2026.09.23 | AC | PCB | Major CCL Maker: M8U/M9 Upgrades

摘要

1. Growth in AI high-speed CCL shipments and M8U/M9 upgrades;
2. Validation and mass production cadence for Rubin Ultra PTFE/Q-cloth solutions;
3. High-end copper foil shortages and global capacity expansion.

The following are the expert's views:

What is the current monthly shipment level for high-speed materials in the AI sector? What is the shipment breakdown across different product grades such as M8, M8+, M8U, and M9?

Monthly shipments of high-speed materials used in AI currently stand at approximately 400,000 sheets. M8 and M8+ together account for roughly 80%. M9 represents between 5% and 10% of shipments and has reached mass production, primarily serving LPU version products. Shipments of M8U have also begun. The remaining 10% to 15% consists of M7 and M6 grade materials, which are supplied primarily to domestic customers, as overseas customers no longer use them.

Looking at the downstream customer structure, who are the primary recipients of high-speed material shipments at present?

Shipments are currently concentrated between two customers. Victory Giant is the largest, followed by Shenzhen Kinwong Electronic. Other accounts, such as Founder and Wus Printed Circuit, account for smaller shares.

What are the current prices for primary high-speed materials like M8, M8U, and M9? Have there been any changes compared with 2026Q2?

M8 is currently priced at around 900 yuan, M8U at around 1,300 yuan, and M9 in a range of 2,000 to 2,500 yuan. These prices have remained unchanged since 2026Q2.

What trajectory do you anticipate for high-speed material shipments from September to December 2026? When do you expect shipments to hit the target of 600,000 sheets per month?

Overall shipments are expected to maintain growth from September through December 2026, with M8U and M9 projected to post month-over-month growth of 20% to 30%. Concurrently, shipments of M8 used in GB products will decline, whereas shipments of M7 materials—primarily supplied domestically—should hold steady. Monthly shipments of high-speed materials are expected to reach over 600,000 sheets by the end of 2026, at which point M8U should make up 80% to 90% of total shipments.

Can current production capacity satisfy the growing demand for high-speed materials?

Capacity is sufficient. High-speed materials share capacity with standard laminates, allowing production lines to switch between them as needed.

Excluding PTFE, what is the current overall capacity level for other PCB laminates?

Apart from PTFE capacity, which is dedicated and cannot be converted, capacity for other laminates is mutually interchangeable. Excluding PP, total capacity for these laminates exceeds 10 million sheets.

Roughly what are the prices of PP products in the M8 and M9 series?

PP products are typically priced at roughly half the cost of the corresponding laminate. Taking M8 as an example, its PP product sells for around 500 yuan. Pricing for M9 series products follows a similar structure.

What is the current shipment breakdown for the M9 series? What proportion do Q-cloth products represent within the LPU category?

Current shipments of the M9 series comprise LowDK Gen 2 and Q-cloth, which go into different board types. In the LPU category, Q-cloth products make up approximately 40% of shipments.

Regarding the application of PTFE materials in next-generation products, what is the status of validation, technical solutions, and supply chain readiness?

PTFE validation results show marked improvement over earlier rounds, primarily targeting two rack architectures within the Rubin Ultra platform. The first is the newly designed Kyber rack, which replaces traditional copper cable backplanes with orthogonal backplanes to connect Switch and Compute Trays. Initial designs tested a 78-layer PTFE board for this orthogonal backplane; however, due to excessive layer count, the design was revised to laminate three 26-layer boards (i.e., two sequential lamination steps) to overcome processing and yield hurdles stemming from high aspect ratios and board thickness. Testing on this concept began in 2024; while electrical performance met specifications, yield has remained the primary bottleneck. As an alternative, M9 Q-cloth failed to qualify due to an approximate 20% electrical performance (insertion loss) deficit. Beginning in 2025, the industry began evaluating M10 solutions, but a ~10% performance gap remains. The most advanced architecture currently pairs M10 PP with PTFE cores in a hybrid lamination, achieving yields of 40% to 50%. However, not all PCB suppliers can run this process, resulting in an inadequate supplier base that could delay the program or restrict it to small-batch trial production. The second architecture is the NVL576 Oberon solution based on the existing Oberon topology, which links eight NVL72 units via CPU optical interconnects. Although these units are distributed across distinct racks, inter-rack link speeds must hit 337Gbps intra-rack scale-up performance. To reach this throughput, the design upgrades the number of switch boards from 9 to 18 while boosting operational speeds. Switch boards for this design are testing hybrid lamination combining PTFE cores with M9 Q-cloth PP. To improve yields and lower processing complexity, the PCB structure was changed from an initial 32-layer high-layer-count board to two bonded 16-layer boards. Current testing indicates links are fully validated, the PCB is mass-producible, and yields can reach 60%. Because no single PCB or CCL supplier can fulfill downstream demand, the supply chain is aggressively positioning itself. Vendors like Elite Material, ITEQ, and Taiwan Union Technology, which historically had limited exposure to PTFE CCL, are already submitting PTFE base material samples. Concurrently, PCB makers such as Zhen Ding Tech and Gold Circuit Electronics are recruiting PTFE PCB processing talent, while Victory Giant, Shenzhen Kinwong Electronic, and others are committing resources to related R&D to ensure production readiness when PTFE scales up. The ultimate selection will be decided by NVIDIA based on PTFE CCL and PCB supply volume, vendor redundancy, and pricing, though the probability of adoption remains relatively high.

