M9 Laminate Qualification: What 224G SerDes Demands of PCB Materials
On 14 August 2026, reports indicated that Shengyi Technology’s M9-grade ultra-low-loss copper clad laminate has passed validation within Nvidia’s supply chain, targeting the 224 Gbps transmission requirements of the Rubin platform, and has entered volume supply. The same reporting noted a roughly 25 percent price increase on M9 products in July, an order schedule extending into the first half of 2027, overseas supply from a Thailand facility, a second Songshan Lake plant representing about 5.2 billion yuan of investment with production planned for 2027, a 700,000 square metre computing PCB project at its Shengyi Electronics subsidiary entering trial production in June, and substrate revenue from domestic AI customers growing more than 200 percent year on year.
Those are commercial facts. The engineering fact underneath them is that material grade, rather than fabrication capability, has become the limiting factor for the highest-speed boards.
M7 to M9 Is Not a Model Number Change
Laminate grades are shorthand for a set of electrical and mechanical properties, and moving up a grade changes several of them at once. The progression from M7 to M8 to M9 is driven by the need for lower dielectric constant, lower dissipation factor, and tighter consistency of both across a panel and between lots.
Those properties are not independent of everything else. Lower loss generally requires changes in the resin system, and often in the filler content, which in turn affects flow during lamination, hardness, drilling behaviour, and thermal expansion. A material that performs better electrically may be more difficult to process. That trade is the reason a grade change cannot be treated as a substitution of one part number for another with the same handling characteristics.
The consequence for manufacturing is that the process window has to be re-established. Drilling parameters tuned for one resin system may not suit another. Press cycles need adjustment because flow behaviour differs. Etch compensation changes because the dielectric surface and the copper adhesion behave differently. Impedance models must be recalibrated because the dielectric constant that was assumed is no longer accurate.
Each of those adjustments is engineering work, and each must be validated before the material can be used in production. That is what a qualification programme actually consists of, and it explains why a newly available high-grade laminate does not immediately translate into available high-speed board capacity.
Why 224G SerDes Makes Material the Bottleneck
As GPU clusters grow, the interfaces inside a server continue to advance toward 112G and 224G SerDes. At those rates, the attenuation contributed by the board becomes a first-order constraint on how far a signal can travel and at what rate it can be delivered.
The loss budget of a high-speed channel has several components. Dielectric loss comes from the laminate and scales with frequency and channel length. Conductor loss comes from the copper, and at high frequency it depends heavily on surface roughness because current concentrates near the conductor surface. Radiation and reflection losses come from geometry: via stubs, impedance discontinuities, and reference plane interruptions.
Fabrication technique can address some of these. Back drilling reduces stub effects. Careful stackup design maintains reference continuity. Tight impedance control reduces reflections. But dielectric loss and conductor loss are properties of the materials themselves. If the laminate dissipates too much energy, no amount of routing discipline recovers it, because the loss is distributed along the entire channel rather than concentrated at a defect.
That is why ultra-low-loss material has moved from a preference to a requirement at the top of the performance range. It is also why low-profile copper foil, which reduces the conductor loss contribution from surface roughness, has become standard alongside advanced laminates rather than an optional upgrade.
Layer Count Multiplies Material Sensitivity
The difficulty compounds with layer count. AI server boards occupy a range extending from sixteen layers to seventy-eight, and the highest-complexity backplanes and midplanes sit at the top of that range.
A longer channel accumulates loss, so every incremental improvement in the laminate’s dissipation factor is worth more on a high-layer-count board than on a short one. At the same time, a high-layer-count board consumes more laminate and more prepreg per unit, so material cost and material availability both scale with layer count.
There is a subtler effect as well. On a thick stackup, small variations in dielectric constant translate into meaningful variations in propagation delay and impedance. If the material’s dielectric constant varies across a panel or between lots, the deviation accumulates over the length of a high-speed channel and appears as a mismatch between channels that are supposed to be identical. When differential impedance is held to a tolerance of roughly plus or minus five percent, the allowable variation in dielectric properties is correspondingly tight.
