---
title: "China's AI Chips Begin on a Wafer Its Own Makers Still Lose Money On"
date: 2026-11-16
category: Chips & Compute
site: NeuroAI
canonical: https://neuroai.site/a/na-chips-12-inch-wafer-gap
language: en
---

# China's AI Chips Begin on a Wafer Its Own Makers Still Lose Money On

> Shanghai's NSIG can polish 850,000 300mm wafers a month and still lost RMB 1.5 billion in 2025 — proof that substrate economics, not lithography, sets the pace.

Picture a disc of pure silicon, 300 millimetres across, polished so flat that a scratch you cannot see ruins it. Before it becomes an AI accelerator, it survives weeks of furnaces, acid baths, and inspection tools. Almost nobody photographs this step. It is also the step where China's chip ambitions are still, quietly, dependent on imports.

## The part of the chip that is not a chip

A silicon wafer (硅片) is not a component. It is the ground floor everything else is built on. Circuits are grown, etched, and deposited onto its surface; the wafer is then diced into hundreds of dies. When a fab says it is "ramping 28nm," it means the line can push a certain number of these discs through without wrecking too many of them.

Diameter matters because area is money. A 300mm wafer has about 2.25 times the usable area of a 200mm one, so the same equipment and cleanroom produce more than twice the chips per run. That is why logic and memory moved to 300mm, and why the shift from 200mm is a one-way door.

The catch is that 300mm wafers are brutally hard to make. The crystal has to be pulled as a single ingot weighing hundreds of kilograms, sawn, lapped, etched, polished, and cleaned to a surface roughness measured in fractions of a nanometre, with strict limits on particles and on how the wafer behaves when heated. Customers qualify a new supplier over nine to eighteen months, because a defect that shows up only at the end of a three-month production run is a very expensive way to learn about a supplier.

For a decade, Chinese fabs bought most of this input abroad. Global supply is concentrated in a handful of houses — Japan's Shin-Etsu Chemical and SUMCO, Germany's Siltronic, Taiwan-based GlobalWafers, and South Korea's SK Siltron — with the Japanese two historically the largest. That concentration is why wafer supply sits in the same risk conversations as lithography.

## What the numbers actually say

Shanghai's National Silicon Industry Group (沪硅产业, NSIG) is the clearest listed proxy for how far the domestic effort has come, and how much it still costs. Its 2025 results, published in April 2026, are a study in that tension:

- Revenue reached RMB 3.716 billion (≈US$516 million / HK$4.0 billion), up 9.69% year on year.

- Net loss attributable to shareholders of RMB 1.508 billion (≈US$209 million / HK$1.6 billion).

- The 300mm segment brought in RMB 2.439 billion (≈US$339 million), up 15.8%, and now accounts for roughly two-thirds of revenue.

- Combined 300mm capacity in Shanghai and Taiyuan reached 850,000 wafers per month by the end of 2025, with 300mm volume up about 27% year on year.

- The 200mm-and-below segment earned RMB 1.125 billion (≈US$156 million), but its gross margin fell to minus 23.37%.

The shape of the problem appears in the margins. The strategic product is growing, the legacy product is bleeding, and the company lost money on both: 300mm gross margin for the year was minus 14.99%.

## Why growing revenue can still mean a bigger loss

Three things happen at once when a wafer maker ramps.

First, depreciation arrives before revenue does. A polishing line is expensive, and the accounting does not wait for customer qualification. Shanghai Xinsheng (上海新昇), the established 300mm plant, ran at high utilization and shipped more; the newer Taiyuan site, operated through Jinke Silicon (晋科硅材料), only began volume sales of prime wafers during the year. Fixed costs are front-loaded, so early output is expensive output.

Second, prices fell. Wafer prices drop faster than fab utilization and recover more slowly. The company also booked impairment charges in 2025, including goodwill and inventory write-downs — a marker of how weak the 200mm market for consumer and industrial chips remained.

Third, the mix is not yet where the money is. The most demanding wafer types — ultra-low-resistivity substrates for high-power devices, high-specification logic wafers for high-performance computing, dToF wafers for radar and physical AI — are the ones that carry pricing power, and they are precisely the ones still being qualified. NSIG says it is developing all of them. Development is not the same as a purchase order.

## The AI connection is real, but indirect

AI does not consume wafers the way it consumes electricity. It consumes them through three doors: logic dies for accelerators and CPUs, DRAM and HBM stacks, and the power and analog chips around them. When those three lines run hot, wafer demand follows with a lag of a quarter or two, and the mix shifts toward the largest, flattest, cleanest discs.

That lag explains a puzzling pattern in the 2025 results. AI demand was the strongest part of the semiconductor market, and NSIG's 300mm revenue grew — but not enough to cover the cost of the capacity built to serve it. A wafer plant is a bet placed two to three years before the demand shows up, and the bill for the bet arrives first.

There is also a structural gap that revenue growth does not close. Even with domestic 300mm capacity rising, the highest-grade substrates, silicon-on-insulator (SOI) wafers, and specialty specifications remain largely imported. NSIG's 300mm SOI pilot line, built through its New Alliance (新傲科技) subsidiary, is rated at roughly 160,000 wafers a year and is still moving through customer validation and small-batch production for radio-frequency, power, and silicon-photonics applications. That is a laboratory-to-factory transition, not a market.

## Honest limitations

- **This piece leans on one company.** NSIG is the most transparent listed wafer maker in China, but it is not the whole domestic industry; unlisted suppliers and the 200mm specialists are not covered here in comparable detail.

- **The financial figures come from the company's 2025 annual report as summarized by broker research**, not from an independent audit I performed. Segment gross margins are as reported and are affected by depreciation policy and impairment timing.

- **Localization rates are deliberately absent.** Public estimates of how much of China's 300mm wafer demand is met domestically vary widely by grade and by who is counting, and I could not verify a defensible figure.

- **I have not modelled wafer pricing.** Without a reliable public price index for 300mm prime wafers, the margin discussion is descriptive rather than predictive.

- **Global share is described qualitatively.** The five-supplier concentration is well established, but precise 2025 shares by supplier were not verifiable from a primary source here.

## What readers can do now

- **If you follow semiconductor earnings, stop reading revenue growth first.** For a ramping materials company, segment gross margin and utilization tell you where in the cycle the business sits; revenue tells you only that somebody is buying.

- **If you are building hardware, treat the substrate as a long-lead item.** Wafer qualification takes nine to eighteen months. Teams that plan around the chip alone discover the schedule risk one layer too late.

- **If you are watching China's chip self-sufficiency story, look past the fab announcements.** Fabs are the visible half. The wafer, the photoresist, the electronic gas, and the metrology tools are where the quiet dependency lives — and where the losses are currently being booked.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-12-inch-wafer-gap
Free to quote with attribution and a link to the original.
