<div id="readability-page-1" class="page"><div><h2>Meta Just Ordered $6 Billion in Fiber Optic Cable From Corning. The Real AI Bottleneck Isn't Software.</h2><p>When Meta announced it would source roughly $6 billion in fiber optic cables from Corning through 2030, Corning's stock surged 16% in a single day. That's one data center customer, one supply contract, and a bigger move than most AI software companies generate with a product launch. It tells you something about where the real constraint is.</p><p>Every AI headline in 2026 focuses on the software: agents, models, job displacement. But AI runs on hardware, and the hardware runs on networks. The bottleneck in AI infrastructure has shifted three times in two years: first GPUs, then power, now connectivity. The companies building the physical layer that connects hundreds of thousands of GPUs into functioning AI factories are becoming the defining investment of this cycle.</p><h3>The Copper Wall</h3><p>Inside data centers, the connection between GPUs still runs largely on copper wiring. As AI clusters have grown from a few thousand GPUs to hundreds of thousands, that copper is hitting physical limits. Dell'Oro Group's VP Sameh Boujelbene told Fierce Network that AI clusters have expanded from thousands to hundreds of thousands of GPUs per cluster, with the future potentially reaching a million per cluster. As those numbers scale, interconnect requirements &#34;scale exponentially.&#34;</p><p>Copper loses signal quality over short distances, wastes energy, requires costly signal boosters, and generates excess heat. Corning's VP of Global Research Claudio Mazzali called the transition from copper to fiber inside data centers &#34;inevitable,&#34; comparing it to the same transition that already happened in submarine, long-haul, metro, and access networks over the past two decades.</p><p>Dell'Oro Group forecasts that AI back-end networks will grow at more than 50% annually to reach $20 billion by 2028. The Optical Transport market posted a 15% year-over-year gain in Q3 2025, and Dell'Oro raised its outlook for both 2025 and 2026. Network backbone capacity growth is back above 25% after stalling for nine quarters.</p><h3>What &#34;Scaling Across&#34; Actually Means</h3><p>The most important infrastructure concept in 2026 is one most investors haven't heard of: scaling across.</p><p>Hyperscalers are no longer building single data centers. They're building gigawatt-scale AI campuses where multi-hundred-thousand GPU clusters span multiple buildings within a region, all treated as a single AI factory. This requires massive fiber deployments between facilities, often up to 100 km apart, using 800 ZR+ optics to tap into different electricity grids.</p><p>Inside these campuses, the fiber density is unlike anything the industry has seen before. Instead of sub-1,000 fiber runs, operators now require thousands of fibers per run and multi-conduit deployments stacking into the hundreds of thousands. Modern GPU clusters specify 200 terabits per second of east-west bandwidth, which translates into installations exceeding 3,000 fibers per fabric link. STL's CEO Rahul Puri told Fierce Network that AI-focused data centers require roughly 36 times more fiber than traditional CPU-based racks.</p><p>The U.S. alone will need to add 213 million more fiber miles by 2029, more than doubling its current 160 million miles. Microsoft has secured more than $8 billion in dark-fiber contracts to reinforce its cloud backbone. At least one major fiber manufacturer has already sold its entire fiber inventory through 2026.</p><h3>Co-Packaged Optics Is the Next Transition</h3><p>The current architecture converts light to electricity at the server's face plate, then sends data through copper to the chip. Co-packaged optics (CPO) keeps data in optical form all the way to the GPU or switch chip itself, eliminating the inefficiency of that last copper link.</p><p>Industry estimates suggest CPO could cut optical component costs by 40%. Some chip makers project more than 50% power savings compared to current technology. Nvidia unveiled silicon photonics-based CPO technology at GTC in March 2025, and Broadcom announced 400G/Lane-class CPO solutions. Corning's Brian Rhoney predicts that 2026 will see the first field trials of fiber-fed scale-up fabrics linking multiple GPUs across multiple racks into a single computing &#34;brain.&#34;</p><p>Corning's own CFO-level view is that CPO will hit actual AI cluster deployments in 2026 after years of demos and pilots. The company's FlexConnect fiber is optimized for the short-length configurations CPO requires. Its GlassWorks AI product suite targets the entire AI network stack, inside and outside the data center. Enterprise sales grew 58% year-over-year in Q3 2025, driven largely by AI network demand.</p><h3>The Sovereign AI Dimension</h3><p>The EU AI Act's transparency requirements are accelerating sovereign AI infrastructure builds across Europe and Asia. Nations are racing to build domestic AI clouds, which means redundant data center capacity in each jurisdiction. Every sovereign AI cluster needs the same fiber, the same optics, the same physical connectivity.</p><p>This creates sustained, geographically diversified demand for infrastructure companies. It's not dependent on any single hyperscaler's capex cycle. Nokia announced a $4 billion investment in U.S. R&amp;D and manufacturing for AI-ready network technologies. Ciena raised its fiscal 2026 revenue guidance to $5.7-$6.1 billion, a significant increase that analysts tie directly to AI-driven demand. The silicon photonics market is forecast to grow from $2.16 billion in 2024 to $7.52 billion by 2029.</p><h3>DeepSeek Changed the Demand Curve, Not the Infrastructure Curve</h3><p>DeepSeek's success in achieving competitive performance at lower cost does not reduce infrastructure demand. It increases it. If high-level AI reasoning becomes affordable, more companies deploy it, which means more inference traffic, more data center load, more fiber. The bottleneck shifts from &#34;who can afford to train a model&#34; to &#34;who has the physical capacity to run the models at scale.&#34; Cheaper intelligence is bullish for the companies that build the network carrying that intelligence.</p><h3>What the Trade Show Actually Showed</h3><p>Big Data &amp; AI World at Tech Show London (March 4-5, 2026) was a general enterprise conference, not the watershed moment the original article described. The shift from chatbots to agentic AI is real, documented extensively across industry research, and accelerating. But the specific claim that &#34;the chatbot era ended this week&#34; at a London trade show overstates what happened there.</p><p>What is genuinely new in early 2026 is the infrastructure investment response to the agentic shift. Agentic AI systems that execute multi-step workflows generate significantly more compute and data traffic than single-query chatbots. The March 16-19 NVIDIA GTC in San Jose is expected to introduce the Feynman architecture, an inference-focused chip designed specifically for agentic workloads. Every new inference chip needs a network to connect it to. The physical layer scales with every software advance.</p><h3>The Infrastructure Thesis</h3><p>Software captures the headlines. Infrastructure captures the revenue. Corning's John McGirr told Fierce Network that hyperscale and AI network loads have &#34;significantly increased&#34; fiber demand expectations. Corning is developing co-packaged optics and glass substrates for opto-electronic circuits. Its Contour Flow Cable fits double the fiber into existing cable diameters. The company is also advancing glass core technology for semiconductor packaging, extending its reach from the network to the chip.</p><p>The AI infrastructure buildout is not a one-year trade. The U.S. needs to more than double its fiber miles by 2029. Hyperscalers are signing multi-year, multi-billion-dollar supply contracts. Data center connectivity demand is growing at 50%+ annually. The companies building this physical layer, Corning, Ciena, Nokia, Prysmian, and the broader optical supply chain, are selling into a demand curve that gets steeper with every model improvement, every agent deployment, and every new sovereign AI mandate.</p><p>The foundation of the AI era is literally under construction. The companies laying cable don't need to win the model race. They just need the race to continue.</p><p>If you liked this, you'll love the daily dose we send to our Substack crew. It's chill, it's sharp, and it actually helps. Hop on here.</p></div></div>

