑learning boom.

He noted that AI’s true transformation lies in the scaling law: increasing model parameters and data volume continuously improves capability, ushering in an industrial‑scale era of large models.

Comparing CPU and GPU eras, Wang highlighted that CPUs follow Moore’s law, offering predictable performance gains, whereas GPUs require tightly coupled hardware‑software co‑design, with frequent incompatibilities across hardware generations.

Wang emphasized that data is becoming scarce; synthetic data generation and reinforcement learning are ways to augment data, but massive compute remains essential to produce, clean, and filter high‑quality data.

Regarding Baidu’s Baige platform, Wang described it as a high‑efficiency AI compute platform supporting up to ten‑thousand‑card clusters, offering "heterogeneous, remote, and cross‑network" capabilities that simplify the use of diverse chips, locations, and networks.

Baige aims for extreme engineering efficiency, achieving up to 10% performance improvements that translate into significant cost savings, and integrates a four‑layer stack from chips to cloud, models, and applications.

Wang argued that building such infrastructure requires large platforms; small teams risk becoming mere ops engineers without the breadth to innovate across storage, compute, and OS layers.

He highlighted the importance of close collaboration between infrastructure and algorithm teams, noting that Baidu’s early SSD projects with the search team quickly delivered value, and that modern talent often possesses full‑stack capabilities across applications, algorithms, and architecture.

Finally, Wang reflected on the global AI‑infra landscape, noting that while companies like Google, OpenAI, and Meta each have distinct strengths, the race ultimately hinges on sustained investment in compute, hardware‑software integration, and the ability to support cutting‑edge models.