In the NVL576 Oberon design, what layer stackup was ultimately selected for the switch board?

Several layer configurations were evaluated for this solution, including 18x18 and 22x22 architectures. After weighing cost-performance tradeoffs and yields, the 16x16 stackup was determined to be the optimal choice.

Is it confirmed that the switch board will upgrade to 18 NVLinks?

Current test platforms show an upgrade to 18 NVLinks, and this setup has demonstrated solid performance. However, it remains in the testing phase and could differ from the final implementation, which will be decided by the end customer.

What is the specific material composition of the switch board? For instance, in a 32-layer board formed by two 16-layer boards, how many layers are PTFE versus hydrocarbon PP?

Across the 32-layer stackup, there are 14 layers of PTFE material and 16 layers of M9-grade hydrocarbon PP. The solution utilizes PTFE.

When is this switch board design expected to be finalized?

It is expected around early next year (2027).

How does the cost structure of the new PTFE-based solution compare with the M9 material solution?

The new solution carries lower costs than the M9 design, primarily on the raw materials side. Key differences include: Resin costs: The resin utilized in M9 is expensive, running over 2,000 yuan per kilogram and reaching up to 5,000 yuan, whereas PTFE resin is comparatively cheap at roughly 100 to 200 yuan per kilogram. With 0.1 to 0.2 kilograms of resin used per square meter of laminate, the cost disparity is substantial. Reinforcement costs: Quartz cloth used in M9 is costly, priced around 300 yuan per meter. That translates to 300 yuan per sheet of laminate, or 600 yuan if two plies are used. The PTFE solution uses no glass cloth, reducing this cost component to zero. Silica powder costs: Silica powder expenses are fairly comparable across both designs. Because the PTFE formulation substitutes silica powder for glass cloth, its higher filler loading leads to slightly higher silica costs, though the difference is modest. Processing costs: Processing fees for the PTFE solution run roughly double those of M8 or M9, adding an estimated 100 to 200 yuan. However, set against the massive raw material savings, the net cost advantage remains pronounced.

What is the projected unit price for this new PTFE-based solution?

Pricing is not yet finalized and will hinge on market supply and demand. In the initial phase, if supplier availability is tight and output falls short of demand, pricing will likely reflect its performance advantage over M9, commanding a premium of 2,000 to 3,000 yuan per square meter. Over time, as adoption expands and more suppliers enter the market, pricing may transition to a cost-plus model, at which point prices should drop below 2,000 yuan per square meter.

Which board fabricators are currently running tests on PTFE products? How involved is your company?

Testing is currently taking place primarily at Shennan Circuits and Wus Printed Circuit. Our company is also participating, but we are not a primary test vendor, and our position in NVIDIA's supplier hierarchy is not near the front.

Does your company face hurdles in penetrating NVIDIA's supplier system?

Gaining entry into their supplier network requires a long lead time. In addition, our current capacity serves as a constraint.

Based on a board structured with 14 layers of PTFE and 16 layers of hydrocarbon PP, what is the estimated dollar value?

One square meter of laminate represents roughly 50,000 yuan in value. Given that switch board panels are cut to a 1-out-of-6 layout—meaning roughly 0.2 square meters per board—the value per switch board comes out to approximately 8,000 yuan.

Why is this upgrade focused primarily on the switch board rather than the compute board?

The primary goal of this upgrade is to boost inter-rack interconnect bandwidth to enable scale-up performance across 72, 144, or even 576 GPUs. In multi-rack systems, data transfer bottlenecks reside largely in inter-rack connections. To allow eight racks to communicate as a single unified system at high speeds, connections must run through CPO optical modules and optical fiber. Once high-speed, high-density optical signals reach the rack, switch boards must instantaneously process data arriving simultaneously from all eight racks, necessitating an upgrade to handle these ultra-high switching rates. Compute boards, on the other hand, mainly handle computational workloads for their four onboard GPUs and rely on switch boards for routing, making switch boards the core focus of the upgrade.

