AI servers have propelled a particular fabric to the forefront of the tech industry.

In the first quarter of 2026, Jensen Huang, the CEO of NVIDIA, took a personal trip to Japan to visit Nitto Boseki, the global leader in electronic fabrics. Meanwhile, Apple has increased its presence in Japan, stationing more personnel there for negotiations. Qualcomm, too, has had to navigate through supply chain hierarchies to actively seek alternative resources. Google, Amazon, and Microsoft are also deeply involved in this competitive landscape.

A thin glass fiber fabric, once confined to the server room, is now making its way back into the spotlight of textile workshops.

Initially, the bottleneck for AI servers was centered around GPUs. However, the focus has now shifted to connectivity as the most critical aspect.

Fast chips necessitate rapid data movement. Any obstruction between chips, servers, or switches can severely limit computing power.

Optical communication has emerged as a key solution. In 2026, expectations for shipments of 800G optical modules have been raised to over 20 million units, with 1.6T module shipments projected to approach 10 million units. Revenues from data communication optical components surpassed $19 billion in 2025, marking a more than 70% increase from 2024.

In March 2026, NVIDIA made significant investments, pouring $2 billion each into U.S. photonics companies Lumentum and Coherent, along with multi-year procurement commitments and future capacity stakes. Where investment flows, the bottleneck lies.

This sector is already heating up, but it's just the beginning.

Regardless of the speed of optical signals, they ultimately need to be received and processed on circuit boards.

AI server motherboards, switch boards, accelerator cards, and packaging substrates all handle high-speed signals. The higher the frequency, the more stringent the material requirements become. Higher losses can lead to signal attenuation, while uneven fabric surfaces can cause transmission delay jitter.

Copper-clad laminates are produced by pressing glass fiber fabric, copper foil, and resin together. In terms of cost structure, glass fiber fabric accounts for approximately 25% to 40% of the total cost of copper-clad laminates. This material, which has remained under the radar for decades, has now become a key factor in determining the lead times and costs of AI servers.

The grade differences in glass fiber directly translate into price disparities. Ordinary E-grade glass fiber serves as a basic material, while NE-grade low-dielectric glass fiber has entered the realm of high-speed communication materials. NER-grade low-dielectric glass fiber is even more specifically tailored for high-speed signal scenarios in AI servers and data centers. The average price of NE-grade glass fiber is about six times that of ordinary E-grade glass fiber, and NER-grade glass fiber is 2.5 times that of NE-grade glass fiber. The closer a material is to the high-speed computing center, the less it resembles an ordinary industrial product in terms of pricing.

Currently, Nitto Boseki holds a dominant position in the global market, with about 90% of the market share for T-glass and 60% to 70% for NER-glass. As AI data centers expand at a rapid pace, the supply of high-end electronic fabrics becomes increasingly tight. Nitto Boseki's new capacity is not expected to significantly alleviate supply pressure until mid-2027 at the earliest.

Nitto Boseki raised prices for its glass fiber products by 20% in August 2025 and plans another increase of about 20% to 30% in April 2026. Price increases for BT substrates will be reflected about a quarter later, and for ABF substrates, about two quarters later.

Although material companies typically operate with a low profile, their bargaining power becomes extremely strong when demand suddenly surges.

Weaving electronic glass fiber fabric requires precise control over tension, speed, flatness, and defect rates. Domestic mainstream electronic fabric manufacturers are highly reliant on Japanese imported looms. A research report by China Merchants Securities points out that the delivery cycle for high-end looms like the Toyota JAT910 is as long as 18 to 24 months, with orders already booked until 2027. The China Glass Fiber Industry Association, in collaboration with the China Textile Machinery Association, has held a special seminar to promote the domestic substitution of high-end weaving equipment.

The pace of future server development is now closely tied to the lead time of looms.

Currently, Hongfa Technology, a domestic company, holds the second-largest global market share for ultra-thin electronic fabrics. Its low-dielectric products have been certified by customers such as Doosan Electronics, Taiguang Electronics, Panasonic Electronics, and Taiyao Technology. As of the end of the first quarter of 2026, the average selling price of electronic-grade glass fiber fabrics represented by Hongfa Technology rose to 9.78 yuan per meter, marking a year-on-year increase of 116.9%. China Jushi is the global leader in the glass fiber industry, with a large-scale production capacity for electronic fabrics covering both ordinary and high-end segments. Its low-dielectric yarn technology has been certified by NVIDIA, benefiting from the growth in AI computing demand. Sinoma Science & Technology is the domestic leader in all categories of specialty electronic fabrics, with low-dielectric, low-expansion coefficient electronic fabrics certified by leading customers such as NVIDIA, and is expanding its production capacity for high-end electronic fabrics.

Dongmai Technology, however, takes a different approach by focusing on high-frequency, high-speed resins. Its M9-grade hydrocarbon resin is the only domestically produced product and one of only two globally certified by NVIDIA. It is already being used in high-end AI servers such as GB300 and Blackwell. In 2025, revenues from high-speed electronic resins reached 591 million yuan, up 125% year-on-year, and in the first quarter of 2026, revenues reached 258 million yuan, up 131% year-on-year.

Looking further upstream, we come to the equipment sector. Titan shares is positioning itself in the high-end loom market. The potential for domestic equipment substitution is significant, but true volume growth depends on factors such as stability, yield rates, customer validation, and sustained orders. High-end electronic fabric manufacturers will only switch equipment if it leads to better yield rates and shorter lead times.

Profits in the AI supply chain are shifting from bustling segments to those with scarce resources.

Customer certification is a slow process, and replacement costs are high. Expansion cycles are long, and yield ramp-ups are gradual. Global supply is concentrated, giving suppliers strong bargaining power when demand surges. High-end lasers, specialty electronic glass fiber fabrics, low-loss resins, high-end copper foils, precision looms, and process databases—each may be small in scale, but each can significantly impact the pace of the entire supply chain.

Electronic fabrics have emerged as a strategic material in the AI computing era.

Global PCB production value reached $85.152 billion in 2025, up 15.8% year-on-year. It is expected to rise to $95.8 billion in 2026, up 12.5% year-on-year. Goldman Sachs predicts that the global market size for AI server PCBs will expand dramatically from $3.1 billion in 2024 to $27.1 billion in 2027.

It may be difficult for the average person to care about T-glass or the weave of electronic glass fiber fabrics.

However, in the coming years, whether AI computing costs continue to decline, whether domestic AI servers can achieve stable volume production, and whether the price of enterprise AI usage can become more affordable will all hinge on these underlying materials.

When technological revolutions unfold, they always pass through factories. Chips, optical modules, PCBs, glass fiber fabrics, copper foils, resins, looms—one link after another. If any link fails to keep up, the costs will ultimately be passed on to the end user.

The most cutting-edge models must ultimately rest on the most concrete industrial details.

Some of the images in this article were generated by AI. Unless otherwise specified, all data in the article comes from publicly available market research reports. If any references are inappropriate or infringe on rights, please contact us for removal. The content is for reference and learning purposes only and does not constitute any investment advice.