Wuxi 2026: AI Data Centers Could Drive $4 Trillion in New Infrastructure Investment by 2028

AI infrastructure was the dominant theme at the 2026 Integrated Circuit (Wuxi) Innovation and Development Conference. Executives and officials presented a view of semiconductor growth in which the largest opportunities—and constraints—sit around data-center chips, high-bandwidth memory, advanced packaging, power delivery and thermal management. The headline figures are forecasts and conference assessments, rather than official guarantees. 5

The projected AI infrastructure buildout

Feng Li, president of SEMI China, said investment in newly built global data-center infrastructure could exceed $4 trillion by 2028. She put semiconductor spending within that total at about $2.8 trillion, meaning chips would account for more than half of the projected buildout. 5

That framing matters because it shifts attention beyond AI accelerators alone. Large-scale AI systems require memory, packaging, interconnect, manufacturing equipment, power components and cooling systems alongside compute silicon.

An earlier-than-expected $1 trillion IC market milestone

Bai Peng, chairman and general manager of Hua Hong Grace, said the AI surge had taken the global integrated-circuit industry past $1 trillion in 2026, roughly four years ahead of a previously cited 2030 expectation. He projected the market could reach about $1.6 trillion in 2026. 5

The projection was presented as evidence of an AI-led structural expansion, not simply a conventional semiconductor-cycle recovery. Still, it should be read as an executive outlook: market size and timing will depend on demand, capacity additions and the wider technology and economic environment.

HBM is growing fast, but supply remains tight

High-bandwidth memory (HBM) emerged as one of the clearest bottlenecks. Feng said the global HBM market was expected to grow nearly 60% in 2026 to $54.6 billion, approaching 40% of the overall DRAM market. 5

Even with Samsung, SK hynix and Micron reportedly allocating 70% of new and transferable capacity to HBM, the reported capacity shortfall remained 50% to 60%. 5 The implication is straightforward: expanding AI compute capacity depends not only on securing processors, but also on obtaining the memory and packaging capacity needed to feed them.

Why packaging, power and cooling are central to the next phase

Cao Liqiang, vice president of the Chinese Academy of Sciences’ Institute of Microelectronics, described advanced packaging as a route to continue improving system performance in the post-Moore era. He said hybrid bonding could increase interconnect density from tens of I/O connections per square millimeter in conventional packaging to more than 1 million I/O connections per square millimeter. 5

Cao also pointed to a potential transition from planar three-stage power delivery to vertical two-stage power delivery. As chip power rises, he argued, thermal management must become more aggressive; Nvidia was cited at the conference as already deploying two-phase immersion cooling. 5

Together, these technologies address a common problem: AI performance is increasingly limited by how efficiently chips, memory, power and heat can be managed as a single system.

China’s equipment position and manufacturing ambitions

The conference also highlighted China’s growing role in semiconductor equipment. Feng said Naura ranked fifth and AMEC eighth in a 2025 global top-10 semiconductor-equipment-company ranking. 5

Bai said China’s semiconductor-equipment sector recorded a 57% compound annual growth rate from 2017 to 2025, compared with 26% globally, and expected domestic equipment growth to run two to three times the global average in the coming years. 5 Separately, Feng cited a SEMI projection that China’s share of mainstream semiconductor manufacturing capacity would reach 42% by 2028. 7

These are reported industry assessments, not measures of self-sufficiency across every manufacturing step. But they illustrate the conference’s emphasis on building capability across the equipment, manufacturing and packaging stack.

Jiangsu and Wuxi’s supply-chain response

Jiangsu used the event to launch its IC Industry Alliance, linking companies, universities, research institutions and public technology platforms through a rotating-chair mechanism. 4 The conference also put the Advanced-Process Materials Key Laboratory into operation and began laboratories focused on high-density optoelectronic microsystem integration and advanced substrates, as well as AI-computing power ICs. 49

Five Jiangsu technologies and products were showcased across advanced packaging, RF, high-end inspection, AI computing and high-speed interconnect. They included a 2.5D/3D advanced-packaging R&D platform from Wuxi Huatian/SHJC, silicon-based gallium nitride RF chips, high-resolution electron-beam defect-inspection equipment, an AI-computing platform and interconnect chips. 510

Wuxi reported that it accounted for roughly 70% of Jiangsu’s leading wafer manufacturers, had monthly capacity above 1 million 8-inch-equivalent wafers, and generated more than RMB 280 billion in IC industry output in 2025. It also said it ranked third among mainland Chinese cities in overall global IC-industry competitiveness. 47

Fudan Microelectronics’ FPGA-agent plan

Fudan Microelectronics chairman and general manager Zhang Wei argued that FPGA development could change materially through collaboration with AI agents. He said this model could shorten product iteration cycles from months to weeks or even days, while helping create a tighter development-and-verification loop. 57

The company said it would open trial access to its fmfpga agent FPGA-development platform in September. 5 The initiative reflects a broader conference theme: AI is not only increasing demand for semiconductor hardware, but is also being positioned as a tool for designing and validating that hardware faster.

The key takeaway

The Wuxi conference’s message was not simply that AI will sell more chips. It was that the AI buildout is reorganizing semiconductor priorities around the system-level constraints of memory, packaging, power, cooling, equipment and design productivity. The $4 trillion infrastructure estimate and $1.6 trillion IC-market projection signal the scale speakers expect through 2028—but the HBM shortfall shows why execution capacity will matter as much as demand. 5