# The Future of AI Is Hitting a Copper Wall. The Answer Is Light.
By 

## Article Content
Twenty-six years ago, I wrote about a future in which manipulating photons would transform the Internet and computing. In 2006, I challenged 3,000 telecommunications engineers to think about what that might mean. Some thought the idea was crazy. Now AI is helping turn that vision into commercial reality. By Jim Carroll | October 2026 A few days ago, I was on stage at the ECIA Executive Conference, talking about the extraordinary technologies that are reshaping the electronics and semiconductor industries. One of my slides carried a rather dramatic headline: copper is hitting a wall that artificial intelligence can't afford. The answer, increasingly, is light. On stage at the ECIA Executive Conference, October 2026 — the slide that sparked this story. The slide outlined how silicon photonics, a technology that uses light to transmit information through microscopic optical circuits, is moving rapidly into the world of artificial intelligence. Major semiconductor manufacturers are committing substantial resources to it. Entirely new categories of optical chips are moving into production. And the global silicon-photonics market, estimated at $2.83 billion in 2025, is projected to reach $13.18 billion by 2031. But as I stood there talking about these developments, I found myself thinking about something rather remarkable. I've been following this exact technological direction for more than a quarter-century. In 2000, my co-author Rick Broadhead and I published a book that explored how the Internet would evolve into something much bigger than the technology of the time could support. We devoted an entire chapter to the extraordinary possibilities that would emerge as scientists learned to manipulate photons, the particles associated with light. The book — From Light Bulbs to Yottabits: How to Profit by Understanding the Internet of the Future — nailed multiple trends with remarkable precision: the coming explosion of bandwidth and data to almost unimaginable scale, the reinvention of the infrastructure beneath the Internet, and, at the very heart of it, the growing importance of learning to manipulate light. That same fascination is why, all these years later, I finally brought YottaBit.com to life — a future-focused science and technology site, now live, where I dig into the big, fast, and sometimes almost unbelievable developments reshaping our world. If the story you are reading now intrigues you, it is exactly the kind of thing you will find more of there. In 2006, I returned to those ideas in a keynote to approximately 3,000 telecommunications engineers, challenging them to think about a future of optical chips, optical routers and extraordinary increases in communications capacity. In 2007, I wrote explicitly about a possible transition from electronics-based computing to photonics-based computing. And here we are in October 2026. Artificial intelligence is creating an enormous new demand for precisely the kinds of optical technologies that I was writing about all those years ago. It's an extraordinary development. And perhaps the most fascinating part is that the biggest breakthroughs might still lie ahead. It began with a book, a ridiculous number, and the future of light. Four documented milestones across twenty-six years: the 2000 book, the 2006 keynote, the 2007 prediction, and the 2026 reality. In 2000, Rick and I published From Light Bulbs to Yottabits: How to Profit by Understanding the Internet of the Future. The Internet was still remarkably primitive by today's standards. Many people accessed it over painfully slow dial-up connections. The idea of streaming high-definition video into every home was ambitious. The thought that we might one day have billions of connected devices exchanging enormous quantities of information was almost impossible for most people to comprehend. But we were fascinated by a different question. What was going to happen when the technologies carrying all that information became dramatically more powerful? We believed that the extraordinary growth of the Internet would eventually demand a fundamental reinvention of the infrastructure underneath it. It wouldn't be enough to simply make existing networks a little faster. We would need radically new capabilities for moving, processing and managing information. And we were particularly fascinated by the potential of light. Chapter Three of the book carried a title that still makes me smile: "Of Photons, Dense Waves, and the Elimination of Accountants: How the High-Speed Internet of the Future Will Soon Be Here . . . Faster than You Think!" The chapter opened with a comment from Bryan McFadden, then a vice-president with Nortel Networks: "We'll be in the business of manipulating photons." That was a rather extraordinary statement for the time. It suggested that an increasingly important part of our digital future would involve learning to control light itself: directing it, combining it, dividing it into different wavelengths, and using it to carry extraordinary amounts of information. Of course, optical fiber was already transforming telecommunications networks. Scientists and engineers had been working with light-based communications for decades. But the future we were exploring went much further. What if we could manipulate light with increasingly sophisticated optical circuits? What if we could develop new technologies capable of carrying enormous quantities of information through ever-smaller physical systems? What if those advances eventually helped us rethink the relationship between optical communication and computing itself? Those were some of the questions that fascinated us. We needed a way to describe the extraordinary scale of information that might eventually move through these networks. So we chose a word that almost nobody had heard of. Yottabit. The prefix yotta represents 10 to the power of 24 — a one followed by 24 zeros. We were trying to get readers to