Despite reporting better-than-expected earnings, NVIDIA, the leading force in AI computing power, saw its stock decline for two consecutive trading sessions. In contrast, shares of key companies in its core supply chain surged again, with related A-share companies also posting significant gains. According to the latest reports from investment institutions, the price of NVIDIA's upcoming new generation of products is expected to nearly double against the backdrop of continuously rising supply chain costs. NVIDIA has not yet responded to this.
Industry analysts indicate that the AI hardware sector has reached a critical juncture of value restructuring. Among the factors, memory and interconnectivity have become key physical bottlenecks, and the interplay between upstream and downstream players in the industrial chain is entering a new phase.
The proportion of GPU costs is declining. Benefiting from the sustained explosion in AI demand, NVIDIA delivered an impressive performance for the first quarter of fiscal year 2027 (February 2026 to April 2026), with core data center revenue reaching $75 billion, a 92% year-over-year increase. Driven by the robust Blackwell architecture products, demand for the GB300 and VL72 racks has been particularly strong. Leading model developers and hyperscale cloud providers have cumulatively deployed hundreds of thousands of Blackwell GPUs, marking the fastest product ramp-up in the company's history.
The highly anticipated next-generation product, Vera Rubin, is scheduled to begin mass production and shipping in the third quarter of this year.
According to a recent Morgan Stanley report, NVIDIA's upcoming Vera Rubin (VR200) rack is being procured from ODMs (Original Design Manufacturers) at approximately $7.8 million, nearly double the price of NVIDIA's current GB300 Blackwell rack.
Memory is seen as a core driver behind the price increase for NVIDIA's products. Coupled with multiple supply chain components also doubling in cost, the share of GPU costs, NVIDIA's core expertise, has actually decreased.
Analysis reports show that, taking the current GB200 NVL72 rack as an example, memory accounts for 5% to 10% of the bill of materials cost. However, for the next-generation VR200 product, the memory share has surged to 25% to 30%, representing a cost increase of up to 435%. Additionally, PCB (Printed Circuit Board) costs have risen by 233%, and MLCC (Multi-Layer Ceramic Capacitor) costs by 182%, among others. Against this backdrop, the share of GPU costs in the total rack cost has decreased from 65% to approximately 51%.
"The sharp price increase for NVIDIA's next-generation Rubin rack is not primarily due to GPUs becoming more expensive, but because the physical bottleneck of AI hardware has shifted from computing power to memory and interconnectivity," stated Chai Daixuan, Managing Director of CIC Burning Knowledge Consulting.
The report analyzes that in NVIDIA's VR200 rack, the number of core chips responsible for data communication and network interconnectivity has significantly increased, leading to a corresponding rise in the usage of components like MLCC substrates.
Earnings reports also confirm this trend. NVIDIA's networking business revenue exceeded $31 billion in fiscal year 2026. In the first quarter of fiscal year 2027, the company's data center networking revenue reached a record $14.8 billion, nearly tripling year-over-year. The scale of NVIDIA's end-to-end Ethernet platform for AI, Spectrum-X, has surpassed the total of all other Ethernet networking vendors.
Amidst the release of positive earnings news, NVIDIA's stock price corrected for two consecutive trading days as investors took profits. Meanwhile, storage suppliers and supply chain companies closely tied to the new Rubin product have seen their stock prices accelerate upward this year.
Industry observer Liu Chao noted that listed companies closely aligned with the Vera Rubin platform, whether in storage, MLCCs, PCBs, or other supply chain areas, have shown significant market performance. For instance, Micron Technology began mass-producing 36GB 12H HBM4 products for the Vera Rubin platform in the first quarter of this year, with its stock price rising 1.38 times. Major MLCC players like Japan's Murata and Taiyo Yuden have seen their stock prices double this year. Within the domestic industrial chain, the PCB sector has shown strong growth. The world's largest copper-clad laminate company, Kingboard Laminates, has seen its stock price increase nearly threefold year-to-date, and PCB equipment companies have also experienced substantial stock price gains.
Despite comprehensive price increases across the supply chain, NVIDIA's control has not been weakened. The latest earnings report shows the company's first-quarter GAAP gross margin at 74.9%, an increase of 14.4 percentage points year-over-year.
NVIDIA's Chief Financial Officer, Colette Kress, recently emphasized in a media interview that the company has collaborated with memory suppliers, including Samsung, SK Hynix, and Micron Technology, for several years in advance to co-design and lock in production capacity. This cooperation model not only ensures massive supply scale but also allows NVIDIA to lock in favorable price ranges early, effectively mitigating the risk of memory price surges.
Furthermore, NVIDIA's founder and CEO, Jensen Huang, previously emphasized that the company controls the hardware resources required for AI, covering packaging, systems, connectors, cables, as well as materials like copper and MLCCs, holding a monopolistic advantage.
In the thriving industrial chain environment with strong supply and demand, end-user cloud providers, as customers for AI computing power, are entering the fray. Beyond developing proprietary chips and purchasing products from NVIDIA's competitors, they may also change procurement models, with hyperscale cloud providers directly purchasing core components.
Morgan Stanley's report analyzes that if cloud providers bypass direct procurement of high-value SOCAMM memory modules, the average price per AI rack could drop from $7.8 million to approximately $6.7 million. Currently, mainstream ODM manufacturers like Foxconn and Quanta are considering a consignment model, where cloud providers procure core components themselves, and ODMs only handle assembly. This could alleviate ODM working capital pressure but may also compress ODM revenue scale.
Chai Daixuan stated that for cloud providers, bringing their own materials bypasses channel markups, reducing total procurement costs. If cloud providers directly procure memory, NVIDIA would cede that markup profit, focusing on its core computing power products. However, for ODM manufacturers, shifting to a consignment model could reduce the cash flow pressure of funding expensive memory purchases, but gross margins might come under pressure.
Some views do not endorse this model. Liu Chao pointed out that NVIDIA sells complete products and services, and memory is just one component. "NVIDIA has its own ecosystem; the integration of hardware and software ensures rack performance. If memory is removed, who guarantees the final quality?"
The domestic AI industrial chain is also closely monitoring and following opportunities related to the launch of NVIDIA's new products.
Chai Daixuan indicated that although it is difficult for domestic companies to directly enter the core high-margin silicon wafer segment, they will benefit from physical support segments, including ultra-high-layer-count PCB materials and comprehensive liquid cooling components. These segments are expected to show significant earnings elasticity with the rollout of Rubin.
Since the beginning of this year, the A-share PCB sector has cumulatively risen over 60%, and the liquid cooling server index has increased by approximately 26%, with related equipment manufacturers showing particularly strong gains.
As a PCB-specific equipment listed company under Han's Laser, Han's CNC executives stated during institutional research that driven by the high-speed transmission demands of AI computing centers, demand for high multilayer boards above 18 layers has significantly increased, evolving towards more layers, greater thickness, higher copper weight, and higher density. The company's CCD six-axis independent mechanical drilling machine, equipped with proprietary 3D back drilling and drilling-testing integrated technology, has passed certification for next-generation AI server PCBs and is in mass production at several leading enterprises.
With the increasing complexity of chip and server motherboard designs, testing solutions and technologies are also continuously evolving. Testing equipment manufacturer Bojie Shares has seen its stock price accelerate upward this year, hitting the daily limit on May 22 and 25. According to the company's annual report, it has developed a complete GPU module thermal management liquid cooling test solution, which has passed major customer certification and begun batch delivery. Additionally, company executives noted during research that a subsidiary covers over 50% of the value of core MLCC manufacturing process equipment.
Comments