The giants are all grabbing chips.

In the first few months of 2026, the pace of mergers and acquisitions in the optical chip industry was so intense that it was somewhat dizzying. On March 2, Nvidia announced investments of $2 billion each in Lumentum and Coherent; three weeks later, they made another $2 billion investment in Marvell; in the same month, Credo Technology announced the acquisition of Israeli silicon photonics company DustPhotonics for $750 million. Just a few months ago, Marvell had completed the acquisition of Celestial AI for a record $5.5 billion. Adding up these transactions, the capital flowing into the optical interconnect field in just six months exceeded the billion-dollar level. A large number of acquisition investments are all pointing in the same direction - optical chips. The rapid expansion of AI computing power has pushed optical chips to the core bottleneck of whether the entire computing system can operate normally. Whoever can establish a barrier at this point will have a moat in the upcoming years of the AI infrastructure competition. Copper wires have reached their limits For AI data centers, traditional copper wires are becoming increasingly inadequate. As the scale of artificial intelligence computing continues to expand, data centers are facing serious power consumption and transmission bandwidth bottlenecks. The transmission distance and efficiency of traditional plug-and-play optical modules have gradually approached the physical limits. This is not an exaggeration. The physical properties of copper wires determine that their transmission loss increases exponentially with frequency. When 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 system's total power consumption, becoming a bottleneck in overall performance. At this point, optical interconnect comes into play. CPO (Co-Packaged Optics) integrates high-speed optical engines with switch chips or large-scale AI computing chips on the same package substrate using advanced packaging technology. This keeps high-speed electrical signal transmission limited to distances of millimeters, with medium to long-distance transmission handled by optical fibers. This approach addresses the shortcomings of traditional plug-and-play optical modules, such as high power consumption, signal loss, and limited bandwidth. Compared to traditional solutions, CPO can reduce power consumption by over 40%, increase bandwidth by 3 times, and reduce latency by 50%. In real-world terms, this means that a large-scale data center operating tens of thousands of GPUs can save hundreds of millions of dollars in electricity costs annually while achieving higher training throughput. Silicon photonics and CPO are not new technologies. Intel began laying the groundwork in this area over 20 years ago and has already shipped over 8 million silicon optical transceivers. Broadcom has also been one of the early promoters of CPO standards. However, for quite some time, there has been a gap between this technology and large-scale commercial deployment. The leap in AI demand for computing power has now brought the market to expect 2026 to be a crucial starting point for the commercial deployment of CPO. In fact, the influx of orders has already given confirmation. Optical and photonics product manufacturer Lumentum expects CPO-related revenue to reach $50 million in the fourth quarter of 2026, with several hundred million dollars of CPO-related orders converting to revenue in the first half of 2027. These orders cover multiple customers, not just a single contributor. Lumentum's CEO also revealed in an interview that despite the company expanding its core manufacturing capacity in Japan by 12 times in the past 24 months, it continues to lag behind demand, with its indium phosphide laser production capacity fully booked and a waiting period of up to 32 months. Almost overnight, optical chips have become a battleground for giants. Nvidia, Locking up the Upstream with $6 Billion As one of the most powerful and cash-rich giants, Nvidia's investment in Lumentum and Coherent of $4 billion together was a bold and generous move. They saw the vulnerability behind the industry chain. The manufacturing of high-end EML chips depends on indium phosphide (InP) substrates. Unlike silicon-based semiconductors, the crystal growth of compound semiconductors like InP is extremely challenging, making yield control very challenging. With the advancement of AI data center construction, the demand for indium phosphide has shown an explosive growth of 40% to 50% annually, but due to the long 18-month machine verification and validation cycle, it is difficult to quickly expand the production capacity of InP substrates in the short term. It is in this background that Nvidia announced investments in Lumentum and Coherent. The agreement with Lumentum includes Nvidia's multi-billion dollar purchasing commitment and future access to high-end laser components; while the agreement with Coherent also includes multi-billion dollar purchasing commitments and future access and production rights to advanced laser devices and optical network products. It is worth noting that both agreements are non-exclusive, but the binding effect of production