There is a moment, standing in a hyperscale data centre, when the sheer physicality of the internet becomes impossible to ignore. Racks of servers stretch into the distance, each one packed with processors, memory, and storage. But what really catches the eye – once you look past the blinking lights – is the cabling. Fibre optic strands, bundled thicker than a fist, carrying pulses of light between machines that must talk to one another millions of times per second. For years, the actual conversion of those light pulses into the electrical signals that chips understand happened inside discrete optical modules, built from exotic materials and assembled by hand. That model is now hitting its limits, and a technology that spent two decades in university cleanrooms is finally stepping into the industrial mainstream.

Silicon photonics – the practice of building lasers, modulators, waveguides, and photodetectors directly onto silicon wafers, alongside ordinary electronic circuits – has crossed a threshold in 2026. It is no longer a promising lab experiment or a niche product for long-haul telecom networks. It is starting to rewire the internet at the rack level, inside the very data centres that power artificial intelligence, cloud computing, and the streaming services billions of people use every day.

The numbers are beginning to reflect that shift. Industry estimates place the global silicon photonics market at around 4.1 billion US dollars in 2025. From there, the trajectory points to roughly 13.8 billion by 2034, growing at a compound annual rate of 10.7 percent. That is not the explosive, triple-digit growth of a consumer app, but for a semiconductor-adjacent hardware market, it tells a story of steady, structural demand that is being pulled forward by forces that show no sign of slowing.

Why copper has started to sweat

The problem that silicon photonics solves is deceptively simple. Copper traces on a circuit board can only carry high-speed electrical signals a few centimetres before the signal degrades beyond recognition. For decades, engineers worked around this by bundling multiple slower lanes together, but as data rates climb past 100 gigabits per second per lane, the physics becomes punishing. You need more power, more equalization, more shielding, and even then you are left with a hot, power-hungry link that cannot stretch across a data centre aisle.

Light does not have this problem. A single fibre can carry multiple wavelengths, each modulated at rates that would melt a copper wire, and it can do so over kilometres without meaningful loss. The trick has always been building the components that generate and detect that light cheaply enough to compete with plugging in a cable. Silicon photonics answers that by hijacking the same manufacturing infrastructure that produces millions of microprocessor chips. Instead of assembling optical components from gallium arsenide or indium phosphide by hand, you fabricate them on silicon wafers using the same lithography tools that make logic chips. The result is optical transceivers that are smaller, cheaper, and more power-efficient than anything that came before, and they can be produced in volumes that are starting to make the copper alternative look expensive.

What happened at the big optical conference this year

The real measure of where a technology stands often comes not from earnings calls but from the corridors of major industry conferences. In March 2026, the Optical Fiber Communication Conference – OFC – took over the San Diego Convention Centre, and silicon photonics was everywhere. Several major transceiver vendors showcased 1.6 terabit-per-second optical modules built entirely around silicon photonic chips. What was notable was not the speed itself, which had been demonstrated before, but the fact that multiple suppliers were now offering the modules as catalogue products with delivery dates, not as prototypes with a plea for funding.

According to reporting from Lightwave, a trade publication that covers the optical networking industry, the conversation at OFC had shifted from “can silicon photonics handle high power” to “how quickly can you scale production.” One major switch maker announced a new line of top-of-rack switches with integrated silicon photonic ports, eliminating the need for pluggable transceivers altogether. The integration path – moving optics closer to the silicon, eventually onto the same package as a network chip or even a processor – is becoming the near-term roadmap rather than a distant vision.

AI training clusters are the new hunger

Behind a great deal of this urgency is artificial intelligence. Training a large language model or a generative video model requires tens of thousands of GPUs or custom accelerators working in parallel, and those chips need to share enormous volumes of data constantly. A single training run can tie up terabits per second of east-west traffic between servers. The electrical interconnects that once handled this load are now responsible for a significant fraction of the total power budget and a growing share of the latency.

In April 2026, Reuters reported that a major cloud provider – widely believed to be Microsoft, though the company did not confirm details – had begun deploying silicon photonic interconnects inside a new AI training cluster being built in Iowa. The move was described internally as a “necessary step” to keep the cluster’s networking energy under 20 percent of the total facility power. That is a threshold that becomes very hard to stay under when using conventional optics, let alone copper. The story, while understated, sent a ripple through the industry because it signalled that hyperscale operators are now willing to bet their most expensive infrastructure on a technology that was, until recently, considered risky.

The economics are starting to make a compelling case. A silicon photonic link that burns a few watts per terabit can replace an electrical link that burns several times that. Multiply that across a campus of a hundred thousand accelerators, and the savings in electricity alone run into millions of dollars a year. For cloud companies that are also facing pressure to meet sustainability targets and manage local grid constraints, that arithmetic is difficult to ignore.

Bringing optics even closer to the chip

The real frontier in 2026 is not the pluggable module, as important as those are. It is co-packaged optics – placing the optical engine directly on the same substrate as a switch ASIC or an AI accelerator. This eliminates the last few centimetres of electrical trace that have been a bottleneck and allows bandwidth densities that are simply impossible with pluggables.

In January 2026, Ayar Labs, a silicon photonics startup with backing from Intel and NVIDIA, demonstrated an optical I/O chiplet that squeezes 8 terabits per second of bandwidth out of a single package. The demonstration, covered by IEEE Spectrum, showed the chiplet feeding data directly into an FPGA, simulating the kind of interface that a future GPU or AI chip might use. While the product is not yet in volume production, the integration with standard semiconductor packaging techniques – the chiplet approach that the broader industry has embraced – suggests that optical I/O is following the same trajectory that chiplets themselves did a few years ago. What was exotic is becoming modular, and what is modular is becoming standard.

The quiet manufacturing build-up

Behind the product announcements, a less visible industrial ramp is underway. Silicon photonics requires a different set of process steps than standard logic – etching waveguides, depositing germanium for photodetectors, bonding lasers made from indium phosphide onto silicon wafers. For years, the capacity for this kind of fabrication was concentrated in a few specialist fabs. Now, the world’s largest semiconductor foundries are paying attention. TSMC, which already makes silicon photonic wafers for multiple customers, has been investing in expanded capacity at its advanced packaging plants. Intel, which has its own deep history in silicon photonics, has been opening its fabs to external customers who need photonic wafers alongside logic chips.

A trade journal report from May 2026 noted that GlobalFoundries had signed a multi-year agreement with a major optical networking company to produce silicon photonic wafers at its facility in upstate New York, doubling its previous output commitment. These are not the kinds of announcements that make the evening news, but they reveal a supply chain that is being built not for a speculative boom, but for a demand curve that the industry now believes is real and durable.

What a $13.8 billion market really represents

A decade from now, when the market size has roughly tripled, silicon photonics will likely be so embedded in data centres, telecom networks, and possibly even inside consumer devices that it will cease to be a distinct category worth counting. The technology will simply be how high-speed connections are built, much as lithium-ion batteries became the default for energy storage without anyone still calling them “novel chemistry.” That 10.7 percent growth rate is the sound of a technology transitioning from breakthrough to background infrastructure.

The story of 2026 is that this transition is no longer a forecast. It is showing up in product catalogues, in factory investments, and in the quiet calculations of data centre operators who are trying to feed ever-larger AI models without melting the grid. The photons that race through those glass fibres may be invisible, but the market they are building is becoming hard to miss.

Discover the Full Report Now: https://semiconductorinsight.com/report/silicon-photonics-market/