The practical implication is that a laminate qualifies for high-speed use not only on its nominal electrical properties but on the consistency of those properties. A material with excellent average loss characteristics but wider lot-to-lot variation is less useful than a slightly lossier material that behaves predictably. Confirming that a manufacturer can obtain consistent material, and can hold process parameters stable enough to preserve that consistency, is a question that belongs in a capability review rather than in a datasheet comparison.
Qualification Is the Real Capacity Constraint
The commercial signals accompanying the M9 qualification point to something worth understanding: price increases on the material, an order book extending more than a year, and simultaneous investment in upstream material capacity and downstream board production.
Historically, PCB supply capability was assessed in terms of equipment: how many lamination presses, drilling machines, and plating lines a factory operates. In the 224G era, that assessment is incomplete. A factory with ample equipment but no reliable access to qualified high-grade laminate cannot produce effective AI board capacity, because the boards it builds will not meet the required loss budget.
This changes what competition in the segment looks like. Advantages accrue to manufacturers that have qualified material, that maintain relationships deep enough to secure allocation, and that have validated alternatives before they are needed. The reporting that a laminate producer is extending capacity abroad while also investing in board production itself reflects the same logic from the other direction: material producers and board manufacturers are moving from a transactional relationship toward joint validation and coordinated capacity.
The pattern is familiar from other constrained parts of the AI hardware supply chain. Advanced packaging capacity and high bandwidth memory availability were identified as bottlenecks early; the boards and the materials inside them now belong on the same list. Each layer of the supply chain can limit the whole system, and the limiting layer moves as other constraints are relieved.
What This Means for Designers and Buyers
Several practical conclusions follow.
Confirm the material before freezing the stackup. If a design assumes M9-class laminate, the schedule depends on that material being available in the required grade and quantity, and on the chosen manufacturer already running it in production. For teams building AI infrastructure boards, treating material selection as a parallel engineering activity rather than a procurement step removes the most common source of late-stage schedule failure.
Design within the process window a material actually supports. A grade change alters drilling behaviour, lamination flow, and etch characteristics. Parameters developed for the previous material must be revalidated rather than copied. Where a programme intends to transition from a lower grade to a higher one, planning that transition with the manufacturer during fabrication planning is cheaper than discovering the incompatibility on a production panel.
Use hybrid stackups where the loss budget allows. Not every layer needs the lowest-loss material available. Placing advanced laminate only on the high-speed layers and using a conventional high-Tg material elsewhere reduces both cost and exposure to constrained grades. The trade-off requires the fabricator to control multiple materials in one lamination cycle, which is a capability question that should be answered rather than assumed.
Ask how consistency is measured. A supplier that can describe how dielectric properties are verified, how coupon panels are used to characterise material lots, and how impedance is monitored in production is managing the variable that matters. Those records belong in a documented quality system, and they are what allow a customer to defend an electrical design across a production year.
Plan verification beyond the bare board. A board that meets its loss budget can still fail once a high-speed device is attached, because assembly introduces its own variables. Paste volume, placement accuracy, and hidden joint quality determine whether the channel behaves as designed. A structured test flow covering paste inspection, optical inspection, X-ray, and electrical verification closes that gap, and it is the difference between a board that measures correctly and a system that runs reliably.
Frequently Asked Questions
What does M9 laminate mean? It denotes an ultra-low-loss copper clad laminate grade with lower dielectric constant and dissipation factor than earlier grades, intended for the highest-speed channels in AI server platforms.
Why is material qualification more important than fabrication capability at 224G? Because dielectric loss is a property of the material distributed along the entire channel. Fabrication techniques can fix stub and impedance problems, but they cannot recover energy dissipated by the dielectric.
Why does layer count increase material sensitivity? Longer channels accumulate loss, thicker stackups amplify the effect of dielectric constant variation, and higher layer counts consume more laminate per board.
Can a lower-grade laminate be substituted to reduce cost? Only where the loss budget permits, and only with measured validation of insertion loss and impedance. Substitution on a design with no margin produces a link that fails later, at maximum temperature or channel length.
What should buyers ask a supplier about material strategy? Which high-speed grades are qualified and in supply, whether alternatives have been electrically validated, how material consistency is verified, and how material changes are controlled and communicated.
