Shenzhen Activates China's Largest Homegrown AI Cluster — and the Numbers Are Staggering

Shenzhen Activates China's Largest Homegrown AI Cluster — and the Numbers Are Staggering

China has quietly crossed a milestone that matters far beyond Shenzhen's city limits. In late March 2026, the southern technology hub switched on its first 10,000-card AI computing cluster built entirely with Huawei's Ascend 910C chips, delivering 11,000 petaflops of processing power. It is a direct answer to US export restrictions — and a signal that China's AI infrastructure ambitions are accelerating faster than many analysts predicted.

The Scale of the Cluster

The new installation is the second phase of Shenzhen's intelligent computing buildout. The first phase, a 3,000-petaflop cluster activated in 2025, had already been fully booked since it went online. Together, the two phases now deliver a combined 14,000 petaflops of computing capacity available to local firms, universities, and research institutions.

Demand has been fierce. Nearly 50 organisations signed framework agreements to access the new cluster before it launched, producing a combined booking rate of 92% across both phases. The tenant roster covers AI startups, robotics companies, and universities — a cross-section that illustrates how broadly Chinese organisations now depend on sovereign compute^✦ infrastructure.

Zhang Luncheng, vice-president of robotics startup X Square Robot, was among those celebrating the expansion.

Significant upgrades to the scale and quality of Shenzhen's computing power had positioned the city as a national leader.

Why the Ascend 910C Matters

Huawei's Ascend 910C is not a straight replacement for Nvidia's H100. A widely cited DeepSeek study found the chip operates at roughly 60% of the H100's capacity on standard benchmarks. Yet for most Chinese AI workloads — fine-tuning^✦ open models, running inference^✦ at scale^✦, training robotics systems — that gap is manageable, particularly when the alternative is acquiring restricted chips at a premium through unofficial channels.

The Shenzhen cluster's architecture sidesteps that entire problem. By building entirely on domestic silicon, operators gain predictability, local support, and freedom from export-control uncertainty. The 92% booking rate suggests the market agrees: 60% of an H100 delivered reliably beats 100% of a chip that may not arrive at all.

The Geopolitical Backdrop

This development does not exist in isolation. Huawei and its domestic rivals have been steadily absorbing market share as US export controls restrict Nvidia's access to Chinese buyers. A Reuters analysis published on 1 April found that Chinese GPU^✦ and AI chip manufacturers collectively captured 41% of China's AI accelerator server market in 2025, with Nvidia's share falling to 55% from a near-monopoly position just two years prior.

Shenzhen's cluster is the physical manifestation of that shift. It demonstrates that 10,000-card deployments — once synonymous with hyperscaler^✦ infrastructure in the US — are now achievable with Chinese silicon. As China's domestic chipmakers continue to take market share from Nvidia, the economics of sovereign AI^✦ compute are becoming harder to ignore.

The story also connects to broader infrastructure investments across Asia. Microsoft's $10 billion AI investment in Japan and Microsoft's $5.5 billion Singapore push represent the Western approach to AI infrastructure expansion in the region — anchor investments by Western hyperscalers. China's approach is different: building publicly accessible compute hubs powered by homegrown chips and targeted at local industrial demand.

What It Means for Asia's AI Race

The Shenzhen activation matters for three reasons beyond the numbers themselves.

First, it validates the "domestic first" compute strategy at scale. Previous homegrown clusters were pilot deployments. A 14,000-petaflop facility with near-full occupancy is a commercial-grade operation.

Second, it accelerates the robotics sector. Among the organisations booking capacity are robotics firms training embodied AI systems — a sector Beijing has designated a national priority. Southeast Asia's data centre buildout is being driven by AI energy demands, but China is solving a different version of the same problem: not energy scarcity but chip scarcity.

Third, it sets a template for other Chinese cities. If Shenzhen can achieve 92% occupancy at 14,000 petaflops with domestic chips, other technology hubs — Beijing, Hangzhou, and Chengdu — will accelerate similar buildouts.

A Test of Domestic Demand

The one open question is performance at the frontier. Huawei's Ascend chips are capable enough for inference and fine-tuning, but training the next generation of foundation models from scratch still benefits from Nvidia's raw throughput. China's largest AI labs have found creative workarounds — distributing training across many more chips — but it remains a genuine constraint.

G42 and FPT's $1 billion Vietnam AI deal shows that infrastructure ambitions across Asia are no longer confined to China. Other nations are investing heavily in sovereign compute. China's edge, for now, is that it has domestic chips to fill its own clusters.

The bottleneck is no longer policy or capital — it is the chips themselves. Whoever controls the compute controls the AI roadmap.