China's Chip Design Sales Projected to Surge 89% by 2026, but AI System Bottlenecks Constrain Scalable Deployment

nashnova research
今天发布阅读约 11 分钟

China's chip-design industry is on track for ¥1.58 trillion in 2026 revenue, up 89.4% year-on-year, but Wei Shaojun warns that stacking more chips cannot substitute for fixing system-level bottlenecks — the real test of whether this growth translates into lasting competitiveness.

01

Where does the ¥1.58 trillion come from?

Wei Shaojun, chair of the IC Design Branch of the China Semiconductor Industry Association, disclosed at the ICCAD-Expo Beijing forum that chip-design revenue is projected at ¥1.58 trillion in 2026, up 89.4% year-on-year, spanning roughly 4,324 firms across 20 cities.
Growth is heavily top-loaded. The ten largest advanced-chip companies are expected to generate nearly ¥800 billion combined — a 218% increase that far outpaces the industry average.
This means → roughly half of total industry revenue comes from ten companies. The gains from scaling are not evenly shared.
02

Hefei's growth hit 2,546% — how can one city move that fast?

Hefei leads the country with a projected 2,546.7% surge; Wuhan jumps to fifth at 218.2%, overtaking Wuxi. Shanghai and Shenzhen rank second and third.
Wei attributes the spikes in Hefei and Wuhan to the rise of China's memory-chip sector, which he estimates at ¥500–600 billion in annual revenue.
Yet ICviews data shows Hefei has only 27 firms with revenue above ¥100 million, far fewer than Shanghai's 150. In plain terms = one or two giants can make a city's headline number explode, but that is not the same as building the layered ecosystem Shanghai or Shenzhen has.
03

What are the three system bottlenecks facing AI chips?

Wei identifies three core constraints on China's AI accelerator chips: compute performance, memory bandwidth, and chip-to-chip interconnect. Restricted access to advanced process nodes and HBM — high-bandwidth memory, a technology that stacks memory dies vertically to boost data throughput — plus gaps in packaging and interconnect compound the challenge.
In large multi-chip configurations, communication overhead alone can cut overall system performance by 30–40%, even when each processor's theoretical compute power is high.
This means → plugging in more chips ≠ linear gains in compute. The cost of chips "talking to each other" deeply discounts on-paper performance.
04

If advanced process nodes are out of reach, what alternatives remain?

Wei argues that manufacturing constraints do not close the door on AI chip performance gains. Advanced packaging, heterogeneous integration, and software optimization can partially offset the absence of leading-edge nodes.
The core challenge is to raise compute efficiency and interconnect bandwidth while cutting system overhead — not to brute-force performance by adding more chips. AI-assisted EDA tools — software that uses AI to help design chips — can also accelerate development and explore new architectures.
In plain terms = when the front gate is blocked, you look for side doors — better ways to assemble chips and smarter software — rather than keep pushing at the lock.
05

An 89% growth rate — what is it actually worth?

Wei also warns that supply-chain security can matter more than technological advancement in certain scenarios — this reflects a structural fragility beneath the headline expansion.
He stresses that semiconductor competitiveness no longer resides in any single company but in the full ecosystem — manufacturing, packaging, memory, interconnect, and software working together. Chip-design engineers must evolve into system engineers.
This means → the revenue surge validates scale, but whether growth converts into durable technological competitiveness depends on domestic suppliers breaking through system-level bottlenecks and shoring up the supply chain. Big numbers and real capability are two different things.

市场有风险,内容仅供研究参考,不构成投资建议。