康寧

Global X未來分析技術ETF

<p>Meta 已向康寧訂購了 60 億美元的光纖電纜，導致康寧的股票上漲了 16%。隨着人工智能集羣的擴展，從銅纜轉向光纖在人工智能數據中心中變得至關重要。預計光纖的需求將顯著增長，美國到 2029 年需要增加 2.13 億英里光纖。共同封裝光學（CPO）技術預計將提高效率並降低人工智能基礎設施的成本。此外，歐盟人工智能法案正在推動主權人工智能基礎設施的發展，增加了各個地區對光纖和光學的需求</p>

The Nervous System Of AI: Why Fiber Optics Is Driving Tech’s Next Wave

- Meta 與 Corning 簽署價值 60 億美元的光纖合同，推動光纖基礎設施需求。  
- AI 基礎設施瓶頸從 GPU 轉向連通性，未來需要更多光纖支持。  
- 企業基礎設施投資正加速，全球 AI 數據中心將需增加 2.13 億光纖英里。

福布斯

Janus Henderson Global Artfcl Intlgc ETF

iShares Future AI & Tech ETF

Wisdomtree Artificial Intelligence And Innovation Fund - ETF

FRNKL EXP DT ETF

Defiance AI & Power Infrastructure ETF

iShares US Digital Infrastructure and Real Estate ETF

Tortoise AI Infrastructure ETF

2 倍做多 META ETF - GraniteShares

Global X Data Center & Dgtl Infrs ETF

人工智能的神經系統：為何光纖技術正在驅動科技的下一波浪潮