In the eight-rack interconnect scheme, do CPO boards also utilize PTFE materials?

Yes, CPO boards are also evaluating PTFE materials. These boards feature higher layer counts—around 40-plus layers—and carry higher manufacturing difficulty. In an NVL72 eight-rack system, a CPO is installed atop each rack; these eight CPOs communicate via the NVLink protocol and operate in tandem with the intra-rack switch boards.

Will Kyber-architecture orthogonal backplanes launch in tandem with other NVIDIA products? What is their market positioning?

That depends on the end customer's roadmap. One possibility is a simultaneous release next year (2027). Owing to capacity and technical considerations, Kyber may target premium deployments, particularly space-constrained environments, because it delivers exceptionally high volumetric compute density with a footprint dozens of times smaller than the NVL576. By contrast, while the NVL576 also reaches high density, it takes up more space and may serve as the standard deployment option for sites lacking strict footprint constraints that require compute upgrades. Shipments of architectures like NVL72 could fall off sharply once that occurs. Mainstream volume would likely shift toward the NVL576 and the Kyber-based NVL144, tracking the historical trend where older rack designs phase down as newer generations ramp.

In the 26-layer board of the Kyber architecture, how are PTFE and M10 materials layered together, and what is the layer count for each?

Within the 26-layer stackup, there are roughly 12 layers of PTFE and 13 layers of M10 hydrocarbon PP. For the M10 material, designs can incorporate either standard or Q-cloth options.

When using the M10 Q-cloth hydrocarbon PP configuration, how does the material cost break down? How does it differ from the PTFE material discussed earlier?

Adopting the M10 solution primarily involves upgrading the resin system, while PTFE pricing remains consistent with prior figures. The resin used in the M10 Q-cloth version is significantly more expensive, running 5,000 to 10,000 RMB per kilogram, with certain high-end grades hitting 20,000 RMB per kilogram. In addition, the Q-cloth may shift to Gen 2, driving costs up from around 300 yuan per meter to 500–600 yuan. It is worth noting that these advanced materials are not yet capable of mass production.

Factoring in material costs, yields, and panel utilization, how do you calculate the final material procurement cost for a 26-layer orthogonal backplane?

You start by calculating bare material costs per board—for example, tallying unit costs across 12 layers of PTFE and 13 layers of PP. Next, factor in board panel dimensions; a 1-out-of-4 panel, for instance, spans roughly 0.25 to 0.3 square meters. Final CCL purchasing volume is calculated from this surface area while incorporating two major yield loss factors: first is PCB fabrication yield, which may be only 40% for orthogonal backplanes compared to roughly 60% for switch boards and around 80% for previous-generation GB300 switch boards; second is panel utilization (edge trim scrap), which typically hovers around 80%. Consequently, total CCL procurement costs are calculated by taking base material costs, dividing by the fabrication yield, and then dividing by panel utilization (e.g., 0.8). Quoted CCL prices already incorporate laminator manufacturing yields, which are generally high—typically 80% to 90%, or around 50% on poorer runs.

What are the current processing fees and pricing trajectories for Gen 3 and Gen 4 high-end copper foil?

Currently, processing fees for Gen 3 copper foil range between 150,000 and 200,000 RMB per ton, while Gen 4 is slightly higher at 200,000 to 250,000 RMB per ton. The high-end copper foil segment is experiencing a supply shortage with upward price momentum. Prices saw an initial upward adjustment early this year of about 10% to 30%, with timing varying by region: overseas and Taiwanese suppliers tend to adjust every six months with larger 30% jumps, while domestic suppliers adjust every one to three months with smaller ~10% increases. By year-end, as Rubin ramps and drives up foil consumption, another wave of price increases is expected, led particularly by overseas suppliers.

When do you anticipate an inflection point in electronic glass fabric pricing? Specifically, will prices climb through mid-2027 before declining, or will they plateau for a period?

The pricing inflection point for electronic glass fabric is expected around mid-2027, at which point prices may plateau or start trending downward. The exact trajectory will depend on market balance: oversupply will trigger steep cuts, while balanced conditions would keep prices flat. Given that substantial new capacity is slated to come online across the sector by mid-2027, a downward price move appears more likely.

How will CCL pricing evolve between now and the end of 2026? Could prices decline if demand falls short of expectations?

