A green board is quietly rewriting the bill of materials for AI compute hardware.
Morgan Stanley’s May 2026 teardown of NVIDIA’s Rubin platform reached a counterintuitive conclusion. Inside a VR200 NVL72 rack priced near $7.8 миллион, the component with the largest value increase was not the GPU. It was not HBM either. The printed circuit board (печатная плата) content per rack climbed from roughly $35,100 в поколении GB300 до $116,700 in Rubin, а 233% jump that led every downstream component. Тем временем, GPU share of the BOM fell from about 65% in GB200 to approximately 51% in Rubin.
This number only makes sense against a longer technical arc. Each NVIDIA architecture iteration moves the PCB from a passive carrier of chips toward an active medium that governs signal integrity and system interconnect density. For sourcing teams evaluating AI server PCB suppliers and hunting for Высокий счет (HLC) и взаимосвязь высокой плотности (ИЧР) емкость, understanding this curve matters more than chasing spot quotes.
The Real Driver Is Specification Jumping, Not Volume Growth
Industry shorthand often reduces PCB value growth in AI servers to “three to five times the volume, eight to twelve times the value” of traditional servers. That framing is accurate but incomplete. The stronger force is a synchronized jump across three dimensions: количество слоев, HDI build-up layer count, and copper-clad laminate (Ccl) grade.
Количество слоев. Traditional server motherboards sit at 16-20 слои. AI server GPU compute boards typically run 20-40 слои, while switch boards reach 32 слои. Rubin adds a 44-layer midplane that cuts tray assembly time from nearly two hours to five minutes. A more aggressive node arrives in the second half of 2027. Rubin Ultra’s Kyber rack replaces tens of thousands of copper cables with a single 78-layer M9-grade orthogonal backplane. That backplane connects 576 GPUs and NVSwitch fabric directly, elevating the PCB from a board-level component to a rack-level interconnect medium. Prismark tracking shows AI server PCB layer counts already reaching 78 layers in premium models, with iteration toward 80 слои и не только.
ИЧР. Build-up layer count is a second steep difficulty curve. Шеннан’ public disclosures describe volume production of 6-order 22-layer HDI and manufacturing capability for 9-order 28-layer HDI. Each added build-up layer tightens laser drilling, plating fill, and layer-to-layer registration tolerances exponentially. IPC-6012 Класс 3 requires layer-to-layer registration within +/-2 мил (50 микрометры) for high-reliability rigid boards. That tolerance window becomes near-critical when registration errors stack across 70 or more layers.
Материалы. Signal speed is the direct pressure behind CCL upgrades. Traditional server board-to-board links run at 25-56 Гбит / с. Blackwell introduced 112 Gbps PAM4 SerDes, and Rubin pushes toward 224 Гбит / с. Each doubling of data rate raises PCB transmission loss exponentially. Стандартный FR-4 (Df around 0.02) cannot meet the loss budget. Rubin compute boards have moved from M7-grade CCL in GB300 to M8-grade material. M8-class CCL delivers Df values around 0.0012-0.0016 в 14 ГГц. Rubin’s switch tray uses M8U-class material (Low-Dk2 plus HVLP4 copper) in a 24-layer HDI structure, while midplanes and CX9/CPX boards incorporate M9-class CCL (Q-glass plus HVLP4) reaching up to 104 слои. M9-class development targets Df near 0.0006 в 10 GHz and tighter CTE for longer, faster channels.
Ccl: The Underappreciated Deciding Factor in Supplier Qualification
Copper-clad laminate accounts for 20%-40% of PCB manufacturing cost. In AI servers, однако, CCL is far more than a cost line. It determines whether a board can pass the loss budget at 224G and whether a PCB fabricator can enter NVIDIA’s or a North American hyperscaler’s approved vendor list.
The M8/M9 supply landscape is highly concentrated. One Chinese CCL producer, Shengyi Technology, has moved its M7/M8+ high-end AI laminates into volume supply for a major North American customer. The company holds a leading share of NVIDIA’s M8-level material and is progressing through M9 qualification. Goldman Sachs raised its global high-end CCL market forecast in August 2026 к $22 миллиард за 2027. The core logic is straightforward: specifications migrate toward M9 and above, unit prices and margins rise, and capacity releases lag demand.
CCL qualification creates a lock-in effect distinct from ordinary raw materials. Once a laminate grade passes a major PCB fabricator’s qualification and enters an NVIDIA part-number system, switching costs become prohibitive. The first qualified CCL suppliers therefore stay embedded in incremental share across an entire high-end PCB capacity expansion cycle. For PCB fabricators seeking stable M8/M9 CCL channels, aligning forecast schedules with laminate suppliers already inside top-tier qualification pipelines is a practical way to reduce material-shortage risk into 2027.
Qualification and Capacity: An 18-to-24-Month Time Moat
The PCB industry never lacks expansion announcements. What remains scarce is effective high-end capacity, meaning lines that simultaneously hold customer qualification, stable yields, and repeatable delivery on high-layer HDI and HLC boards.
NVIDIA, AMD, and other compute giants run a three-stage PCB supplier qualification process. It starts with laboratory performance testing, moves to small-batch pilot runs, and ends with long-term supply qualification review. The full cycle takes 18 к 24 месяцы. Rubin-class AI PCB qualification exceeds one year on its own. New entrants may complete capacity build-out and still find the supply chain closed, because industrial tracking suggests at least two to three years are needed to enter mainstream AI PCB supply.
