China has quietly constructed more than 24 gigawatts of datacenter capacity dedicated to AI workloads, a figure that exceeds all of Europe's computing infrastructure and represents a fundamental shift in the global AI infrastructure landscape. This buildout spans over 1,000 facilities across more than 60 operators and excludes an additional 20 gigawatts of projects in development and 30 gigawatts of announced plans, according to newly released analysis from SemiAnalysis.

For years, the scale of China's AI infrastructure investment remained largely unmeasured outside the country, obscured by fragmented ownership, private companies that file no public financial reports, and primary sources published in Chinese. The result was a market settled on lazy assumptions: China is big, but China is mostly empty. Published estimates of China's datacenter capacity differed by as much as 15 times, with many reports citing high vacancy rates as evidence of overcapacity.

That narrative is now colliding with reality. The three largest Chinese tech companies, Alibaba, Tencent, and Baidu, collectively spent $20 billion on datacenter infrastructure in the second quarter of 2026 alone, more than doubling their year-over-year spending and marking the largest capital expenditure step-up in the sector's history. For the first time on record, all three companies posted negative free cash flow, a sign of how aggressively they are investing in compute infrastructure.

Why Is China Building Datacenters at This Pace?

The acceleration reflects a fundamental reality: training and deploying large language models (LLMs), the AI systems that power chatbots and reasoning engines, requires enormous amounts of computing power and storage. As models grow larger and more capable, they demand more memory to store model weights, more bandwidth to move data quickly, and more sophisticated storage architectures to manage the intermediate data generated during inference, the process of running a trained model to generate predictions or responses.

China's homegrown AI models are driving much of this demand. GLM 5.3 and Kimi K3 are among the latest open-weight model releases, while ByteDance's Doubao serves 345 million monthly users as China's equivalent to ChatGPT, and Seedance has released state-of-the-art video generation technology. Every one of those models runs on a datacenter, and the infrastructure race to support them is intensifying.

The challenge facing AI infrastructure builders is what experts call the "memory wall." As models become more complex, the bottleneck shifts from raw computing power to the speed at which data can be moved between processors, memory, and storage. This is driving innovation in storage architecture, advanced packaging, and memory technologies that can handle the demands of large-scale inference.

"As large model inference drives up demand for capacity, bandwidth, and cost efficiency, high-bandwidth flash and heterogeneous multi-media storage architectures stand to play a larger role, combining HBM, NAND, and RRAM to strike a better balance between performance, capacity, and cost for AI inference," explained Yimao Cai, Dean of School of Integrated Circuits at Peking University.

Yimao Cai, Dean of School of Integrated Circuits at Peking University

How Is China Building Datacenters Faster Than the West?

One of the most striking aspects of China's infrastructure buildout is the speed of construction. China routinely delivers 100-megawatt datacenter facilities in under 12 months, a pace that reflects both regulatory efficiency and standardized design practices. Modular datacenters, where facilities are built from prefabricated components that can be rapidly assembled, have become the standard playbook. Tencent deployed its third-generation modular design in 2014, and what was considered cutting-edge in China a decade ago is only now being adopted at scale in the United States.

The buildout is also a nationwide effort involving not just private tech companies but state-owned carriers and power grid companies. China's three state carriers still own roughly one-third of the nation's datacenter capacity, a legacy of the telecom-dominated era when datacenters were primarily used for hosting websites and online gaming services. The national power grid companies are investing aggressively as well, with combined capital expenditure accelerated from 2024 and now exceeding the original blueprint for the 14th Five-Year Plan by 24 percent, with the 15th plan (2026 to 2030) layering another 40 percent on top, totaling over $746 billion.

Steps to Understanding China's AI Infrastructure Strategy

• Capacity Scale: China has built 24 gigawatts of delivered datacenter capacity across 1,000+ facilities, with an additional 50 gigawatts in development or announced, making it the second-largest datacenter market globally after the United States.
• Tenant Concentration: ByteDance, the private company behind TikTok and Doubao, occupies roughly one-fifth of all delivered datacenter capacity in China and rents nearly all of it, making it the single most important customer for wholesale colocation providers.
• Geographic Expansion: The "Eastern Data, Western Compute" policy is driving capacity inland, away from coastal regions, with construction speeds and unit costs that the West cannot match, while overseas leasing by Chinese hyperscalers is set to double from 2026 to 2029.

The concentration of capacity among a small number of tenants reveals another critical dynamic. ByteDance alone occupies roughly one-fifth of delivered datacenter capacity in China, and it rents nearly all of it rather than building its own facilities. This makes ByteDance the single most important customer for every wholesale colocation provider in the country. Meanwhile, the two largest publicly listed Chinese datacenter landlords, GDS and VNET, signed 1.3 gigawatts of wholesale orders in the first half of 2026, but captured barely one-third of the orders from ByteDance and Alibaba combined during 2024 through mid-2026.

The implication is stark: the visible players in China's datacenter market are only the tip of the iceberg. Unlisted operators and state-owned carriers are capturing the majority of AI infrastructure investment, making the true scale of China's buildout difficult to measure from outside the country.

What Does This Mean for the Global AI Race?

China's infrastructure advantage extends beyond its borders. Chinese hyperscalers are leasing capacity from Western cloud providers at massive scale, renting hundreds of thousands of graphics processing units (GPUs), the specialized chips used to train and run AI models, from providers like Oracle. This strategy allows Chinese companies to access Western computing resources while simultaneously building domestic capacity, creating a dual-track approach to the global AI infrastructure race.

The market does face headwinds. Vacancy rates remain high in legacy retail datacenter segments, and datacenter developers compete heavily on price, which can pressure margins. Chip supply is also constrained due to export restrictions, limiting the availability of the most advanced processors needed for cutting-edge AI work. Yet neither constraint has stopped AI datacenters from being built and filled at a remarkable speed, driven by the urgent need to support China's rapidly advancing AI models and the massive user bases they serve.

The broader implication is that the global AI infrastructure race is no longer primarily about the United States versus the rest of the world. It is increasingly about how China's massive, state-supported buildout is reshaping the competitive landscape for AI compute, storage, and inference capabilities. As memory and storage technologies evolve to handle the demands of large-scale AI inference, the infrastructure that supports those technologies will determine which countries and companies can build and deploy the most capable AI systems.