Walk through the final inspection area of a substrate plant and the product looks underwhelming: a square panel, roughly the size of a dinner plate, covered in copper lines too fine to see. Hold it up and the light does not pass through.
That panel is the layer between a chip and a circuit board. It carries power, distributes thousands of signals, and absorbs the mechanical stress of a package that may be a hundred times the size of the die it holds. As AI accelerators grow, this unglamorous slab has become one of the tightest links in the supply chain.
Shennan Circuits (深南电路, 002916.SZ), a Shenzhen company better known for high-end printed circuit boards, is now one of the Chinese firms trying to make it.
A record year built on two different businesses
The 2025 results show a company running two engines at once. Group revenue reached RMB 23.65 billion (≈ US$3.33B / HK$26.0B), up 32%, and net profit RMB 3.28 billion (≈ US$461M / HK$3.60B), up 74.5%.
The first engine is PCBs for AI servers, switches and optical modules: RMB 14.36 billion (≈ US$2.02B / HK$15.8B), up 36.8%, with a gross margin of 35.5%. The second is IC substrates: RMB 4.15 billion (≈ US$584M / HK$4.56B), up 30.8%, at a gross margin of 22.6%.
The oddity is that the more technologically demanding product earns the lower margin. Substrate lines take years to stabilize, and Shennan is still climbing the yield curve on its newest plant in Guangzhou.
What a substrate actually does
A modern processor does not connect directly to a motherboard. It sits on a package substrate, which fans out a pitch measured in tens of micrometres to a pitch measured in hundreds. Memory stacks, power delivery and high-speed links all route through it.
For AI chips, two things change. The package gets much larger, which makes warpage a first-order problem, and the number of routing layers climbs because hundreds of high-speed lanes have to reach the edge without interfering.
That is what the layer count means. Shennan says FC-BGA substrates of 22 layers and below are now in mass production, with 24-layer-and-above products still in development and sampling. Global leaders in this segment are already shipping higher layer counts for the largest AI packages.
Closing that gap is not a matter of buying newer machines. It is a compound problem of fine-line lithography, dielectric materials, warpage control and yield — each of which improves on its own schedule.
The memory detour that paid for the climb
Shennan's substrate business grew faster than its FC-BGA progress alone would suggest, because of memory. As DRAM demand tightened through 2025 and 2026, orders for BT-based substrates used in memory packages rose sharply, lifting utilization at existing plants.
That is a useful reminder about how this industry actually develops. A company funds its entry into the hardest segment by first winning volume in an easier one. Memory substrates are less demanding than processor substrates, but they are produced in enormous quantities and they build the process discipline that later gets applied to finer lines.
The company's Guangzhou plant, which had been a drag, narrowed its losses as BT substrate output ramped. Its Thailand plant and a fourth-phase expansion in Nantong started production in the second half of 2025 and are now in early ramp.
The bill for moving up
R&D reached RMB 1.59 billion (≈ US$224M / HK$1.75B), or 6.73% of sales, with part of it aimed squarely at FC-BGA capability. In April 2026 the company approved a further RMB 4.6 billion (≈ US$648M / HK$5.06B) project in Wuxi for high-layer-count circuit boards — capacity for the board business rather than the substrate business, but funded from the same balance sheet.
The strain is visible elsewhere in the accounts. Outsourced processing costs jumped 79% as internal capacity saturated; management expenses rose nearly 47% on new plant costs; and asset impairment reached RMB 287 million (≈ US$40.4M / HK$316M), much of it inventory.
None of that is unusual for a company building capacity ahead of demand. It does mean that the profit growth of 2025 should not be extrapolated without watching margin at the new plants.
Why this matters beyond one company
China has become the world's largest producer of PCBs, but the substrate layer directly beneath advanced chips has remained a weak point, dominated by Taiwanese, Japanese and Korean suppliers. Every AI accelerator built domestically needs one, and the substrate is often the part with the longest lead time.
Shennan is not the only Chinese entrant, but it is one of the few with a PCB business large enough to fund a multi-year substrate ramp. If 24-layer FC-BGA reaches volume, the constraint shifts from "can it be made in China" to "can it be made cheaply enough and at sufficient yield" — a much more ordinary industrial problem, and a much better one to have.
Key takeaways
- Shennan Circuits (深南电路) grew 2025 revenue 32% to RMB 23.65 billion (≈ US$3.33B / HK$26.0B) and net profit 74.5%, driven by AI-server PCBs and memory substrates.
- Its FC-BGA substrates are in mass production at 22 layers and below, with 24-plus still in development — the layer count is the clearest public marker of how close China is to the advanced-package supply chain.
- Substrate gross margin (22.6%) sits well below PCB margin (35.5%), a direct measure of what climbing the yield curve costs.
Honest limitations
Financials come from the 2025 annual report, its summary and the company's investor-relations record, cross-checked against Securities Times and National Business Daily. The 22-layer figure is the company's own description of mass-production capability; layer count alone does not capture line width, dielectric performance or yield, so it should not be read as a direct comparison with any specific overseas supplier. I have not verified how much of the substrate revenue is memory versus processor product, and the claim that global leaders ship higher layer counts is a general industry observation rather than a sourced comparison. Revenue concentration by customer is not disclosed in the materials I reviewed. The Wuxi project figure comes from the April 2026 board approval as reported by Securities Times, and its final cost may differ.
What readers can do now
- Track substrate layer counts and yield disclosures rather than "advanced packaging" headlines; the substrate is where package ambition meets manufacturing reality.
- For supply-chain risk work, treat IC substrate as a separate line item from PCB — different suppliers, different lead times, different bottlenecks.
- When evaluating a capacity build, compare capex to revenue and to depreciation already booked; that ratio tells you how much of the next two years' profit is already spoken for.