china's ai revolution hits a snag: gpu shortage threatens hyperscalers
Beijing, Thursday, 24 July 2025.
chinese hyperscalers are facing a critical shortage of gpus, essential for ai training. nvidia’s h800 and h20 chips are in high demand, but export restrictions are causing scarcity. this forces companies like alibaba and tencent to seek alternative solutions and chinese domestic ai chip suppliers are predicted to gain market share, reaching 55% by 2027. the ai chip market in china is expected to reach $38 billion in 2025, an 85% increase year-over-year.
export restrictions and market impact
U.S. export controls restrict Chinese companies’ access to high-performance GPUs [1]. These restrictions and high ai workload demand are forcing tech companies to rethink their strategies [2]. H3C, a server manufacturer and nvidia oem partner, reports its h20 inventory is nearly depleted, with shipments delayed until mid-2025 [2]. Nvidia’s h20 chip, designed to comply with u.s. export restrictions, performs approximately 70% as well as the original h100 [3]. Market analysts report mixed reviews on the h20’s performance in specific ai workloads [3].
domestic alternatives and ai infrastructure
Chinese hyperscalers are adopting cpu-optimized solutions from domestic providers like feiteng, kunpeng, and hygon [2]. The ‘made in china 2025’ policy mandates 70% domestic chip usage in public-sector data centers [2]. Feiteng and hygon have shipped 10 million units combined [2]. Total ai capital expenditure in china is projected to reach 655 billion yuan ($91 billion) in 2025, a 51% increase [3]. This shift highlights the need for chinese companies to balance high performance with export restrictions [1].
nvidia’s strategic moves and challenges
Nvidia plans to release updated versions of the h20 in q4 2025 to improve performance in key ai tasks [3]. However, tsmc production line adjustments have led to h20 chip supply shortages, potentially delaying revenue inflow for nvidia [5]. Relaunching the h20 production line could take at least nine months [5]. Nvidia might sell b30 chips in china, potentially generating $2.8 billion in sales from 400,000 chips [3]. Nvidia’s cfo indicated that china will transition from the hopper architecture to the blackwell architecture [3].
hyperscaler strategies and market dynamics
ByteDance, alibaba, tencent, and baidu are the main purchasers of h20 chips, accounting for over 80% of total demand in 2024 [3]. Alibaba cloud is expanding into southeast asia, opening a global ai capability center (aigcc) in singapore and a third data center in malaysia [2]. The ai infrastructure market in china is projected to grow at a 22.3% cagr [2]. However, the h20 chip shortage has intensified, with prices increasing by up to 20% recently [6]. Several chinese ai companies are experiencing project delays due to chip shortages [6].
investment landscape and future outlook
Investors are advised to focus on domestic cpu providers like huawei and feiteng, cloud infrastructure leaders like alibaba cloud and tencent, and semiconductor enablers like smic and xingsen technologies [2]. The ‘east data, west compute’ initiative and the rise of open-source llms signal a future where china’s ai sector continues to thrive [2]. A new government initiative to subsidize domestic chip manufacturers is planned [6]. These developments suggest that overcoming challenges and integrating domestic solutions will define china’s ai development [1].