access rights far exceeds that of regular purchasing contracts. The investment will help Lumentum expand its manufacturing capacity and research and development to meet future AI data center demands. Lumentum also plans to build a new wafer production facility for this purpose. Coherent will also expand its manufacturing base in the United States. In short, Nvidia has exchanged $4 billion for the priority use of the most scarce upstream optical chip production capacity in the coming years, ensuring that competitors will find it more difficult to obtain the same under the same conditions. Subsequently, Nvidia made an additional $2 billion strategic investment in Marvell, officially integrating their cooperation into the NVLink Fusion rack-level infrastructure platform. This money was used not only to purchase products, but also to secure Marvell's customized AI chips and silicon photonics technology in a deeply integrated position within Nvidia's ecosystem. It is also worth noting that Nvidia had previously participated in the Series D ($155 million in 2024) and Series E ($500 million in 2026) financing rounds of the optical chip unicorn Ayar Labs, focusing on the potential of their TeraPHY optical engine to replace traditional copper interconnects. The combination of these transactions has allowed Nvidia to secure its position in the optical interconnect industry chain in just a short span of two to three years. Marvell, Betting $5.5 Billion on Photonic Mesh As one of Marvell's investment targets, Marvell, like Nvidia, has completed its layout after acquiring Celestial AI, seeing the more fundamental technology behind it. The core contradiction facing AI training clusters is that the growth rate of GPU computing power far exceeds that of memory bandwidth, known as the memory wall. As the scale of model parameters continues to increase, the time each GPU spends waiting for data to be read from memory becomes a larger proportion, resulting in a waste of computing power in waiting. The traditional solution is to stack more HBM memory chips, but the limitation of physical packaging area will soon be reached. The path proposed by Celestial AI is to directly connect processors to memory using optical signals, allowing thousands of GPUs to access pool memory as if it were local memory. Celestial AI's photonic fabric technology, which uses optical signal transmission, has an efficiency that is over two times that of copper interconnects, with nanosecond-level latency and excellent thermal stability, enabling it to reliably operate in extreme thermal environments of several thousand watts of XPU power, supporting 3D vertical stacking to increase XPU's HBM capacity. Its first photonic chip has a single-chip bandwidth of 16Tbps, which is ten times that of existing mainstream solutions. According to the terms of the transaction, Marvell will pay $1 billion in cash and 27.2 million Marvell common shares worth $2.25 billion, with an important earn-out clause: if Celestial AI can achieve cumulative revenue of $2 billion by the end of the 2029 fiscal year, Marvell will pay additional amounts to reach a total transaction value of up to $5.5 billion. At the same time, Marvell issued stock options to Amazon through its subsidiary Digimoc Holdings, allowing Amazon to purchase Marvell's stock based on its purchases of products by the end of 2030. Amazon could receive up to $90 million worth of Marvell stock. This arrangement effectively binds the most important customer into the deal, while also serving as an endorsement of Amazon AWS for this technological direction. Marvell estimates that the photonics interconnect market in the AI and data center fields could reach $10 billion. It is expected that Celestial AI will start contributing revenue in the second half of the 2028 fiscal year, reaching annual revenue of $1 billion by the fourth quarter of the 2029 fiscal year. This is a bet that will take more than two years to see a return, but Marvell clearly believes that the timing is crucial. Once the landscape of optical interconnect is set, the window for newcomers will be very limited. Broadcom's Play Unlike Nvidia and Marvell, who quickly filled their gaps through acquisitions, Broadcom is taking a route that emphasizes independent control. Broadcom is the world's first commercial CPO switch manufacturer, and its Bailo platform integrates optical engines with Tomahawk 5 series switch chips, offering significant advantages in efficiency over traditional solutions. At OFC 2026, Broadcom further showcased the industry's first 400G/lane optical DSP (code-named Taurus), paving the way for 1.6T and even 3.2T networks. Broadcom's competitive advantage comes more from its system-level integration capabilities and its deep, long-standing collaboration with top cloud companies like Google, Meta, and others. By 2027, Broadcom expects AI-related business revenue to reach $60 to $90 billion, with optical interconnect technology at its core in the DRIVE. With this scale, Broadcom is not in a hurry to fill in technology gaps through large-scale acquisitions. Instead, relying on internal research and development and the support of TSMC's advanced packaging platform, has been sufficient