Capacity constraints are equally rigid. New high-layer HDI and HLC lines typically need 12 к 18 months to reach stable volume production. Complex AI server products stretch the ramp to 24 месяцев или больше. Initial yields often sit below 60%. Laser drilling and precision plating equipment lead times add another layer of delay. Prismark expects the premium-board shortage to persist at least through 2027.
Qualification cycles, yield ramps, and equipment lead times stack into the sector’s most durable moat. This is why only a handful of manufacturers worldwide can consistently supply AI server high-layer PCB and HDI boards. Buyers with qualified high-layer HDI or HLC demand should engage fabricators already holding completed top-tier AI customer qualification for technical review and a quote, rather than waiting for new lines to finish construction.
China’s PCB Position: Volume and Quality Rising Together
Mainland China is already the core base for global PCB manufacturing. Measured by the output of Chinese-funded enterprises, global market share rose from 34.9% к 37.6% в 2025, with output around $34.18 миллиард. The more important signal, однако, is where high-end capacity is flowing. Of the industry’s 2025-2026 capacity expansion capex, 82% is concentrated in mainland China and Taiwan. The spending targets HLC, mSAP HDI, and ABF substrates.
Prismark’s regional data confirms a division of labor: mainland high-end lines paired with Southeast Asian volume facilities. Thailand and Vietnam mainly absorb mid-range communications and automotive PCB work. High-end HLC, Мсап, and substrate capacity remain in mainland China, where the complete supply-chain ecosystem is difficult to replicate offshore. Mainland PCB output is forecast to reach approximately $47.71 миллиард в 2026, outpacing the global market.
Для AI server PCB and PCBA buyers, this landscape has a practical meaning. Leading Chinese manufacturers now cover volume experience from 20 к 78 слои. Selection therefore hinges less on whether capacity exists and more on whether a specific layer-count and HDI build-up combination already has a qualified, volume-production case at the end customer.
A Curve Still Steepening
From GB300 to Rubin, per-rack PCB value grew 233%. By the time Rubin Ultra NVL576 reaches volume in 2027, Kyber’s orthogonal backplane, 800V DC power PCB, and CoWoP packaging paths will redistribute value again across three dimensions. Per-board value could reach two to three times that of conventional PCBs. Prismark projects a 24.8% CAGR for AI server PCB from 2025 к 2030, with AI board layer counts iterating from 22-plus toward 80-plus.
Two variables set the slope of this curve: how many more times signal rates can double, and how many PCB/CCL suppliers can keep pace with each doubling. For readers inside the procurement chain, the answer does not sit in the BOM number itself. It sits in the qualification list and the volume-production record.
Data Sources
- Rubin VR200 NVL72 per-rack PCB value ($116,700) and GB300 comparison ($35,100, +233%): Morgan Stanley Rubin platform BOM teardown, Может 2026, cross-referenced with Yahoo Finance and multiple brokerage research notes.
- VR200 rack BOM ($7.803 миллион), layer upgrades (26-слой HDI / 44-layer midplane), and Rubin Ultra 78-layer orthogonal backplane: Hebei Yuanda Securities PCB industry report, Сентябрь 2026, based on Morgan Stanley and supply-chain data.
- Global PCB output ($85.836 миллиард в 2025; $101.954 billion forecast for 2026) and AI server PCB market size/growth: Prismark PCB and Substrate Market Report, Q1 2026.
- AI server PCB layer range (78 layers in premium models, iterating beyond 80), HDI build-up layer count (6-order 22-layer volume production / 9-order 28-layer capability): Shennan Circuits investor disclosure, январь 2026; Prismark Q1 2026.
- CCL grades (M8 Df around 0.0012-0.0016 в 14 ГГц; M9 Df around 0.0006 в 10 ГГц; CTE and thermal targets): BestPCBs M8/M9 CCL comparison, Сентябрь 2026; Morgan Stanley Rubin BOM teardown; Hebei Yuanda Securities.
- Rubin switch tray (M8U-class, Low-Dk2 + HVLP4, 24-слой HDI) and midplane/CX9/CPX boards (M9-class, Q-glass + HVLP4, до 104 слои): New Electronics, citing TrendForce research, ноябрь 2025.
- Shengyi Technology M8 share in NVIDIA systems and M9 qualification progress: Goldman Sachs research, Август 2026; Zhongcai News, Сентябрь 2026.
- PCB supplier qualification cycle (18-24 месяцы) and new-entrant timeline (2-3 годы): Securities Times industry survey, cited by Lanjing Finance; Tianfeng Securities AI-PCB supply-chain update.
- High-end PCB line yield ramp (12-18 месяцы; AI high-layer HDI 24+ месяцы; initial yield below 60%): Industry expert interview in Securities Times, cited by Lanjing Finance.
- Chinese-funded PCB enterprise global share (37.6% в 2025, $34.18 billion output) and mainland output forecast ($47.71 миллиард в 2026): TPCA data, cited by Sange Pijiang Report.
- IPC-6012 Класс 3 layer-to-layer registration tolerance (+/-2 мил / 50 микрометры): МПК-6012, Квалификационные и эксплуатационные характеристики жестких печатных плат; AtlasPCB technical reference.
- UL 94 В-0 flammability requirement (self-extinguishment within 10 секунды, no flaming drips) и UL 796 board-level safety standard: UL 94 и UL 796, Андеррайтеры лаборатории; AtlasPCB and PCBSync certification guides.
- AI server PCB usage at 3-5x traditional servers and value at 8-12x: EET-China industry report, synthesizing Prismark and supply-chain estimates.
