CITIC Securities: Rapid Growth in High-Frequency and High-Speed Demands Such as Computing Power, Electronic-Grade PTFE Expected to Be Widely Applied 中信建投：算力等高频高速需求快速增长，电子级PTFE有望大规模应用

When NVIDIA engineers repeatedly test the Rubin Ultra server backplane in laboratories, the true competition may not be about computing power, but about who can more quickly master a type of plastic called PTFE. This is not science fiction; it is a materials revolution unfolding in real time—the revelry of computing infrastructure has finally reached the most overlooked foundational material layer.

PTFE, short for polytetrafluoroethylene, is the material used for non-stick pan coatings. But don’t underestimate it: in the world of servers, it is the ultimate carrier capable of withstanding high-frequency signal “sprints.” Its excellent thermal stability, strong chemical inertness, and near-perfect low dielectric constant make it tailor-made for high-speed data transmission. When CITIC Securities’ research report starts championing it, the market recognizes the familiar formula: an undervalued niche sector is about to be ignited by the roaring fire of AI computing.

Yet there is a subtle irony here. We often talk about AI’s “brain” (chips) and “limbs” (algorithms), yet we frequently overlook its “nervous system”—the high-speed cables and backplanes that weave like spider webs inside server chassis. As signal frequencies push toward 112G or even 224Gbps, traditional PCB materials are like trying to run high-speed trains on mud roads, with signal losses becoming intolerably high. PTFE’s low-loss properties make it a “must-have” rather than an “option.” The industry’s hot discussion of the “orthogonal backplane” architecture fundamentally aims for shorter, more direct signal paths—which precisely requires a “clean” material like PTFE to pave the way.

The moves of domestic manufacturers are intriguing. Shengyi Technology, a leading PCB maker, is actively pursuing NVIDIA’s validation. This is not merely about positioning for technological advantage but also reflects a survival anxiety. In the high-end server market, if one fails to secure the next generation of technical standards in advance, it will be too late to join the game by the time Rubin Ultra enters mass production. This reveals a harsh reality: in the AI hardware ecosystem, the window for material suppliers is shorter than that for chip manufacturers, competition is more covert, and losers don’t even get a chance to make the news.

However, amid the euphoria, a dose of cold water is needed. We’ve seen broker research reports’ enthusiasm too many times. Can PTFE really “redefine downstream sectors” as claimed? Its large-scale adoption faces at least two major hurdles: first, cost—PTFE is highly challenging to process and nowhere near as cheap and accessible as FR-4; second, process—the existing PCB manufacturing system needs fundamental adjustments to accommodate it. A more likely path is initial deployment in scenarios where cost is no object, such as supercomputers and top-tier AI training clusters, followed by gradual penetration. Expecting it to appear on consumer-grade motherboards next year is pure fantasy.

But regardless of short-term speculation, this trend strikes at the essence of the AI race: it is a comprehensive arms race ranging from transistors to plastic molecules. When we cheer for every successful tape-out of domestic GPUs, we should also recognize that these silent foundational materials form the skeleton supporting the entire architecture. The story of PTFE reminds us that true self-reliance and control lie not only in designing more powerful chips but also in whether we can secure every meter of high-frequency cable and every high-speed backplane material in our own hands.

So, the next time you hear the term “computing bottleneck,” consider those electrical signals struggling within millimeter-thick layers. Whether they can run faster and more stably may well depend on an apparently ordinary “plastic.” This materials upgrade battle is without smoke or fire, but its outcome may determine who can truly harness the next tide of artificial intelligence.

当英伟达的工程师在实验室里反复测试Rubin ultra服务器背板时，他们真正较量的或许不是算力，而是谁能更快搞定这种叫PTFE的塑料。这不是科幻，这是正在发生的材料革命——算力基建的狂欢，终于烧到了最不起眼的基础材料层。

PTFE，全称聚四氟乙烯，就是那种做不粘锅涂层的东西。但别小看它，在服务器的世界里，它是能扛住高频信号“狂奔”的终极载体。热稳定性好、化学惰性强、介电常数低到近乎完美，简直是为高速数据传输量身定制的。当中信建投的研报开始为它摇旗呐喊时，市场嗅到了熟悉的配方：一个被低估的细分赛道，即将被AI算力的大火点燃。

但这里有个微妙的讽刺。我们总在谈论AI的“大脑”（芯片）和“四肢”（算法），却常常忽视它的“神经系统”——那些在机箱里蜿蜒如蛛网的高速线缆和背板。当信号频率冲向112G甚至224Gbps，传统的PCB材料就像试图用泥巴路承载高铁，信号损耗大到无法忍受。PTFE的低损耗特性，让它成了那个“必选项”而非“可选项”。产业内热议的“正交背板”架构，本质上就是在追求更短、更直接的信号路径，而这恰恰需要PTFE这种“干净”的材料来铺路。

国内厂商的动作很有意思。生益科技作为PCB龙头，积极跟进英伟达的验证，这不仅是技术卡位，更是一种生存焦虑。在高端服务器市场，如果不提前绑定下一代技术标准，等到Rubin ultra量产时再想挤上牌桌，黄花菜都凉了。这揭示了一个残酷现实：在AI硬件生态里，材料商的窗口期比芯片厂商更短，竞争更隐蔽，失败者连上新闻的机会都没有。

然而，狂欢之下需要一点冷水。券商研报的热情，我们见过太多次。PTFE真的能像宣称的那样“重新定义下游领域”吗？它的大规模应用至少面临两座大山：一是成本，PTFE加工难度高，完全不像FR-4那样便宜易得；二是工艺，现有的PCB制造体系需要为它做出根本性调整。更可能的路径是，在超算、顶级AI训练集群等不计成本的场景率先落地，然后缓慢渗透。指望它明年就出现在消费级主板上，纯粹是幻想。

但无论短期炒作与否，这个动向戳中了AI竞赛的一个本质：它是一场从晶体管到塑料分子的全方位军备竞赛。当我们为国产GPU的每一次流片成功欢呼时，更应看到这些沉默的基础材料才是撑起整个架构的骨架。PTFE的故事提醒我们，真正的自主可控，不仅在于设计出更牛的芯片，也在于能不能把每一米高频电缆、每一块高速背板的材料都握在自己手里。

所以，下次听到“算力瓶颈”这个词时，不妨想想那些在毫米级厚度里挣扎的电信号。它们能否跑得更快、更稳，答案可能就藏在一种看似普通的“塑料”里。这场材料升级战没有硝烟，但它的胜负，或许将决定谁能真正驾驭下一代人工智能的洪流。