When a single chip hits its limit, build "a bigger computer"

At the 26th China International Industry Fair, which opened October 12, the IC exhibition appeared as a standalone zone for the first time, gathering 300+ companies. Per 36Kr-IT Times and Jiwei, the Chinese GPU makers in this "National Chip" zone all chose the same direction: connect more chips together — from competing on single-chip performance to telling how a complete supernode system runs large models.

CIIF supernode scoreboard (cross-verified)

· Iluvatar Tianshui 300 IXL144: 144 GPUs + 18 CPUs in a 44U rack, ~10.1TB total memory, 100-nanosecond-class single-hop latency, +15% effective bandwidth in MoE All-to-All traffic; 128 compute + 16 spare cards, designed for future 100-trillion-parameter models

· MetaX Sjing S600: 64 GPUs fully interconnected per rack via orthogonal connectors and a cable-free architecture; plus a 16-card S600 Lite

· Enflame CloudSailing ESL64-O: 64-card solution co-built with ZTE — Enflame supplies chips, ZTE does rack architecture, integration and delivery

· Reference frame: 128 cards per rack was the recent mainstream (Moore Threads, Biren, Alibaba Cloud; ZTE's OEX supports up to 128); after Ascend 384, Huawei floated clusters of up to a million cards at HC 2026

Why bigger is mandatory

The answer is on the model scaling curve: Moonshot's Kimi K3 already totals 2.8 trillion parameters; at Yunqi, Alibaba said the next Qwen is training with a 50–100 trillion parameter target. Training exchanges gradients and intermediate results relentlessly; even sparse-MoE inference must store and schedule enormous weights; longer contexts keep inflating KV Cache. The single-chip performance story no longer covers the next generation of models.

Who pays: the commercial exam for supernodes

The bigger the system, the sharper the "who pays" question. An IDC report cited by Reuters on October 13: China's data-center AI accelerator shipments reached 2.008M units in H1 2026 — NVIDIA 678K (34%), Huawei 572K (28%), Alibaba T-Head about 7%; GPU shipments fell 26% to 991K while non-GPU accelerators grew nearly 80% to 1.016M. For independent GPU vendors, supply stability, software ecosystems and the ability to bring real business to compute centers decide orders more than peak specs. At the fair, a local compute-center investor's complaint was typical: two years from planning to operation, and the chips have already turned over. Meanwhile, as domestic open-weight models mature, "Chinese chips + Chinese models" on-premise deployment is becoming a genuine incremental market for security-sensitive state-owned enterprises.

Yesterday vendors had to prove "my GPU can run the model"; today they must prove "my supernode can keep producing tokens that cover its cost."

For inference-side scheduling teams like NineZenith's, the exam upgrades too: when the unit of account shifts from per-card utilization to per-rack token cost, scheduling and caching strategies gain value — at the fair, a new venture already sells operations for supernode-native compute systems, focused on multi-level KV Cache management and cache-aware routing.

Trend watch: China's compute race is moving from device specs to systems economics; late this year through next year is the commercialization window for supernodes, and "who delivers dependable token output" decides the next round. (Information synthesized from public reports by 36Kr-IT Times, Jiwei and Securities Times)