imagine information volumes that seemed almost absurd at the time. In the book, we wrote about the enormous quantities of data that future networks might eventually need to accommodate. I registered YottaBit.com that same year. The domain sat largely unused for 26 years, until I finally launched it as a substantial future-focused science and technology site in 2026. But the thinking behind the original book never went away. In fact, I think the story of light and information is becoming more fascinating now than it was back then. In 2006, I told 3,000 engineers where light might take us. Six years after the book was published, I delivered the opening keynote at the Society of Cable Telecommunications Engineers Conference on Emerging Technologies in Tampa, Florida. The audience consisted of approximately 3,000 telecommunications professionals, many of whom spent their careers designing, maintaining and operating the networks that carried the world's rapidly growing volume of digital information. I challenged them to think far beyond the incremental improvements that dominated many industry conversations. At the time, people were debating how fast broadband connections might become. Would consumers eventually have connections operating at 100 megabits per second? Perhaps 300? I thought that was the wrong scale of thinking. I suggested that they should be thinking about the enormous increases in network capacity that would eventually follow from advances in optical technology. And then I began discussing something rather unusual. Researchers were discovering ways to dramatically slow and manipulate light. Laboratories were experimenting with optical components that suggested entirely new ways to control information. I wondered aloud what might happen when those advances eventually found their way into practical optical routers and other networking equipment. Could we move beyond the limitations of conventional electronic systems? Could we manipulate optical signals in ways that enabled entirely new network architectures? What might happen when the capacity of optical communications increased by orders of magnitude? And then something happened that I have come to appreciate over the years. They laughed at me. I had suggested, among other things, that one day we would build a router able to stop light in its tracks, examine its packets, and in doing so unlock enormous new increases in bandwidth. To a room full of working engineers, that sounded ridiculous. A few people complained to the conference organizers — "How can you bring in someone with such crazy ideas?" Others took their skepticism to the trade press, and an industry publication ran an article that examined my "crazy observation" about the speed and capacity of future networks and essentially dismissed it. That piece, Optical Futures: Another ET Look at Light, appeared in CT's Pipeline on January 24, 2006. I did not let it go. I followed up more than a few times over the years — in print, on stage, and on my blog. And here is the part I still enjoy: before long, a number of leading optical scientists came to my defence, pointing out that the underlying science was real and the broader direction was sound. The idea that had been laughed off in the room was, in their view, a lot less crazy than it had seemed. But here's something particularly interesting. I still have the original article that I wrote for Broadband Library after that keynote. It was published in spring 2006 under the title Are We Thinking "Fast" Enough? In it, I wrote about research into slowing light and the possibilities for entirely new optical routers. I argued that rapidly advancing science would continue to produce enormous improvements in communications capacity. And I made the case that a world of zettabits and yottabits would arrive faster than many people expected. That article wasn't written last week. It wasn't a retrospective assembled after the technology had become commercially important. It was published twenty years ago. Read my original 2006 article in Broadband Library — PDF I find that rather remarkable. In 2007, I put the optical-computing prediction into words. On February 13, 2007, I published another article on my blog, titled What Happens When Light Stops? Looking back, one particular observation jumps out at me. I wrote: "Think optical-chips based on light, not today's model-T's based on electrons." I was exploring a profound technological possibility: what happens when light begins taking on more of the work traditionally assigned to electronic systems? At the time, it seemed like a dramatic proposition. Electronic computing had spent decades becoming faster, smaller and more powerful. The semiconductor industry was extraordinarily successful at continuously improving the technologies upon which modern computing depended. And yet, the basic technological architecture still depended on electrons. Scientists were already exploring optical alternatives and hybrid systems that might eventually allow light to perform functions that previously required electrical circuitry. I believed that the long-term implications could be staggering. To be clear, I wasn't predicting the specific AI systems we use today. Nor was I identifying the precise optical architecture that NVIDIA or other semiconductor companies would eventually bring to market. I was identifying a much more fundamental possibility. The future of digital technology might eventually require us to move beyond electronics alone and begin relying much more heavily on photons. That was the important idea. And it is precisely the direction in which a substantial part of today's AI infrastructure industry is now moving. Why artificial intelligence is forcing us to rethink the connection. To understand why light has suddenly become so important, consider what is happening inside the enormous computing systems being built for artificial intelligence. For decades, we've concentrated on making processors faster. We put more transistors onto silicon