Nvidia 已决定，其未来前景取决于改进的硅基板和封装供应链，因此动用其巨额现金储备，以确保能获得所需零部件，方法是确保供应商拥有所需资源，在维持 GPU 和 AI 处理器强劲供应流中发挥各自的小作用。制造商投资供应商并不罕见，因为这有助于创造确定性并分担财务风险。但在一家大型科技公司的大会上宣布此类交易 —— 正如 Nvidia 在其柏林 GTC 大会上所做的那样 —— 尚属罕见，这表明该公司感到有必要向投资者展示，它正在努力确保能满足其产品的需求，且这种需求将持续数年。
Nvidia has decided that its future prospects rest on an improved silicon substrate and packing supply chain, so has dipped into its giant pile of cash to ensure it can get the parts it needs by ensuring its suppliers have the resources needed to play their small part in keeping the flow of GPUs and AI processors strong. It’s not unusual for manufacturers to invest in their suppliers, because it helps to create certainty and spread financial risk. For those deals to be announced at a major tech company’s conference – as Nvidia did at its Berlin GTC gabfest – is somewhat novel, suggesting the company feels the need to show investors it is working to ensure it can meet demand for its products, and that demand will persist for years.
Nvidia 的多项宣布涉及硅基板，即晶圆。无论怎么称呼，它们都是空白的、圆形的硅片，芯片制造商在其上蚀刻电路。它们由向台积电和英特尔等晶圆厂销售的公司制造，后者再将其制成各种类型的芯片。在柏林 GTC 大会上，Nvidia 首席执行官黄仁勋指出，AI 硬件需求巨大，但“产能是制约因素”。具体而言，是制造基板、先进硅封装、测试和制造光连接器等环节的产能。
Several of Nvidia’s announcements pertained to silicon substrates, aka wafers. Whatever you call them, they’re blank, round, silicon slabs onto which chipmakers etch circuits. They’re made by companies that sell to foundries like TSMC and Intel, which then turn them into chips of various flavours. At GTC Berlin, Nvidia CEO Jensen Huang observed that demand for AI hardware is vast, but that “Capacity is the constraint.”Specifically, capacity for activities like building substrates, advanced silicon packaging, testing, and creating optical connectors.
因此，大绿厂计划在未来三年内将其产品的基板产能提高 2.5 倍。为实现这一目标，Nvidia 已将三星电机业务部门纳入其基板供应链。三星显然在韩国一家工厂拥有即时产能，Nvidia 打算立即利用。根据该协议，到 2028 年，三星约四分之一的基板产量将供应 Nvidia。Nvidia 已与日本基板厂商 Ibiden 达成协议。
Big Green therefore plans to increase substrate capacity for its products by 2.5x over the next three years. To make that happen, Nvidia has added Samsung’s Electro-Mechanics business unit to its substrate supply chain. Samsung apparently has immediate capacity at a factory in South Korea and Nvidia intends to tap that now. The deal means that by 2028, about a quarter of Sammy’s substrate output will go to Nvidia. Nvidia already has a deal with Japanese substrate outfit IBIDEN.
在柏林，这家 GPU 巨头指出将进一步扩产，将一家 Ibiden 工厂的产量提高两倍（即增至三倍）。
In Berlin, the GPU giant pointed to further expansion that will triple production at an IBIDEN plant.
台湾基板厂商华通电脑也因英伟达的慷慨而受益，形式为产能预留协议及相关厂房扩建计划。英伟达想要“囤积”的不仅仅是晶圆。该公司还与韩国封装测试厂安靠签署了产能预留协议。同样，绿厂的介入将使其合作伙伴扩大产能，此次幅度达300%。英伟达表示，这项协议“将支持当前和下一代平台的更高产量”。
Taiwanese substrate player Unimicron is also benefiting from Nvidia’s generosity, in the form of a capacity reservation agreement and associated plant expansion plans. It’s not just wafers that Nvidia wants to tuck into. The company has also signed a capacity reservation with Korean silicon packaging player Amkor. Again, Big Green’s involvement will see its partner expand capacity, in this case by 300 percent. Nvidia says this deal “will support higher production volumes for current and next-generation platforms.”
AI系统需要高速网络，许多厂商认为光连接是构建能够应对现代工作负载所产生和需要的数据洪流的系统的最佳方式。因此，英伟达制定了一项计划，以提升光学封装厂商SPIL的制造产能。同样，绿厂已与该公司达成协议，但在柏林，它扩大了该安排，希望到2027年年中将产能扩大15倍，然后到2028年底再翻一番。
AI systems need speedy networks, and many players see optical connections as the best way to build systems capable of handling the data deluges modern workloads create and require. Nvidia has therefore hatched a plan to boost manufacturing capacity at optics outfit SPIL. Again, Big Green already has a deal with the company, but in Berlin it expanded that arrangement in the hope it expands production capacity 15x by mid-2027, then doubles that again by the end of 2028.
所有这些迹象表明，英伟达相信AI基础设施繁荣还将持续数年。®
All of which suggests that Nvidia believes the AI infrastructure boom has years left to run. ®