AliCloud, an Alibaba subsidiary, partners with Nvidia to develop GPU-based computing cloud platform in China
Context & Ripple Effects
This 2016 deal is the opening move of a decade-long arc: AliCloud brought Nvidia GPUs into Chinese cloud infrastructure just as AI training demand was forming, giving Nvidia a distribution channel inside Alibaba's data centers. A year later Nvidia doubled down with an AI acceleration software push alongside Huawei, Lenovo, and all three major Chinese clouds.
The relationship has since inverted. Alibaba began developing its own Ali-NPU neural network chip within two years of this partnership, and by 2025 its GPU pooling system cut the number of Nvidia H20s needed by 82% when serving large LLMs — before moving to open-source its chip software alongside Huawei and Moore Threads to challenge CUDA itself.
First-order effects
- AliCloud gains GPU-backed compute it can sell to Chinese developers and enterprises without building silicon itself, while Nvidia secures volume deployment inside one of China's largest cloud platforms.
Second-order effects
- Rival clouds follow: by late 2017 Baidu and Tencent had joined Alibaba in Nvidia's China partner program, making GPU supply a table-stakes feature across the Chinese cloud market.
Third-order effects
- The customer-supplier dependency proved temporary — Alibaba's path from renting Nvidia GPUs to designing its own chips, pooling them more efficiently, and open-sourcing alternatives to CUDA shows how cloud partnerships seed the domestic capability that eventually displaces the supplier.
- If the pattern holds, compute becomes strategic leverage in China's tech policy: state-backed hard-tech funds and homegrown GPU software stacks point toward a bifurcated AI infrastructure market where Nvidia's platform advantage no longer transfers automatically across borders.
The trend: Cloud providers' early Nvidia partnerships are becoming the on-ramp for China's push toward self-sufficient AI compute, from rented GPUs to domestic silicon and open-source software stacks.