to maintain its dominance in the field of switch chips. Intel, the Mighty Old Giant with Great Potential Let's talk about Intel, the pioneer and leader in the field of silicon photonics. Intel has accumulated over 20 years of deep technical expertise in this field, witnessing and driving the entire process of optical chip technology from the laboratory to industrialization. In recent years, Intel has adjusted its strategic focus, transitioning from a pure supplier of optical modules to a core provider of silicon photonics components and integrated chips (OCI), consolidating its core competitive edge in the high-end optical chip field through strategic acquisitions, investments, business optimization, and technological breakthroughs. At the core acquisition level, as far back as 2019, Intel acquired Barefoot Networks, which became a key move for Intel in the optical chip field, combining the latter's programmable Ethernet switching chips with its own silicon photonics engine to successfully launch the industry's first Co-Packaged Optics (CPO) solution, featuring an integration of 1.6 Tbps silicon photonics engine with a 12.8 Tbps switch chip. In 2023, although Intel terminated its $5.4 billion acquisition plan for the world's top silicon photonics foundry Tower due to regulatory reasons, they reached a close foundry agreement, which further strengthened Intel's integrated IDM manufacturing capabilities in the field of photonic chips, filling the core gap in the foundry segment. At the strategic investment level, similar to Nvidia, Intel has been an early and core investor in Ayar Labs, participating in the company's Series C ($130 million in 2022), Series D ($155 million in 2024), and Series E ($500 million in 2026) rounds, focusing on supporting their TeraPHO optical engine technology. In the face of financial pressures and industry competition, Intel has also undergone strategic contraction and restructuring of its photonics business, transferring the production and sale of pluggable optical transceivers to Jabil in 2023. By divesting the lower-threshold, lower-profit standardized optical module assembly business, Intel can save $1.8 billion in annual expenses, concentrating resources on its core areas. On the technological breakthrough level, in 2024, Intel showcased the first fully integrated OCI (Optical Compute Interconnect) chip at the Optical Fiber Communication Conference (OFC), supporting bidirectional transmission of 4 Tbps per second, achieving an ultra-high efficiency level of 5 pJ/bit, and support for direct co-packaging with CPU/GPU, completely solving the data traffic bottleneck in AI clusters and providing core support for high-performance AI infrastructure. This OCI chip integrates silicon photonics integrated circuits (PICs) with on-chip lasers, optical amplifiers, and electronic integrated circuits, supporting 64 channels of 32Gbps each, with transmission distances of up to 100 meters. Its energy efficiency is increased by approximately 67% compared to traditional pluggable optical transceiver modules, effectively alleviating the high power consumption issue in AI applications and promoting the elevation of optical interconnect technology to chip-level integration. Overall, Intel's approach in the field of optical chips is to discard the low-threshold, low-profit margin business of standard optical module foundries, invest continuously in startups like Ayar Labs to lock in cutting-edge optical I/O technology, and leverage its IDM advantage in Intel Silicon Photonics to focus its efforts on conquering core areas of high-end chip-level optical transmission such as CPO and OCI, steadily consolidating its leading position in the silicon photonics field. TSMC, the Downstream Positioning As the largest foundry, TSMC has not expanded aggressively in the field of optical chips, but has relied on its irreplaceable foundry advantage and advanced packaging capabilities to position itself as a core provider of global AI photonic integrated infrastructure. One of the most noteworthy developments is the silicon photonics engine COUPE platform created by TSMC. This platform uses SoIC-X chip stacking technology to stack electrical and photonic chips directly on top of each other in a 3D stack. Compared to traditional packaging, the COUPE platform has significantly lower impedance, a 5 to 10 times increase in efficiency, a 10 to 20 times reduction in latency, meeting the demands of high bandwidth, low power consumption, and low latency in AI scenarios. TSMC successfully validated small pluggable optical modules in 2025 and began large-scale production in 2026. It also integrated the COUPE optical engine into its CoWoS packaging in CPO form, a crucial step in the successful deployment of CPO. To standardize and lower the design threshold for silicon photonics industry, TSMC joined forces with more than 30 upstream and downstream companies, including industry giants like Foxconn, MediaTek, and Quanta, to establish the SEMI Silicon Photonics Industry Alliance (SiPhIA). This alliance aims to unify global silicon photonics technology specifications, set industry standards, promote the mainstream adoption of "optical over copper," and address some key