chips. We developed increasingly sophisticated architectures. We found ways to accomplish extraordinary amounts of computation with less electricity and in less space. That progress continues. But AI has introduced an additional and increasingly difficult challenge. Modern AI infrastructure does not simply consist of one extraordinarily powerful processor working alone. It involves thousands, sometimes tens of thousands, of highly specialized processors operating together. Those processors constantly exchange information. Large quantities of information must travel between processors, memory systems, networking equipment and storage devices. And the more powerful these systems become, the more demanding those connections become. Think of it this way. Imagine building a factory filled with machinery that gets twice as productive every few years. But imagine that the conveyor belts moving materials between those machines cannot keep up. Eventually, the performance of the entire factory becomes limited not by the machines themselves, but by the connections between them. Something similar is happening inside large AI computing installations. We are creating tremendous new capacity to process information. But moving that information around the system is becoming increasingly difficult. For many years, the industry has relied heavily on copper-based electrical connections. Copper is cheap, reliable, familiar and extraordinarily effective over appropriate distances. It isn't going away. But as speeds increase, electrical signals become harder to maintain over longer connections. Engineers must compensate for signal degradation through increasingly sophisticated circuitry. That consumes electricity. It creates heat. It adds complexity and cost. And at the extraordinary speeds required by tomorrow's AI systems, these limitations become a serious concern. In October 2026, Broadcom published an engineering discussion of this very issue, describing how electrical connections were approaching practical limits involving distance, power consumption and the density of information they could carry. The company outlined the growing need to bring optical connections closer to the computing devices themselves. It's essentially the same technological direction I was urging people to consider in 2006. Except this time, it isn't primarily a question about what scientific research might eventually make possible. It's an engineering problem that major technology companies need to solve right now. The future of AI is hitting a copper wall. The answer is light. What actually changes when we start using light? The wonderful thing about this story is that the underlying idea isn't terribly difficult to understand. Traditional electronic circuits use electrical signals to move information. Optical systems use light. Instead of sending information entirely through metal conductors, we can encode it onto light and transmit it through optical fibers or microscopic optical pathways. That's the basic idea. But the real opportunity comes from what we can do with light once we learn to manipulate it. Different wavelengths can carry separate streams of information through the same fiber. Optical circuits can guide and modulate light at extraordinary speeds. Advanced photonic components can be placed increasingly close to the processors and switches that generate the data. And increasingly, some of those components can be manufactured using techniques developed by the semiconductor industry. This is the promise of silicon photonics. It combines aspects of optical science with the manufacturing expertise that has made silicon electronics so successful. One of the most important developments is something called co-packaged optics. Co-packaged optics moves the optical engine closer to the switch chip — shorter electrical paths, not the elimination of electronics. Traditionally, a computer networking system might move information electrically across a circuit board before converting it into light at a separate optical module. Co-packaged optics brings the optical components much closer to the central computing or switching chip. The electrical connection becomes shorter. The amount of power required to compensate for signal degradation can be reduced. And the system can potentially move more information using less energy and physical space. This is not because light automatically travels faster than every electrical signal. That's an oversimplification. The advantage comes from the way optical technology can carry enormous volumes of information across useful distances with attractive power and bandwidth characteristics. It's a different approach to solving the problem. And that's why it matters. The vision is no longer theoretical. It's moving into production. Here's where the story gets particularly interesting. In March 2025, NVIDIA announced new networking switches using silicon photonics and co-packaged optics, designed to connect enormous numbers of AI processors. The company reported substantial improvements in the power efficiency and reliability of the optical connections compared with traditional pluggable architectures. By May 2026, NVIDIA was reporting that its Spectrum-X Ethernet Photonics technology had moved into production as part of its Vera Rubin infrastructure initiative. In its published technical comparisons, the company has described optical networking that can achieve as much as five times better power efficiency for particular connections compared with conventional approaches. These are specific vendor-reported comparisons, not promises that an entire data center will use five times less electricity. But think about what they represent. One of the world's most important AI infrastructure companies is moving sophisticated optical technology into commercial systems designed to connect vast numbers of processors. That's an extraordinary milestone. And NVIDIA is hardly alone. In July 2026, United Microelectronics Corporation and SILITH announced their first mass-produced silicon-photonics wafers from a semiconductor facility in Singapore. The technology supports optical connections operating at 1.6 terabits per second, intended for AI and large-scale data-center networks. That's 1.6 trillion bits of information per second. In September, Tower Semiconductor and NewPhotonics announced high-volume shipments of specialized optical chips integrating lasers and other photonic components, supporting optical connections rated at 800 gigabits and 1.6 terabits per second. These aren't simply laboratory demonstrations. They are real manufacturing developments. And there was another announcement that caught my attention. Tower Semiconductor revealed in May 2026 that it had signed customer contracts representing $1.3 billion in silicon-photonics revenue planned for 2027, supported by $290 million in customer prepayments for manufacturing capacity. Contracted future revenue, not money already earned — but a powerful signal of commercial commitment. That's future contracted revenue, not money already earned from delivered products. But it's powerful evidence of commercial commitment. Customers are reserving production capacity for a technology that was once discussed largely in specialized research circles. The science is becoming engineering. The engineering is becoming manufacturing. And the manufacturing is becoming an increasingly important part of the AI industry. There's a big difference between a prediction and a manufacturing order. I've been writing about technological trends and their impact on business for more than three decades. One of the patterns I've learned to watch is the moment when an interesting scientific possibility starts attracting substantial industrial investment. Research is important, but it doesn't guarantee commercial success. Promising prototypes are fascinating, but they don't automatically become practical products. The real transition often happens when companies begin making long-term commitments. They build factories. They acquire specialized technology businesses. They develop manufacturing processes. They sign customer contracts. They begin shipping products. Those are the signals that a technological possibility is becoming an industrial reality. That's what makes the current developments in photonics so fascinating. A 2026 market analysis by Mordor Intelligence estimates that the silicon-photonics industry will grow from $2.83 billion in 2025 to $13.18 billion by 2031. That's almost fivefold growth. Of course, forecasts are just forecasts. The final numbers might be quite different. But the direction of investment, product development and manufacturing is already visible. And artificial intelligence is helping accelerate it. The enormous demand for AI computing has created a practical need for communications technologies that can move vastly greater quantities of information while controlling the associated energy and infrastructure costs. Silicon photonics is emerging as one of the technologies that can help meet that demand. I find it fascinating that something I was exploring in a book in 2000 is now becoming part of the industrial planning of some of the world's most important technology companies. The next possibility is even more extraordinary: computing with light. From moving information with light to performing calculations with it. So far, much of this story has been about moving information using light. But there's another frontier that takes us even closer to the possibilities I was discussing in 2007. What if light could do some of the computing itself? That's an entirely different challenge. Modern AI systems perform enormous numbers of mathematical calculations. Specialized electronic processors have become extraordinarily capable at performing those operations. But researchers have been exploring whether some of those calculations might be performed using optical components. In December 2024, scientists at MIT described a photonic processor capable of performing the key calculations involved in a particular neural-network task in less than half a nanosecond. The demonstration achieved more than 92% accuracy on its classification task. Half a nanosecond. Less than half a billionth of a second. That's extraordinary. Now, this doesn't mean that we can replace today's general-purpose electronic processors with entirely optical computers tomorrow. The experiment involved specialized calculations in a research environment. Building complete photonic computing systems that outperform conventional technology on practical workloads is an entirely different challenge. But the scientific possibility is fascinating. We are beginning to explore how the physical properties of light might be used not only to transmit information but also to perform certain mathematical operations. I explored this development in a recent article on my YottaBit site: What If the Next AI Breakthrough Was a Chip That Calculates With Light? That story examines the possibility of new computing architectures in which electronic systems handle some types of work while photonic components handle others. Perhaps the future isn't a choice between electrons and photons. Perhaps it's a sophisticated combination of the two. And perhaps the most important advances will come as scientists learn how to bring these different technologies together. That strikes me as a particularly exciting possibility. I wasn't predicting NVIDIA in 2000. I was predicting the direction of change. There's an important distinction worth making here. I didn't sit down with Rick Broadhead in 2000 and predict that NVIDIA would introduce co-packaged optical networking switches in 2025. I didn't know that artificial intelligence would create precisely the kinds of computational challenges the industry now faces. I certainly didn't predict every technical detail of modern silicon photonics. And not every idea I explored has become commercially practical. For example, the particular stopped-light router that intrigued me in 2006 has not become the foundation of modern AI networking. I acknowledged that in my own review of earlier predictions. But that's not what makes this story remarkable. What makes it remarkable is the underlying technological direction. In 2000, we were writing about a future in which rapidly growing information demands would require extraordinary improvements in network capacity, with the manipulation of photons playing an increasingly important role. In 2006, I was suggesting that optical science could create fundamentally new opportunities in computing and communications. In 2007, I wrote explicitly about the potential transition toward photonics-based computing. And now, in 2026, AI is driving major investment in technologies that use photons to address limitations of conventional electronic connections. That is a remarkable continuity of thinking across more than a quarter-century. It also demonstrates something important about how we should approach the future. Don't become fixated exclusively on the products and technologies that exist today. Pay attention to the scientific capabilities being developed underneath them. Because those capabilities often become the foundation of entirely new industries. So where does light take us next? I suspect we are still in the early stages of understanding how extensively optical technology might reshape computing infrastructure. Today, one of the biggest opportunities involves moving information more efficiently between large numbers of processors. Tomorrow, optical connections may move closer to individual computing chips, memory devices and increasingly complex processing systems. Researchers are continuing to explore advanced optical switches, integrated lasers, new photonic materials and ways to combine optical and electronic circuitry. Others are investigating entirely different computing architectures that rely on specialized optical components. There will be engineering difficulties. Manufacturing costs will matter. Reliability will matter. Heat will remain a challenge. Different technologies will compete, and some will fail. Copper will continue to be important wherever its particular characteristics make it the best choice. But the boundary between conventional electronics and optical technology is clearly shifting. And that shift will create new opportunities throughout the semiconductor industry, data-center infrastructure, communications, advanced manufacturing and scientific computing. That's the bigger story. AI is accelerating the need for optical technology. Optical technology may, in turn, help enable more powerful and efficient AI infrastructure. And advances in that infrastructure could open the door to computing possibilities we have barely begun to imagine. That's how major technological transformations happen. One scientific development creates opportunities for another. Different technologies begin reinforcing each other. And suddenly, the future starts moving much faster than anyone expected. Twenty-six years later, the future is catching up with the photons. I keep coming back to that original book. From Light Bulbs to Yottabits. Published in 2000. An entire chapter about photons, dense waves and the extraordinary technologies that might transform digital communications. A quote about learning to manipulate light. A prediction that the world would eventually need to deal with information volumes that seemed almost absurd. Then came the 2006 keynote. The skeptical engineers. The industry article. My 2007 prediction about optical chips. And now, the extraordinary developments unfolding in 2026. The details are different from some of the possibilities I imagined all those years ago. The specific technologies have evolved in ways that were difficult to anticipate. But the underlying direction has become remarkably clear. As our hunger for computing power and communications capacity continues to grow, we're going to need far more than incremental improvements in conventional electronics. We're going to need new ways to move information. New ways to manipulate it. And perhaps, eventually, fundamentally new ways to compute with it. Increasingly, we're turning to light. There's something rather satisfying about seeing a scientific possibility I've followed for 26 years moving into the industrial mainstream. Not because it means I was right about every detail. But because it reinforces something I've believed throughout my career. The most important technologies of tomorrow are often already being developed today. The challenge is recognizing how consequential they might eventually become. Twenty-six years ago, I was writing about the extraordinary possibilities of manipulating photons. Today, I'm watching major technology companies invest billions of dollars in bringing optical technologies into the next generation of computing. And the most fascinating part? I suspect the biggest breakthroughs are still ahead of us. The future of AI is hitting a copper wall. The answer is light. Historical articles and original sources From Light Bulbs to Yottabits: The 26-Year Story Behind YottaBit Are We Thinking "Fast" Enough? — Jim Carroll, Broadband Library, Spring 2006 (PDF) What Happens When Light Stops? — February 13, 2007 People Will Always Laugh at Your Crazy Ideas: They Are Usually Wrong! — March 2017 What If the Next AI Breakthrough Was a Chip That Calculates With Light? — YottaBit Modern developments NVIDIA — Spectrum-X Photonics and Co-Packaged Optics, March 2025 NVIDIA — Optical Networking Enters Production, May 2026 UMC and SILITH — Mass-Produced Photonic Wafers, July 2026 Tower Semiconductor — $1.3 Billion in Silicon-Photonics Contracts Tower and NewPhotonics — High-Volume Optical Chip Shipments, September 2026 Broadcom — Scaling AI Infrastructure With Integrated Optics, October 2026 MIT — Photonic Processor for AI, December 2024 Mordor Intelligence — Silicon Photonics Market Forecast

---
Source: https://jimcarroll.com/2026/10/the-future-of-ai-is-hitting-a-copper-wall-the-answer-is-light/