Three months, five companies, a total investment of 6.5 billion dollars! Nvidia is leading a revolution in AI transformation.

Since entering 2026, the global computing power industry chain is undergoing a profound bottom-up reconstruction. As AI training clusters scale from tens of thousands of cards to hundreds of thousands and even millions of cards, traditional electrical interconnection technologies - copper cables and pluggable optical modules - are approaching physical limits. Power consumption surge, signal attenuation, and latency accumulation are becoming core bottlenecks restricting computing power release. Facing this challenge, chip giant NVIDIA Corporation (NVDA.US) is betting heavily on a new direction: silicon photonics and Co-Packaged Optics (CPO) technology. Three months, at least $6.5 billion: NVIDIA Corporation's "lightning-fast layout" In the spring of 2026, NVIDIA Corporation's investment pace in optical interconnection field was so intense that it dazzled the market. On March 2nd, NVIDIA Corporation announced investments of $2 billion each in the U.S. optical technology companies Lumentum (LITE.US) and Coherent (COHR.US), totaling $4 billion on the same day. The non-exclusive agreement with Coherent includes billions of dollars in purchasing commitments and future capacity rights for advanced laser components; at the same time, NVIDIA Corporation also supported Lumentum's research and capacity building, assisting in the construction of a new wafer factory in the U.S. Three weeks later, NVIDIA Corporation once again invested $2 billion in Marvell Technology, Inc. (MRVL.US), with both parties jointly developing silicon photonics technology and customized XPU solutions. In early May, NVIDIA Corporation also announced a $500 million investment in fiber optic giant Corning Inc (GLW.US), gaining priority access to its advanced optical connectivity capacity. Corning Inc plans to increase its optical connection product capacity in the U.S. to ten times the current level and expand its fiber capacity by over 50%. In addition, NVIDIA Corporation participated in Ayer Labs' Series E financing of up to $500 million, with the optical startup's valuation entering the unicorn ranks. This series of investments has been interpreted by the market as a highly systematic grand layout: NVIDIA Corporation covers the three levels of capabilities from photon chips to system interconnection - optical components (Coherent and Lumentum), optical interconnection architecture (Marvell), and optical I/O chips (Ayer Labs) through strategic investments rather than acquisitions. Analysts point out that NVIDIA Corporation is using its huge profits to build a strong ecosystem to support the development of advanced AI technology. By directly investing in key optical component companies, the overall market demand for its computing infrastructure is enhanced. Approaching the power wall: Why we must shift from "electricity" to "light" Behind this series of heavy investments is the imminent power consumption and bandwidth crisis in AI data centers. The physical limits of copper wires are emerging. According to industry analysis, as the scale of AI operations continues to expand, data centers are facing severe power consumption and transmission bandwidth bottlenecks. The physical properties of copper wires determine that transmission losses increase exponentially with frequency, and as the cluster size expands from thousands of GPUs to tens of thousands or more, the connections between chips begin to consume an increasingly large share of the total system power, becoming a bottleneck for overall performance. In contrast, the advantages of optical interconnection are significant. CPO integrates high-speed optical engines with switching chips or AI computing chips through advanced packaging technology on the same package substrate, limiting the transmission of high-speed electrical signals to a millimeter-level range, with longer-distance transmissions being handled by optical fibers. Compared to traditional pluggable optical module solutions, CPO can reduce power consumption by over 40%, increase bandwidth by 3 times, and shorten latency by 50%. When applied to large-scale data centers, this means saving hundreds of millions of dollars in electricity costs annually, while achieving higher training throughput. Brian Colello, senior equity analyst at Morningstar, points out that while copper cables are currently the main connection solution due to their low cost and high reliability, photonics technology will become increasingly important in AI infrastructure. "NVIDIA Corporation's roadmap for the next generation of AI rack-level solutions will require more and more optical interconnection to meet the exponential growth in bandwidth demand brought about by new models and higher usage." At the ISSCC 2026 conference held recently, NVIDIA Corporation once again emphasized the essential difference in interconnection requirements between AI data centers and traditional cloud data centers - the inherent characteristics of electrical interconnections are becoming one of the three core bottlenecks in computing power release. "Photonics technology allows NVIDIA Corporation to expand its AI infrastructure while avoiding the high energy costs associated with continued reliance on electricity and copper cables," said Forrester senior analyst Alvin Nguyen in an interview. "By investing in photonics technology companies, NVIDIA Corporation ensures ongoing progress in this area, avoiding expansion and performance bottlenecks due to cling to electricity and copper cable technologies." 2026: The "industrialization year" of CPO from concept to order In fact, silicon photonics and CPO are not entirely new concepts. Intel Corporation (INTC.US) began laying out in this area over 20 years ago and has now shipped over 8 million silicon photonics transceivers; Broadcom Inc. (AVGO.US) is also one of the early advocates of CPO standards. However, what truly brought this technology to a turning point was the exponential increase in AI's demand for computing power. The market widely expects 2026 to be a key starting point for CPO commercial deployment, with Sinolink explicitly stating in a recent research report, "2026 is the industrialization year of CPO." According to a survey by TrendForce, by 2030, CPO penetration in optical communication modules for AI data centers is expected to increase from approximately 0.5% to 35%. In terms of industry rhythm, Scale-out scenarios are expected to be deployed first in 2026, with potential extension to Scale-up scenarios in 2027-2028. The industrialization expectations of CPO have triggered a chain reaction in both the primary and secondary markets. The shortage in the supply chain is intensifying. Lumentum revealed that its Indium Phosphide laser capacity has been fully booked, with a scheduling period of up to 32 months, and the supply-demand gap continues to widen. Tower Semiconductor Ltd (TSEM.US), an Israeli silicon photonics foundry, saw its silicon photonics business revenue in the first quarter of 2026 triple year-on-year and signed long-term supply contracts worth $1.3 billion, with over 70% of its silicon photonics total capacity booked or in the booking process until 2028. Indium Phosphide substrate is monopolized by three companies - Sumitomo in Japan, Tong Mei in Beijing, and JX in Japan - with over 90% market share. The delivery cycle of high-end EBL equipment generally exceeds one year - these are practical bottlenecks that restrict downstream expansion. The capital markets have also turned enthusiastic. Lumentum's stock price surged 339% throughout 2025 and has risen by approximately 140% this year, approaching a market value of $70 billion and being included in both the S&P 500 and Nasdaq 100 indices. Coherent's stock has risen by about 110% since the beginning of the year, reaching a historic high. Marvell has seen a 140% increase since the beginning of the year, and Corning Inc has risen by 110%. In the Taiwan and A-share markets in China, silicon photonics concept stocks are also one of the most sought-after sectors in this round of AI infrastructure boom. However, the industrialization year does not mean that large-scale deployment is imminent. Nick Patience, head of AI business at the Futurum Group, said, "The technology itself is reliable, but the challenge lies in manufacturing at scale. The production yield of complex co-packaged optical components is still a major challenge because of the low fault tolerance between optical and silicon-based components. Once a problem occurs during the packaging process, it is usually impossible to rework. "The industry transformation process has already begun but is still in its early stages. I expect large-scale applications to begin around 2028," Patience added.