China's AI Super-Node Computing Power Accelerates Expansion, 2026 Market Size Estimated at Nearly 90 Billion Yuan

Nashnova编辑部
Published 2026-08-15About 13 min read

China's smart-compute capacity surged 177% year-on-year to 2,185 EFLOPS in H1 2026, as super-node architecture — clustering thousands of accelerator cards into one compute domain — shifts from proof-of-concept to volume delivery. Guohai Securities estimates the market at ¥88.9 billion in 2026, crossing ¥1 trillion by 2028.

01

What exactly is a super-node, and why does it matter now?

A super-node links hundreds or thousands of accelerator cards via high-speed interconnects and unified memory pooling into a single compute domain — essentially turning a rack farm into one giant machine.
This means → China's domestically made AI chips still trail global leaders on a per-card basis. The super-node thesis is scale for efficiency: more cards, tighter links, shared memory — closing the gap at the system level rather than the chip level.
In plain terms = if one card can't beat the competition, wire a few thousand together and compete as a single system.
02

How big could this market get?

Guohai Securities projects super-node market size at ¥88.9 bn in 2026 → ¥415.1 bn in 2027 → ¥1.109 tn in 2028. Super-nodes' share of domestic AI chip shipments is expected to rise from 15% in 2026 to 55% in 2028.
Huatai Securities is more conservative, pegging 2028 at roughly ¥341.4 bn — a threefold gap between the two forecasts.
This reflects an industry still in early ramp: brokerages disagree sharply on penetration speed, and the actual trajectory hinges on whether systems can ship at volume.
03

Who is building them, and how far along are they?

Sugon scaleX640: 640 accelerator cards in a single rack, with claimed compute density 20× peers. Sugon 8000 is China's first fully domestic 100,000-card AI supercomputing cluster connected to the National Supercomputing Internet.
Huawei Atlas 950 SuperPoD: scales from 64 NPUs per rack to a maximum of 8,192.
Alibaba Pangu AL128 integrates 128 cards; Baidu Tianchi 256 integrates 256. ZTE and Moore Threads have also shipped products.
This means → super-nodes are no longer one vendor's concept rig — at least five or six vendors have entered the delivery phase.
04

Where is the infrastructure money coming from?

The *Economic Information Daily* cites estimates that compute-network buildout during the "15th Five-Year Plan" period could drive roughly ¥4 trillion in new direct investment.
Project-level examples: GDS (万国数据) committed over ¥30 bn across five years for a high-density data center in Ulanqab; Jinko Power is involved in a 1 GW compute-center project in Zhongwei, Ningxia, with an investment of about ¥24.5 bn.
In the first seven months of 2026, over 150 publicly tendered smart-compute and data-center projects each exceeded ¥100 million in contract value — this reflects capital expenditure moving from plans to actual procurement.
05

How much are telcos and internet giants spending?

China Mobile's Wuhan smart-compute center has deployed over 700 racks with 2,876 PFLOPS of installed capacity. H1 2026 capex reached ¥61 bn, with compute-network investment up 71% YoY. AI-storage synergies lifted long-sequence token output efficiency by 50%.
China Unicom's Wuzhen center launched Phase 1 at 2,000 PFLOPS; the Wujiang project secured ¥5.1 bn in approved investment. Full-year capex budget is ¥50 bn, of which over ¥17.5 bn (35%) is earmarked for compute.
Tencent Q2 2026 capex hit ¥52.78 bn — up 176% YoY and 65% QoQ. Lenovo's AI-related revenue rose 60% YoY to ¥63.4 bn in the quarter ending June 2026.
06

Can this market deliver on the forecasts?

Super-nodes currently account for roughly 30%–40% of compute procurement at major internet companies. Broader industry penetration is expected to begin in 2027.
This means → the market is still in an early-adopter phase led by top-tier buyers; the real volume window has not yet opened.
Two make-or-break checkpoints: whether leading systems can sustain volume delivery, and whether domestic chip interconnect performance can keep up with expanding system scale. In plain terms = can the cards be manufactured fast enough, and can the "roads" between them be built wide enough — that determines whether the trillion-yuan forecast materializes or stays on paper.

Content is for reference only, not financial advice.