CCL prices should maintain an upward trajectory through the end of 2026, with likely one more price hike; a sharp downturn is improbable. Even if demand comes in below expectations, prices are more likely to fluctuate at elevated levels. This rests on two factors: first, year-end marks the peak season for consumer electronics, supported by holiday demand across National Day, Christmas, and New Year; second, most CCL makers will have met their annual targets by year-end, leaving little incentive to discount aggressively for market share. Furthermore, AI-related demand continues to expand. While rising component and memory costs have pushed up prices on handsets and PCs and dampened some consumer demand, core baseline demand remains intact.

With PCB fabricators passing cost pressures downstream, will gross profit per CCL sheet hold steady, or could it expand further?

With numerous small and mid-sized PCB shops having shuttered, orders have consolidated toward tier-one and mid-tier fabricators, producing a capacity crunch in PCB manufacturing and prompting fabricators to raise prices sharply upstream. However, these increases have begun chilling end-market consumer electronics demand, softening hardware purchases and curbing demand for PCBs. The market will likely find a new balance point, keeping margins per CCL sheet largely steady or flat.

What is the outlook for gross margins on advanced materials like PTFE, M8, M8U, and M10 once they hit high-volume production?

Final gross margins will hinge on competitive dynamics: more market entrants will compress margins, whereas limited competition will keep them elevated.

What is the progress on the samples your company submitted to Google for qualification? Which product was submitted, and when do you expect results?

The qualification process with Google is ongoing, with primary samples focused on V8-series M8. Further results will likely emerge in the first half of 2027.

How reliant is your company on domestic versus foreign suppliers for high-end copper foil? How do you expect domestic supplier share to trend?

Our high-end copper foil sourcing currently relies heavily on overseas suppliers. For example, Fukuda's capacity is largely locked up by Elite Material and our company, while Mitsui allocates volume on a quota system. To grow our footprint in AI CCL, incremental raw material demand must be satisfied by suppliers in mainland China and Taiwan, including Co-Tech, Defu Technology, and Tongguan. As incremental demand grows, the share held by mainland and Taiwanese vendors should gradually expand.

What share of HVLP3 and HVLP4 copper foil supply can domestic suppliers capture by 2027?

Incremental supply of HVLP3 will be very limited by 2027. For HVLP4, incremental demand will be met primarily by domestic suppliers. Assuming domestic production ramps smoothly, their shipments are expected to account for 30% to 40% of total incremental volume.

Will your company continue qualifying additional domestic copper foil suppliers? Which vendors look most promising?

We will continue testing viable domestic vendors whenever their products near production readiness. Suppliers that remain far from volume production capabilities will be kept on a watch list for now.

Will processing fees for high-end copper foil decline as domestic capacity expands? What is the main capacity constraint?

Standard copper foil prices have not seen meaningful increases. For high-end foil, the primary bottleneck is the limited availability of Mitsui surface treatment machines. Even if foil makers add raw capacity, output cannot be realized without Mitsui treatment equipment, capping real capacity expansion. While vendors like Defu Technology are pursuing in-house treatment tools, internal yields remain low and costs high, ruling out volume production for now. By contrast, Mitsui equipment delivers higher stability and superior yields.

What level of market demand do you project for PTFE laminates in 2027 and beyond? What factors determine usage?

PTFE laminates should enter mass production in 2027, with annual consumption likely hitting 1 to 2 million sheets. By 2028, as target rack architectures ramp, monthly demand could surge to 1 to 2 million sheets, reaching tens of millions of sheets annually. Actual volumes will tie directly to end-customer shipment numbers (such as NVIDIA Rubin Ultra), final hardware architectures (e.g., orthogonal backplanes versus switch boards), and the specific adoption scope of PTFE on various PCBs.

What challenges is NVIDIA encountering in selecting high-end materials? What trends do you foresee in material adoption?

NVIDIA's primary challenge is the tight supply of single high-end material types; neither legacy capacity nor upcoming additions can fulfill its massive demand. Processing equipment is the core expansion bottleneck, and even new capacity enabled by toolmakers will fail to supply NVIDIA's needs if it relies on a single material across all product lines. Consequently, NVIDIA must make tactical trade-offs by diversifying material usage across different boards—deploying Q-cloth on some PCBs and PTFE on others. The likely direction is a hybrid architecture where PTFE and Q-cloth (such as M9 Q-cloth or M10 Q-cloth) coexist.

What is the current status of Fangbang Electronics' peelable copper foil? What are its primary target applications?

Fangbang Electronics' mid-to-low-end peelable copper foil has cleared major technical hurdles and is shipping normally, targeting mainstream memory chips like DDR3 and DDR4. Adoption remains limited for its higher-tier peelable foil aimed at premium chips like HBM and DDR5.