NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

China's chip packagers are adding AI capacity fast — the workaround while lithography lags

Chinese OSATs JCET, TF Microelectronics, and Hua Tian are pouring tens of billions of yuan into advanced packaging, betting that heterogeneous integration offsets what blocked lithography takes away.

2026-11-12 · 1126 words · NeuroAI
China's chip packagers are adding AI capacity fast — the workaround while lithography lags

Everyone watches the lithography machine. The part of China's chip supply chain actually moving fastest right now is the step that comes after the wafer is printed. Across the Yangtze delta, packaging and test houses are breaking ground on fabs whose entire purpose is to squeeze more compute out of chips that can no longer shrink the easy way.

Key takeaways

  • TrendForce counted 13 major advanced-packaging facilities ramping by 2025 globally; several sit in China, including TF Microelectronics' Nantong base (7.5 billion yuan / ≈ US$1.04 billion / HK$8.1 billion), Hua Tian's Jiangsu Pangu Semiconductor project (3 billion yuan / ≈ US$417 million / HK$3.2 billion), SJSEMI (10.09 billion yuan / ≈ US$1.40 billion / HK$10.9 billion), JCET's Jiangsu wafer-level microsystem project (10 billion yuan / ≈ US$1.39 billion / HK$10.8 billion), and WYsemi (2.37 billion yuan / ≈ US$329 million / HK$2.6 billion).
  • China's outsourced assembly and test (OSAT) leaders — JCET (长电科技), TF Microelectronics (通富微电), and Hua Tian (华天科技) — already rank among the world's largest packagers by industry rankings, giving them a head start in scaling advanced packaging.
  • The strategic logic: as Moore's Law slows, AI-accelerator performance increasingly comes from chiplets, 2.5D/3D integration, and heterogeneous packaging — steps less exposed to the strictest lithography export controls.
  • Hua Tian has flagged 2.5D/3D as a strategic priority and set up a dedicated advanced-packaging entity in 2025, with an eSinC 2.5D platform positioned as a domestic counterpart to TSMC's CoWoS, according to industry reports.

Why packaging became the battleground

For two decades the headline race was the nanometer number on the logic die. That race is now both expensive and, for China, partially fenced off: the most advanced exposure tools are restricted. But performance for AI accelerators does not only come from shrinking transistors. It also comes from how you connect them — stacking dies, bridging chiplets, and packing high-bandwidth memory next to logic in the same package.

This is the domain of advanced packaging, and it is where China's existing strengths line up with its constraints. The country already hosts some of the largest packaging and test houses on the planet. Rather than fight the lithography fight head-on, the practical move is to wring more system-level performance out of chips built on nodes China can already produce.

The Chinese capacity wave

The build-out is concrete and large. TrendForce's January 2025 survey of advanced-packaging capacity identified 13 major facilities ramping that year; the China-based entries alone represent tens of billions of yuan in committed investment. TF Microelectronics broke ground on a Nantong advanced-packaging base with total investment of 7.5 billion yuan (≈ US$1.04 billion / HK$8.1 billion), targeting multi-layer stacking, flip-chip, wafer-level, and panel-level packaging, with full production scheduled for 2029. SJSEMI's 3D multi-chip integration project carries a 10.09 billion yuan (≈ US$1.40 billion / HK$10.9 billion) price tag and aims for monthly capacity of 80,000 metal-bump products and 16,000 3D multi-chip integrated packages. JCET's Jiangsu wafer-level microsystem integration project is a 10 billion yuan (≈ US$1.39 billion / HK$10.8 billion) effort whose first phase is designed to produce 6 billion high-end advanced-packaging chips a year. WYsemi is building a two-phase 3D wafer-stacking line, with a first phase investment of 2.37 billion yuan (≈ US$329 million / HK$2.6 billion) intended to establish a domestic 12-inch wafer-level advanced-packaging line. Hua Tian's Jiangsu Pangu Semiconductor project is planned at 3 billion yuan (≈ US$417 million / HK$3.2 billion) with partial production targeted for 2025.

These are not pilot lines. They are volume bets on a specific thesis: that AI and high-performance computing demand will keep pulling advanced-packaging capacity tight worldwide, and that Chinese customers, wary of supply risk, will prefer domestic capacity when it is available.

What advanced packaging actually buys AI

The payoff is density and bandwidth. A chiplet design lets a vendor mix compute, I/O, and memory dies built on different nodes, then link them through a silicon interposer or bridge at micron-scale pitches. That is how modern AI accelerators reach the memory bandwidth their models need without printing the whole thing on the most advanced node. For China, the appeal is double: it reduces dependence on the single most-controlled piece of equipment, and it plays to an OSAT base that is already world-class in scale.

It is also where the "system vs. single tool" logic from China's equipment push meets the "integration vs. lithography" logic of its fabs. Advanced packaging is the integration layer — the place where domestic deposition, etching, and inspection tools, domestic OSAT capacity, and domestic design talent all converge.

According to Li An, chief scientist at BrainNet (脑机网), advanced packaging is where China can realistically close the AI-compute gap fastest, because it depends more on engineering integration than on the single most-controlled piece of equipment. Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).

The limits of the workaround

Packaging is not a free pass. Leading-edge 2.5D/3D still relies on tools and materials — interposers, hybrid bonding, thermal management — where yield and maturity decide whether the economics work. A chiplet built from mature-node dies costs more and draws more power than the same function on an advanced monolithic die; packaging narrows the gap but does not erase it. Export controls have also reached deeper into packaging-adjacent areas over time, so the "safer" assumption should be treated as a moving target.

There is also a customer question. Much of the world's advanced-packaging demand is tied to a handful of AI-chip designers and foundries with established CoWoS and EMIB supply. Chinese OSATs must prove they can hit the yield and reliability bar at scale, not just build the halls.

Honest limitations

This article leans on TrendForce's January 2025 capacity survey for the investment and timeline figures; we have not independently verified each fab's construction status or ramp schedule, and some 2025/2029 dates are company-stated targets that may slip. The Hua Tian 2.5D/3D strategic emphasis and dedicated entity are drawn from a single 2025 industry report and are presented as reported, not confirmed. We did not assess the technical yield or power penalty of specific Chinese chiplet packages, nor model how export-control changes might reshape the packaging equipment supply. Node-specific mass-production claims that appeared in secondary wikis were excluded as unverified.

What readers can do now

  1. Track utilization and yield at the new Chinese advanced-packaging lines, not just groundbreaking ceremonies — capacity announcements say little until they ship paid volume.
  2. Watch whether Chinese AI-chip designers actually route flagship parts through domestic OSATs at scale, which is the real test of the "packaging workaround."
  3. For engineers, learn the chiplet and 2.5D/3D design mindset (UCIe-style interconnects, interposers, hybrid bonding); it is becoming the practical skill set for AI hardware in a constrained supply environment.

Related coverage

Cite this article

NeuroAI (脑机网), "China's chip packagers are adding AI capacity fast — the workaround while lithography lags," 12 November 2026, https://neuroai.site/a/na-chips-osat-advanced-packaging

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Chips & Compute” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories