On August 14, 2026, optical interconnect technology entered NVIDIA's product line for the first time in mass production form: its Spectrum-X Ethernet silicon photonic switch entered full-scale manufacturing, with products having been in production between May and July of that year. This marks the world's first mass-produced switch built on co-packaged optics (CPO) technology—placing optical engines directly beside the chip—and serves as a core component of NVIDIA's next-generation AI supercomputing clusters.

In recent years, NVIDIA, which has enjoyed remarkable success, has faced repeated delays and adjustments, revealing significant headwinds. The new flagship computing platform Vera Rubin was postponed due to SK Hynix HBM4 memory issues; the four-memory-die high-end version of the Rubin platform was cut to two dies at launch because of TSMC interposer yield problems; and the rack-level solution Kyber NVL144, critical to next-generation computing clusters, has been pushed back to 2028 due to challenges with a one-square-meter PCB. The reasons for these setbacks vary, but they share a common underlying logic: with Moore's Law for pure semiconductor chips having reached its limits, components surrounding the chip in top-tier AI clusters are now taking on the responsibility of high-speed inter-chip communication, effectively combining them into a virtual "mega-chip."

These surrounding components, pressed into service, have expanded computing power but also caused efficiency declines: when thousands of GPUs train a large model together, data movement consumes 30% to 50% of training time. Even more striking is the energy perspective—data centers spend 90% of their energy on moving data, with only 10% going to actual computation. Looking ahead, the challenges will only grow: chip computing power scales with area, while interconnect bandwidth scales only with edge length (chip I/O is typically distributed around the chip's perimeter), creating a gap that grows by a factor of the square.

Optical interconnect is precisely the solution NVIDIA, as the industry leader, has proposed. In March 2026, NVIDIA directly invested $6 billion across three optical communications companies, with Jensen Huang elevating the interconnect layer to "the critical path determining NVIDIA's revenue." The Spectrum-X co-packaged optical switch showcases the latest evolution in NVIDIA's AI clusters: through the application of the new switch, per-port power consumption drops from 30 watts to 9 watts, the number of lasers is reduced by four times, and the entire cluster's switching side is upgraded to optical communications.

At GTC Taipei this June, Jensen Huang also revealed NVIDIA's next step in optical communications: leveraging CPO to advance micro-ring modulator (MRM) technology maturity, ultimately moving toward optical input/output (OIO). The most critical component in the entire system is the silicon photonic micro-ring modulator, which offers the smallest size and lowest power consumption among current optical communication modulators, delivering the highest density deployment capability. Thanks to the MRM's integration, the switch's total bandwidth exceeds 400Tb/s—transmitting 5TB of data in a blink of an eye (0.1 seconds)—while improving signal quality by dozens of times across the communication process.

The improvement on the switching side is only the first step. According to the optical communication technology roadmap, micro-ring modulators will ultimately lead to the OIO stage of chip-to-chip direct connection, truly linking tens of thousands, hundreds of thousands, or even larger numbers of high-performance chips together, potentially bringing a new exponential leap in AI computing power. This goal may sound distant, but Chen Chao, co-founder and CEO of OptoLink, disagrees: "OIO will expand into the entire supernode architecture, and the timeline will be faster than most people think—three to five years is sufficient." OptoLink recently completed a Series A+ funding round of nearly 1 billion RMB, becoming the newest unicorn in the global optical interconnect sector. The company's ultimate goal is precisely the OIO, the final form based on micro-ring modulators.

Four People All-In on OIO

If optical communications are compared to building new highways, the difference between generations lies mainly in the "access points." The future generations of optical communications—LPO (linear-drive pluggable optics), NPO (near-package optics), CPO (co-packaged optics), and OIO (optical input/output)—roughly correspond to "suburbs," "edge of the city center," "downtown," and "at the doorstep of every home." In practical deployment, current pluggable solutions involve inserting a module shaped like a large USB drive at the back of the rack, while OIO's ultimate goal is to package this structure directly onto the core computing chip, linking core computing chips with optical pathways to create an independent high-speed connection that reaches right up to the chip's doorstep. Because the optoelectronic conversion point sits close to the chip core with near-zero distance, interconnect efficiency approaches the theoretical limit.

The optical communications industry's original plan was to advance steadily, moving forward one generation every few years. But the AI boom has caused demand to surge: with massive orders and investment pouring in, a leapfrog approach has become the preferred choice. OIO is the ultimate and most cutting-edge—once achieved, its value is enormous, but the difficulty is correspondingly greatest. Because the entire optical communication system must be compressed from an external discrete component into a chip-level product, OIO can no longer rely on the traditional industry division of labor, where components like light sources, modulators, and DSPs are produced separately and then assembled. It requires simultaneously conquering optoelectronic chip co-design, process yield, multi-wavelength light sources, advanced packaging, and other key technical areas, with each step tightly coordinated to ultimately deliver a chip-level product.

The demand accompanying the "most cutting-edge" positioning is equally cutting-edge, which in turn poses immense challenges to human capability. Zhang Weiwei, co-founder of OptoLink, candidly admits: "No expert in any single field can independently cover all areas. Even at the world's top research institutions, no one has ever single-handedly completed the entire chain." This makes integrating top talent in the OIO track and fusing them into a cohesive fighting team the core challenge of OIO entrepreneurship. The four co-founders of OptoLink are the best example of this logic: they met over many years, each with different life trajectories, but ultimately came together for the shared technical ideal of OIO. All four describe the process almost identically: "We have an extreme reverence for technology, as well as an entrepreneurial mindset to transform top-tier lab results into industrial strength. Therefore, we trust each other, and we are highly aligned in both technical judgment and commercial beliefs."

Chen Chao, co-founder and CEO of OptoLink, graduated with a bachelor's degree in electronics from Harbin Institute of Technology, ranking first overall in his class and earning the only qualification for direct doctoral admission to Tsinghua University. Yet he gave up that opportunity to pursue a master's degree in chips at Duke University. His advisor presented him with two paths—pursue a PhD and become a 50-year-old architect, or go into business—and he chose the latter without hesitation. He then completed a full-time MBA at MIT Sloan School of Management, where he encountered classmates working in optical communications and learned about OIO, which at the time was a purely science-fiction concept, planting the seed of "light." After returning to China with his MBA, Chen Chao gained experience in core strategy teams at major domestic internet companies. After leaving the corporate world, he joined Zhenzhi Ventures as managing partner while pursuing a doctorate in innovative leadership engineering at Tsinghua University. Zhenzhi Ventures' "entrepreneur-producer and deep incubation" model essentially involves doing entrepreneurship through investment, and this experience allowed him to bridge the gap from "industry perspective" to "entrepreneurial perspective," ultimately driving him to personally go all-in on founding OptoLink. Zhenzhi Ventures has also invested in three consecutive rounds from the start.

Zhang Weiwei, mentioned earlier, studied under Professor Graham Reed, hailed as the global "father of silicon photonics," a Fellow of the Royal Academy of Engineering and a professor at the University of Southampton's Optoelectronics Research Centre (ORC). Before joining OptoLink, Zhang Weiwei stood at the pinnacle of the academic ivory tower. He held a tenured professorship in the UK, was a core researcher in Reed's group on silicon photonic modulators, and published first-author papers in Nature Photonics. His collaboration with Reed forms the solid academic foundation for OptoLink's OIO technology route. In 2024, Zhang Weiwei made a decision that shocked his academic peers—giving up his tenured position to return to China and start a company. He poured over a decade of deep academic expertise in silicon photonics into OptoLink without reservation, devoting all his energy to silicon photonic process development, enabling OptoLink to handle complex micro-ring yield optimization with ease. During a late-night conversation with Chen Chao, Zhang Weiwei once shared his thinking: "If a technology remains only at the level of papers and academia, only producing beautiful data on precision optical benches in the lab, without truly landing in industry to solve real-world physical problems, then in a sense, it cannot be considered a true success."

Hu Zhipeng, a co-founder with a PhD in optoelectronics from Zhejiang University, conducted a series of optical communication device research at university and after graduation led participation in multiple national-level inter-chip/on-chip optical interconnect projects. In China, few people have fully participated in such projects. He met Chen Chao earlier, engaging in intensive technical and industry trend discussions at industry exhibitions years ago. The final co-founder, Ye Yafei, is an outstanding graduate of Tsinghua University's School of Integrated Circuits. He has been responsible for high-speed communication design at overseas major companies and has two entrepreneurial experiences, with ample expertise in electrical chips, high-speed interconnect, and team management. Chen Chao, as a fellow alumnus, got to know Ye Yafei through classmates early on and communicated with him for over a year to persuade him to join.

The "CEO + three co-founders" structure allows each to play their role: Chen Chao translates both industrial and strategic perspectives into the company's development pace, deciding where to place bets and how fast to advance. Zhang Weiwei, Hu Zhipeng, and Ye Yafei focus on technological breakthroughs and engineering implementation, turning optoelectronic chips step by step into viable products with fully integrated design, manufacturing, and packaging. Beyond building one of the rare "full-strength" optoelectronic integration teams in the optical communications track, they share a common driving force: ambition and a desire to leave something behind for China's semiconductor industry. In Chen Chao's view, this is precisely the essence of "entrepreneurship"—the Latin root meaning "to see, to explore, to discover." Technology entrepreneurs are essentially explorers expanding frontiers, where technical ideals and entrepreneurial spirit are both indispensable.

Tackling the Hardest Challenges

Compared to credentials and experience, overcoming obstacles and final results best demonstrate a team's strength. In the optical interconnect track, there are two main technology routes: the traditional Mach-Zehnder modulator (MZI) route and the micro-ring modulator route representing the next frontier. The MZI route is relatively mature but suffers from large size, high power consumption, and limited bandwidth density, unable to support dense optical interconnect for hundred-thousand-GPU clusters on space-constrained chips. In contrast, the MRM is extremely small (one-hundredth or even one-thousandth the size of MZI), consumes minimal power, and offers high bandwidth density, widely recognized as the ideal solution for next-generation CPO and OIO. However, precisely because the MRM is so tiny and controls optical signals differently from traditional devices, it faces manufacturing challenges. Micro-ring devices demand extremely high process yields. Although the overall size requirements can be met with 65nm process technology, microscopic structural defects cannot be completely eliminated even with expensive 3nm processes. Structural defects directly affect the micro-ring's actual operating wavelength, rendering the entire optical chip unusable. As a result, the nanoscale manufacturing tolerances of traditional fabs are enough to cause wavelengths across an entire wafer to be uneven, unable to precisely align with optical signal wavelengths, with yields too low for large-scale commercialization. Chen Chao revealed a figure: "Current process yields may be below 10%."

How to improve yield after micro-ring optical chips are manufactured, bringing wavelengths up to spec, is precisely the direction Zhang Weiwei studied academically. His 2023 paper in a Nature sub-journal, "Harnessing plasma absorption in silicon MOS ring modulators," and his 2025 publication in Advanced Photonics Research, "Towards High Resolution Trimming of Silicon Photonic Waveguides," provided academic-level technical solutions. During his time under Professor Graham Reed, Zhang Weiwei also gained exposure to European silicon photonic production lines, giving him deep familiarity with manufacturing equipment. This deep accumulation quickly translated into tangible results: in just over six months, Zhang Weiwei developed the first-generation "silicon photonic maskless programmable lithography machine" specifically for OIO optical chip trimming. This lithography tool can project with high speed and precision without a mask, combined with a high-precision laser annealing system, to achieve wavelength trimming of micro-rings on optical chips. Importantly, this equipment is not for technical validation but for real industrial-scale mass production.

Zhang Weiwei shared two sets of equipment performance figures: after wavelength trimming, yield on wafers improved from below 10% to 99%—from mostly scrapped to nearly all usable; the equipment is also highly automated, completing "inspection-trimming-shipment" across an entire wafer in just a few hours. With production-stage yield solved, the next challenge is temperature control during operation. Optical chips are sensitive to ambient temperature—a 0.1°C fluctuation is enough to shift the micro-ring wavelength and cause signal interruption. To solve the temperature sensitivity challenge, OptoLink independently developed a high-performance micro-ring electrical driver chip and pioneered a microsecond-level wavelength locking algorithm, creating a "temperature control system" with microsecond response speeds. Whether data center servers run at full load generating high temperatures or at idle with low temperatures, the system monitors the actual state of the micro-rings in real time through the electrical chip and completes dynamic wavelength adjustment and locking in an extremely short time.

The defining characteristic of these challenges is that they often involve both electrical and optical design simultaneously, requiring alignment at the team level from the very start. In Hu Zhipeng's own words: "The optical and electrical teams have been in deep communication all along. I've been helping Yafei (Ye Yafei) build understanding of optical devices, and Yafei has been helping me build understanding of electrical chips." Ye Yafei describes it in more detail: "We understand each other's difficulties and advantages. We figure out how to use your advantages to make things easier on our side, or use our strengths to make your challenges easier."

A Precious Window of Opportunity

What OptoLink aims to do, in Chen Chao's own words, is not "to grab a share of the track, nor to be a domestic substitute, but to become the 'definer' of the next-generation computing architecture—participating in and defining the industry standards and technology direction for optical I/O in next-generation computing architecture globally." Inevitably, OptoLink faces competition from overseas players. Global OIO leaders like Ayar Labs and Lightmatter, after trying various other routes, have ultimately converged on micro-ring technology. NVIDIA, in March of this year, invested $2 billion each in Lumentum, a leader in optical communication components, Coherent, a leader in lasers and optical modules, and Marvell, a company focused on data center interconnect and custom chips. The flood of investment is rapidly accelerating the pace of development and competitive intensity in the OIO track. Inside OptoLink, the reaction is more excitement than worry.

As results have emerged, OptoLink is gaining increasing recognition from capital. In the recently completed Series A+ round, the company's valuation exceeded $1 billion. Investors include China Life Capital, China Merchants Capital, Xinglian Capital, CSC Financial, Henan Huirong, and others, along with well-known entrepreneur individual investors such as Pei Zhenhua and Zhao Hongxiu. Existing shareholders including Sequoia, Hillhouse, and Legend Capital all oversubscribed in follow-on rounds. Additionally, China holds an overall advantage in the OIO track: from the perspective of OIO optical chip demand, domestic 55/65nm processes are sufficient once trimming processes mature, and China's most advanced logic processes are adequate for electrical chips, not to mention the country's strong capabilities in advanced packaging. OIO is not just feasible—it is becoming an opportunity for China's semiconductor industry to overtake on the curve.

Chen Chao still remembers the pivotal moment: at GTC 2025, Jensen Huang singled out MRM optical interconnect, declaring that "next-generation AI infrastructure must fully transition to optical connectivity." This declaration brought immense joy to Chen Chao, who founded OptoLink in 2024. He recalls his state of mind at the time: "When we started the company in 2023 to 2024, we only had a 'vague sense of correctness' about the direction. After Jensen Huang's announcement, we finally confirmed we had seen correctly, seen accurately." But this excitement didn't slow the team down; instead, they plunged back into the next phase of technology and product development. On the day OptoLink's self-developed micro-ring modulator version 3.2T NPO successfully passed the 112G PAM4 test eye diagram, the result was shared in the group chat by Ye Yafei. The team's celebration was understated—just simple replies of "Awesome!" and "Drinks tonight!" Hu Zhipeng explained: "The reason is simple—we're all confident in our team, because when we went to tape-out, we already felt it would succeed, so this is just the expected result."

Chen Chao's explanation is more poetic: "What we're advancing, in the larger sense, is the frontier of human civilization." Over the past year, the computing power storm brought by AI large models has been more ferocious than anyone imagined, and OptoLink's technological evolution and engineering iteration speed have equally exceeded industry expectations. At last year's CIOE (China International Optoelectronic Expo), OptoLink didn't have its own booth. Its product was displayed in a corner of a partner's laboratory booth, occupying about 1.5 square meters: a first-generation 8-channel micro-ring evaluation board, physically functional, displayed statically. Micro-rings—something nobody else had. Internally, the third generation had already completed iteration, two generations ahead. Many visitors came to ask questions. A representative from a major domestic tech company walked over, looked, and exclaimed: "You're moving faster than we are." That evening, the four founders left the convention center and sat down at a street stall. It was raining, and in the half-indoor, half-outdoor space, they ate skewers and talked about next year—whether they could have their own booth; and about one year, many years later, where this company might go.

One year later, as CIOE 2026 approaches, OptoLink now has its own booth, with globally debuted results ready to be unveiled. From 1.5 square meters to 100 square meters. Exactly one year. About this appearance, Chen Chao says, "I'm holding myself back, waiting for the milestone event, waiting for the right moment to make a splash." In OptoLink's laboratory, there's a piece of equipment covered with a cloth. Chen Chao says, "This is prepared for this year—something truly stunning, which may create a domestic and even world first while advancing the global optical interconnect industry."