T-Glass Shortage Threatens AI Server PCB

The Hidden Bottleneck in AI Infrastructure

Everyone talks about GPU shortages and HBM memory constraints limiting AI infrastructure buildout. But there’s a less visible bottleneck emerging in the supply chain: the specialty glass fiber used to manufacture the PCB laminates inside every AI server.

T-glass — a low-dielectric-constant glass fiber critical for maintaining signal integrity at 56 Gbps and 112 Gbps data rates — is in severe shortage. And without T-glass laminates, you cannot manufacture the 24–40 layer motherboards, switch fabrics, and backplanes that connect AI accelerators together.

Why Standard Glass Won’t Work

The physics are straightforward. Standard E-glass fiber has a dielectric constant (Dk) of approximately 6.2. T-glass (NE-glass) has a Dk of 4.4 — a 30% reduction. This difference translates directly into:

Lower insertion loss: At 28+ GHz Nyquist frequency (112G PAM4), every 0.1 dB/inch of additional loss matters. T-glass laminates deliver 15–25% lower loss than equivalent E-glass constructions.
Tighter Dk tolerance: E-glass Dk varies ±5% due to manufacturing variation. T-glass achieves ±2% tolerance, enabling more predictable impedance control across 40+ layers.
Higher bandwidth density: Lower Dk enables wider traces at the same impedance, or equivalent impedance at narrower traces — both improving routing density in space-constrained server designs.
Reduced skew in differential pairs: More uniform Dk reduces intra-pair skew, which is the dominant limiter for 112G PAM4 reach.

The Demand Explosion

AI Server Board Specifications

Modern AI server platforms require extraordinary PCB specifications:

Platform	Layer Count	Material Class	T-Glass Required
NVIDIA GB200 NVL72	28–36 layers	Megtron 7 / I-Speed	Yes (all signal layers)
NVIDIA GB300 NVL72	32–40 layers	Megtron 7N	Yes
AMD MI400 host	24–32 layers	Megtron 6 / IS680	Yes
Google TPU v6	28+ layers	Proprietary spec	Yes
Microsoft Maia 200	30+ layers	Ultra-low-loss	Yes

Each of these platforms requires T-glass-reinforced laminate for every signal layer. Power and ground planes can tolerate standard E-glass, but signal layers — which comprise 60–70% of total layers in these designs — must use T-glass for adequate channel margin.

Volume Mathematics

Consider a single hyperscaler deploying 10,000 AI server racks in a quarter (a typical buildout pace for major cloud providers in 2026):

10,000 racks × 36 boards per rack = 360,000 boards
360,000 boards × 2 m² T-glass per board = 720,000 m² of T-glass fabric
That’s 720,000 m² per quarter from a single customer

Global T-glass production capacity is estimated at approximately 15–20 million m² annually. When multiple hyperscalers simultaneously expand AI infrastructure, the demand-supply imbalance becomes acute.

Supply-Side Constraints

Why T-Glass Is Hard to Make

T-glass manufacturing differs from standard E-glass in several critical ways:

Furnace chemistry: T-glass uses an alumina-rich, boron-free formulation that requires higher melting temperatures (1,350°C vs 1,250°C for E-glass). This increases energy consumption and furnace wear.

Fiber drawing: The higher viscosity of T-glass melt makes fiber drawing more challenging. Yield rates are typically 70–80% compared to 90%+ for E-glass.

Qualification cycles: Switching a glass furnace from E-glass to T-glass requires complete drain-and-restart, consuming 6–8 weeks. Once running T-glass, the furnace cannot easily switch back.

Capital investment: A new T-glass furnace costs $50–$80 million and takes 18–24 months from investment decision to qualified production output.

Current Manufacturers

The T-glass supply base is concentrated among a few Japanese and Chinese producers:

Nittobo (Nitto Boseki): Largest T-glass producer, estimated 40% market share
AGC (Asahi Glass): Second-largest, focused on highest-purity grades
CPIC (China Jushi subsidiary): Rapidly expanding, primarily serving Chinese laminate makers
Taishan Fiberglass: Growing capacity, targeting mid-range applications

All four are investing in capacity expansion, but new furnaces won’t reach full production until late 2027 at earliest.