challenges in the large-scale production of CPO. It is clear that TSMC's positioning in the optical chip field is very clear: not engaging in end-optical module manufacturing, but focusing on its core strengths, using the COUPE platform to become the standard setter and core foundry for the optical chip era. Credo: $750 Million Links the Final Puzzle Piece Compared to the mainstream players, Credo Technology is much less well-known. However, the acquisition they announced on April 13 is also worth serious consideration. Credo is the inventor of Active Electrical Cables (AEC) technology, and its ZeroFlap optical transceiver has established a stable customer base in the AI cluster field. However, they have been reliant on external suppliers for the key process of silicon photonics integrated circuits (PICs), resulting in a lack of complete vertical integration in their product range. The acquired company, DustPhotonics, founded in Israel in 2017, is a fab-less company. DustPhotonics has developed a differentiated portfolio of SiPho PIC products covering 400G, 800G, and 1.6T, with a technology roadmap extending to 3.2T, supporting integrated and external laser configurations, a team of approximately 70 people, with deep expertise in photonic integration. The core value of their products lies in integrating multiple optical functions onto a single silicon chip, reducing component complexity, improving manufacturing yield, and as data rates exceed 800G, cost advantages will become increasingly significant. This acquisition will give Credo the ability to span across SerDes, digital signal processing, silicon photonics, and system integration, covering both electrical and optical connections in expanding networks. In other words, Credo, which has been relying on copper technology until now, has finally completed its core capabilities in photonics by acquiring DustPhotonics. Credo expects the combined optical business portfolios to generate over $500 million in optical revenue by the end of the 2027 fiscal year. With a market cap of around $24 billion, Credo's stock has risen by over 200% in the past year. The $750 million cash acquisition, along with stock options, makes this the largest bet the company has ever made. More Followers, Already Here? Apart from the mainstream players, other giants are also accelerating their entry into the optical chip market. Let's start with AMD; their optical chip layout is not much slower than Nvidia's. In May 2025, AMD completed the acquisition of Silicon Valley silicon photonics startup Enosemi, whose core team has top-notch technology expertise and talent in the field of photonic integrated circuits (PICs) production and has had a long-term collaboration with AMD in photonics. Brian Amick, AMD's Senior Vice President of Technology and Engineering, said that the core value of this acquisition is to accelerate the innovation iteration of Co-Packaged Optics (CPO) technology, creating high-bandwidth, low-power interconnect solutions for the next generation of AI systems and overcoming the bottleneck of traditional electrical interconnects in high-speed transmission. In addition to direct acquisitions, AMD, as a core strategic investor in Ayar Labs, participated in the company's $500 million Series E financing round in March 2026. The core goal of their cooperation is to integrate Ayar Labs' optical I/O chiplets into AMD's Instinct AI accelerator, replacing traditional copper interconnects with optical transmission to significantly increase the per-watt throughput of AI systems. Similarly, as a giant in the global mobile chip industry, MediaTek has also started its foray into the field of optical chips. In February 2026, MediaTek invested approximately $90 million through its subsidiary Digimoc Holdings to acquire a stake in leading silicon photonics company Ayar Labs, obtaining about 2.4% of the company. The key strategic goal of this investment is to deepen their collaboration with top companies in the field of optical interconnects, focusing on application-specific integrated circuits (ASICs) and high-speed optical interconnect technologies to build core capabilities for the technological upgrade of their AI and communication businesses. It is worth noting that in the AI data center field, MediaTek plans to apply silicon photonics technology to AI ASIC chips and is expected to secure more orders from giants in the future. Other companies within the industry are also accelerating their movements in the silicon photonics field and driving the industrial application of optical interconnect technologies. Astera Labs, a provider of semiconductor interconnect solutions for rack-level AI infrastructure, recently announced that it has reached a final agreement to acquire aiXscale Photonics. Astera Labs stated that this acquisition is expected to combine aiXscale's glass optical coupler technology with its own interconnect and signal processing products, helping them develop optical-scale extension solutions to unleash the full potential of rack-level AI deployment. Sanjay Gajendra, Chief Operating Officer and President of Astera Labs, stated: "The transition to AI infrastructure 2.0 requires optical solutions specifically designed for the