WAIC 2026: AI Computing Competition Shifts from Chips to Efficiency, Domestic Full-Stack Solutions Accelerate Deployment

From July 17 to 20, 2026, the World Artificial Intelligence Conference (WAIC 2026) was held in Shanghai. This year's conference sent a clear signal: AI computing competition is shifting from single-chip parameter battles to system-level efficiency, full-stack collaboration, and ecosystem openness. Multiple vendors showcased solutions such as supernodes, heterogeneous hybrid inference, direct GPU connectivity, and token optimization factories, aiming to address the structural challenge of 'strong individual chips but weak systems' in domestic computing.

Supernodes Emerge as New Computing Infrastructure

Huawei, ZTE, Baidu AI Cloud, Enflame Technology, and others exhibited supernode products. ZTE released the OEX orthogonal architecture supernode, integrating 128 GPUs per cabinet with zero-cable, backplane-free design, and jointly launched the Matrix supernode with Lightelligence and Biren Technology, winning the WAIC SAIL Star award. Baidu AI Cloud's Tianchi supernode showcased 256/512 configurations, achieving 25% throughput improvement and 50% inference efficiency gain, with adaptation to mainstream models like Wenxin and DeepSeek.

Breakthrough in Direct GPU-RDMA Connectivity

Qimoer and Biren Technology jointly demonstrated the IBGDA solution, enabling direct connectivity between domestic GPUs and domestic RDMA network cards. Real-world tests showed significantly reduced All-to-All communication latency and improved small-packet bandwidth, eliminating the 'small-packet penalty' of traditional solutions. Based on the Ethernet ecosystem, this solution provides a high-performance, low-cost Scale-Out network option for domestic computing clusters.

Heterogeneous Hybrid Inference and Token Optimization Key to Efficiency

SenseTime's SenseCore launched a heterogeneous hybrid inference solution, separating Prefill and Decode phases—using high-end resources for bottleneck stages and domestic chips for throughput tasks—boosting MFU of mainstream domestic chips by 85%-152% and increasing token output per unit cost by 2.5x. Shishi Technology released the 'Tuoyuan' Token Optimization Factory, compatible with over 10 domestic chips, achieving daily throughput of hundreds of billions of tokens, reducing inference costs through KV cache compression and long-context optimization.

Full-Stack Self-Development and Ecosystem Openness

Baidu AI Cloud emphasized a 'new full-stack AI cloud,' forming a closed loop from Kunlun chips and Tianchi supernodes to the Baige platform and Wenxin large models, driven by real business loads for continuous optimization. Cix Technology unveiled the AGX Agentic Compute strategy, launching the desktop supercomputer AGX Station capable of local inference on 70B-150B parameter models, with support for hot-swappable multi-brand AI accelerators. TsingMicro differentiated with reconfigurable computing architecture, launching a 4K supernode that reduces interconnection costs by approximately 90% and supports over 200 models through the RAISA software stack.

Computing-Power Synergy and Optical Interconnect for the Future

ZTE introduced an 800V high-voltage DC solution and liquid cooling, supporting over 200kW per cabinet with PUE as low as 1.15. SenseTime's SenseCore released a 'Computing-Power Synergy Agent,' increasing token output per unit electricity cost by 80%. Qimoer jointly published the 'Co-Packaged Optics (CPO) Technology White Paper' and deployed OSU IO chiplets, exploring the evolution of optical interconnect from board-level to chip-level.

Overall, domestic AI computing is transitioning from 'chips being usable' to 'systems being easy to use,' with full-stack collaboration, efficiency-first, and ecosystem openness becoming the core themes of the next phase of competition.