Impact on PCB Laminate Availability

The T-glass shortage cascades through the laminate supply chain:

Premium Laminate Lead Times (as of May 2026)

Laminate	Normal Lead Time	Current Lead Time	Price Change
Panasonic Megtron 7	6–8 weeks	18–24 weeks	+45%
Panasonic Megtron 7N	8–10 weeks	20–28 weeks	+55%
Isola I-Speed	6–8 weeks	16–20 weeks	+35%
Isola I-Tera MT40	8–10 weeks	20–24 weeks	+50%
EMC EM-891K	6–8 weeks	14–18 weeks	+30%

These extended lead times mean that AI server PCB manufacturers must commit to material purchases 5–6 months before board production — a dramatic shift from the previous 2-month planning cycle.

Building AI Infrastructure PCBs?

AtlasPCB maintains strategic laminate inventory including Megtron 6, I-Speed, and EM-891K for high-layer-count AI server boards. We provide material availability guidance and DFM optimization to minimize material consumption without compromising signal integrity.

Inquire About Material Availability →

Design Strategies to Mitigate T-Glass Shortage

Hardware engineers facing T-glass laminate constraints can consider several mitigation strategies:

1. Hybrid Stackup Design

Not every layer needs T-glass reinforcement. A practical approach:

Signal layers (high-speed): T-glass laminate (Megtron 7, I-Speed)
Power/ground planes: Standard E-glass FR-4 or mid-tier laminate
Low-speed signal layers: Standard laminate acceptable

A 32-layer board might use T-glass on 20 signal layers and standard glass on 12 power/ground layers — reducing T-glass consumption by 37% with minimal signal integrity impact.

2. Spread Glass (Flat Glass) Alternatives

Spread-glass E-glass weaves achieve Dk closer to T-glass by reducing the glass-to-resin ratio:

Standard E-glass: Dk = 6.2 at 10 GHz
Spread E-glass in high-resin laminate: Dk = 3.8–4.2 at 10 GHz
T-glass: Dk = 4.4 at 10 GHz

For applications at 56G or below, spread-glass constructions may provide adequate performance at lower cost and better availability.

3. Reduce Layer Count Through Advanced Routing

Use via-in-pad and blind/buried vias to increase routing density per layer
Employ wider bus topologies (fewer, faster links vs. many slower links)
Consider chiplet architectures that reduce board-level interconnect length

4. Trace Geometry Optimization

Advanced fabrication techniques can partially compensate for higher-Dk materials:

mSAP (modified Semi-Additive Process): Enables smoother copper with lower conductor loss
Ultra-low-profile copper foil: HVLP or reverse-treated foil reduces roughness loss
Combined effect: 3–5 dB/m improvement at 28 GHz, partially offsetting the E-glass vs T-glass delta

Market Outlook

The T-glass shortage represents a structural challenge that will persist through 2027:

Q3–Q4 2026: Tightest supply period as AI infrastructure buildout accelerates and new capacity hasn’t arrived
H1 2027: Incremental capacity from Nittobo and CPIC expansions provides partial relief
H2 2027–2028: New furnaces reach full qualification, supply-demand rebalances

For hardware teams, the strategic implication is clear: treat T-glass laminate availability as a long-lead constraint item (similar to leading-edge semiconductors) and plan material procurement accordingly.

Frequently Asked Questions

What is T-glass and why does AI hardware need it?

T-glass (technically NE-glass, a boron-free aluminosilicate glass) has a dielectric constant of 4.4 and dissipation factor of 0.0030, compared to standard E-glass at Dk 6.2 and Df 0.0050. At 112 Gbps PAM4 signaling speeds used in AI accelerator interconnects, the additional loss from E-glass makes signal integrity unachievable over the required trace lengths. Every major AI server platform (NVIDIA GB200/GB300, AMD MI400, Google TPU v6) specifies T-glass laminates for their host boards and switch fabrics.

How much T-glass does a single AI server consume?

A typical AI server motherboard uses 24–40 layers, each requiring T-glass-reinforced prepreg and core. At approximately 0.1 mm per prepreg layer, a 40-layer board consumes about 2 m² of T-glass fabric per board. A single NVIDIA GB200 NVL72 rack contains 36 compute trays plus switching boards — potentially consuming over 100 m² of T-glass fabric per rack. With hyperscalers ordering racks by the thousands, the aggregate demand is staggering.

When will the T-glass shortage resolve?

Major glass fiber manufacturers (Nittobo, Nitto Boseki, AGC) are investing in new T-glass furnaces, but glass furnace construction takes 18–24 months from groundbreaking to qualification. Industry analysts expect supply-demand balance to improve in late 2027 to early 2028. Until then, lead times for T-glass-based laminates will remain elevated and pricing will stay 30–50% above 2024 levels.