Huawei used its Shanghai summit to push a simple message: AI networks, not just chips, will decide who scales. The company’s Xinghe AI Fabric 2.0 packages switches, optics, orchestration software, and liquid-cooled racks into an end-to-end data center fabric aimed at large AI clusters. It is a product launch and a policy signal. Under tightened U.S. controls and slowing access to foreign interconnects, China’s buildout is shifting to domestic Ethernet at 400G and 800G. The question now is execution at scale: can Huawei’s fabric deliver performance close to Nvidia’s Infiniband while meeting China’s energy and cost targets.
Strip out the show-floor superlatives and the positioning is clear. Huawei wants to own the AI plumbing, from Ascend accelerators and MindSpore software to CloudEngine and XH switches, StarryLink optics, and an AI-aware scheduler. The new fabric layers “AI Brain” automation tools over “AI Connectivity” and “AI Network Elements,” with claims of 95 percent network throughput and a tenfold reliability improvement. These are vendor numbers; proof will come from live clusters, not testbeds. But the stack matters. Chinese state media have framed computing power as a national capability, tied to “new quality productive forces.” At Shanghai’s World AI Conference, Nvidia was absent; Huawei’s Ascend took the prime booth. The open-sourcing of Pangu models is designed to pull developers onto a domestic stack and, by extension, Huawei’s hardware. In a market where systems integration risk has risen, a single-vendor fabric is a commercial answer to a geopolitical problem.
The more interesting shift is architectural. China’s internet platforms historically leaned on Nvidia’s Mellanox Infiniband for top-end AI training because of its low latency and mature software. Export controls and supply tightness have hardened the case for high-speed Ethernet with RDMA. Huawei is betting that 800GE Ethernet, tailored with lossless features and load balancing algorithms, can close enough of the gap for most training and almost all inference. Its XH-series 800GE switches and optics are the spearpoint. Domestic research institutes and the Ministry of Industry and Information Technology have promoted “computing power networks,” with standards pushing toward interoperable, software-defined Ethernet fabrics. The catch is the stack: congestion control, telemetry, and reliability in Ethernet at 800G are hard to perfect outside lab conditions. Even Chinese tech press concede Nvidia still sets the pace at cluster scale. If Huawei’s NSLB and “rock solid” redundancy work as advertised in production, it would move the cost-performance frontier for state clouds and private AI parks.
Beijing has been laying the groundwork. The 14th Five-Year Plan and follow-on digital economy plans call for new infrastructure and data center upgrades aligned with dual carbon goals. The “East Data, West Computing” initiative encourages compute to flow to inland provinces with cheaper land and power, stitched to coastal demand via fast networks. Regulators have set tighter benchmarks for energy efficiency in data centers, pushing power usage effectiveness down and nudging operators toward liquid cooling. Huawei’s liquid-cooled 400G and 800G gear slots directly into this policy lane. Procurement rules in sectors like telecom, finance, and energy favor secure and controllable tech, where domestic vendors score higher on supply security. The result is a captive early market: provincial computing hubs, SOE clouds, and regulated industries are the likely first buyers of a Chinese 800GE AI fabric—especially if it lowers integration risk and simplifies compliance.
Hardware is only half the story. Huawei is trying to make its network fabric “AI-native” with an orchestration layer that maps jobs to resources, automates provisioning across security domains, and visualizes bottlenecks. That is an attempt to bind network and workload. The company’s open-source posture with Pangu, and the continued investment in MindSpore and the Ascend toolchain, are meant to reduce the friction for developers who might otherwise default to CUDA-centric workflows. But openness here is bounded. China’s operators run mixed estates with legacy IB islands, white-box switches, and multiple security zones. Claims of “end-to-end automation over heterogeneous networks” will be tested in messy environments with compliance constraints. Real-world stickiness will depend on how well Huawei’s fabric coexists with non-Huawei elements, not just how elegantly it runs in an all-Huawei rack.
Procurement teams will run the total cost math. In large AI clusters, the network can account for 20 to 30 percent of capex when optics are included. At 800G, the optics bill is often the swing factor. Huawei’s push to localize 800G modules and deploy co-packaged or low-power solutions is about cutting that curve. If domestic optics and switches deliver consistent yields and power budgets, operators can scale without importing expensive modules or burning through energy quotas. The company is pairing this with liquid cooling to meet provincial energy caps. Against this, white-box Ethernet with domestic optics remains a price pressure point, and H3C and others will not cede share quietly. Under export constraints, Cisco and Arista are marginal in China’s AI buildout, but software maturity and ecosystem depth still matter. A single-vendor fabric may reduce integration headaches, yet it can create lock-in premiums over time; buyers will push for benchmarked performance and transparent lifecycle costs.
Three risks are hard to ignore. First, export controls remain dynamic. Even if AI networking is less restricted than accelerators, advanced DSPs for 800G optics, high-end FPGAs, and certain EDA flows for SerDes can get caught in new rounds. Building a resilient supply chain for lasers, modulators, and PAM4 chips at volume is still a grind. Second, performance credibility. Vendor claims of 95 percent throughput and 10x reliability gains need third-party validation at thousands of nodes under realistic congestion. Without that, operators will cap cluster sizes or keep Infiniband where latency spikes hurt model convergence. Third, software gravity. Despite progress, Nvidia’s ecosystem lock still pulls. Chinese firms are narrowing the gap, but catching up in networking telemetry, congestion algorithms, and scheduler integration is a multi-year project. The market will gauge progress in deployments, not conferences.
A few signals will show if AI Fabric 2.0 is more than a showroom demo. Look for named wins at China’s top cloud providers and SOEs with cluster sizes above 2,000 accelerators running production training. Watch MIIT-backed standards on computing power networks and liquid cooling move from guidance to compliance and whether Huawei’s designs map cleanly to those specifications. Track domestic 800G optics volumes and failure rates; if yields improve and prices fall, TCO tilts in Huawei’s favor. Finally, check whether major internet platforms decommission or cap Infiniband in new builds. If Ethernet-based fabrics dominate new AI parks under East Data, West Computing, Huawei’s bet pays off. If not, expect hybrid estates and cautious scaling. The race is less about headline chip TOPS and more about the steady, expensive work of wiring China’s AI era. That is the niche Huawei is trying to own.
